JP7576290B1 - TOPIC MODULE SET CREATION DEVICE, INTERACTION DEVICE, TOPIC MODULE SET CREATION METHOD, INTERACTION METHOD, AND COMPUTER PROGRAM - Google Patents

Abstract

To continue a conversation even when complex topic transition patterns cannot be predicted.
[Solution] A dialogue device comprising: a topic module set creation unit that creates a topic module set including multiple topic modules indicating topics to be provided to a user who is the dialogue partner when certain conditions are satisfied, based on one or more scenarios that describe a one-way dialogue flow; a topic determination unit that determines a topic according to the user's status or the system based on the topic module set created by the topic module set creation unit and status information that represents at least the status of the user or the system; and an output unit that outputs content according to the determined topic.
[Selected Figure] Figure 1

Description

Article 30, paragraph 2 of the Patent Act applies. Publication date (publication date) May 3, 2023 Publication name Journal of the Robotics Society of Japan Vol. 41, No. 3 April 2023, pp. 291-302 Robotics Society of Japan [Web publication] URL: https://www.jstage.jst.go.jp/article/jrsj/41/3/41_41_291/_article/-char/ja DOI: https://doi.org/10.7210/jrsj.41.291 <Materials> Journal of the Robotics Society of Japan Academic and technical papers published

The present invention relates to a topic module set creation device, a dialogue device, a topic module set creation method, a dialogue method, and a computer program.

Conventionally, research has been conducted on dialogue systems between humans and dialogue robots (see, for example, Non-Patent Document 1). Approaches to implementing dialogue systems include methods using machine learning and methods using finite state machines. Methods using machine learning can control the system using learned patterns and can return some kind of response in response to human speech, allowing for continued dialogue. On the other hand, methods using machine learning rely on learned patterns, and therefore may not be able to carry out the dialogue intended by the designer. Furthermore, even if a specific behavior needs to be modified, it cannot be easily done.

In contrast, with the method using finite state machines, the designer writes the rules manually, making it easier to realize the interactions the designer intended. Furthermore, if a specific behavior needs to be modified, it is only necessary to modify the rules related to that specific behavior. For this reason, the method using finite state machines is also widely used.

Kazunori Komatani, “Configuration and Future of Speech Dialogue Systems”, [online], [Retrieved June 9, 2023], Internet <URL: https://system.jpaa.or.jp/patent/viewPdf/3307>, Vol. 72 No. 8

When using a finite state machine, it is necessary to design all contexts for topic transitions. As a result, the design can become complicated as the number of anticipated transition patterns increases. In actual operation, it is difficult to design a system that covers all transition patterns, and in such cases, there is a problem that topic transitions cannot be made in undesigned areas, causing the dialogue to break down. This type of problem is not limited to spoken dialogue between humans and conversational robots, but is also common in text-based dialogue.

In light of the above, the present invention aims to provide technology that can continue a conversation even when complex topic transition patterns cannot be fully anticipated.

One aspect of the present invention is a topic module set creation device that includes a topic module set creation unit that creates a topic module set including a plurality of topic modules indicating topics to be provided to a user who is the other party of a conversation when certain conditions are satisfied, based on one or more scenarios that describe a one-way flow of dialogue, and at least one or more topic modules are set with an activation condition including at least one combination of candidate states and variables for representing the state of the user or system, and the topic module set creation unit creates a plurality of topic modules that are sets of the candidate states and the topics by corresponding each topic module to content corresponding to the topic indicated in the topic module, and creates the topic module set by arranging the created plurality of topic modules in a hierarchical structure in a predetermined priority order .

One aspect of the present invention is an interactive device comprising the above-mentioned topic module set creation device , a topic determination unit that determines a topic corresponding to the state of the user or the state of the system based on the topic module set created by the topic module set creation device and state information representing at least the state of the user or the system, and an output unit that outputs content corresponding to the determined topic.

One aspect of the present invention is the dialogue device described above, in which at least one or more topic modules are set with a start condition including at least one combination of a candidate state and a variable for expressing the state of the user or the system, and the topic module set creation unit creates the topic module set by associating content according to the topic indicated in the topic module with each topic module and arranging multiple topic modules in a hierarchical structure in a predetermined priority order.

In one aspect of the present invention, the dialogue device further includes a state information creation unit that creates the state information by associating a variable with each candidate state, with one scenario being composed of multiple topics, and with each topic constituting the scenario being treated as a candidate state.

One aspect of the present invention is the dialogue device described above, in which the state information creation unit, when creating the state information, integrates multiple topics that have similar meanings among the topics that make up the scenario into one topic.

One aspect of the present invention is the above-mentioned dialogue device, further comprising a state update unit that updates state information related to the user in response to at least the content of the user's utterance or the content input by text, and updates state information related to the system in response to the content of the utterance or an action in the system, and the state update unit updates the state information by updating variables of candidate states in response to the content of the user's utterance or the content input by text, or the content of the utterance or an action in the system.

One aspect of the present invention is the dialogue device described above, in which the topic determination unit determines, based on the state information, a topic with a high priority among the topics associated with the satisfied activation conditions as a topic corresponding to the state of the user or the system.

In one aspect of the present invention, in the dialogue device described above, when there is no topic module that satisfies the activation condition, the topic determination unit selects a dialogue module for which the activation condition is not set.

One aspect of the present invention is the dialogue device described above, further comprising an operation control unit that controls the operation of a robot connected to the device, an agent displayed on a display device, or a voice output device based on the topic or information based on the topic determined by the topic determination unit.

One aspect of the present invention is a topic module set creation method in which a computer creates a topic module set including a plurality of topic modules indicating topics to be provided to a user who is the other party in a conversation when a certain condition is satisfied, based on one or more scenarios describing a one-way dialogue flow, at least one or more topic modules are set with an activation condition including at least one combination of candidate states and variables for representing the state of the user or system, multiple topic modules which are sets of the candidate states and the topics are created by matching content corresponding to the topics indicated in the topic modules with each topic module, and the topic module set is created by arranging the created plurality of topic modules in a hierarchical structure in a certain priority order .

One aspect of the present invention is a dialogue method in which a computer determines a topic corresponding to the state of the user or the system based on the topic module set created by the above-mentioned topic module set creation method and status information representing at least the state of the user or the system, and outputs content corresponding to the determined topic.

One aspect of the present invention is a computer program for causing the topic module set creation device or the dialogue device to function as described above.

This invention makes it possible to continue a conversation even when complex topic transition patterns cannot be fully anticipated.

FIG. 1 is a diagram illustrating an example of a configuration of a dialogue system according to an embodiment. FIG. 4 is a diagram illustrating an example of state information according to an embodiment. FIG. 1 is a diagram showing an example (part 1) of a topic module set in an embodiment. FIG. 4 is a diagram illustrating an example of output language information in the embodiment. FIG. 2 is a diagram illustrating an example of a scenario in an embodiment. FIG. 11 is a diagram for explaining a method for creating state information in the embodiment. FIG. 11 is a diagram for explaining a method for creating state information in the embodiment. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. 11 is a diagram for explaining a method for creating a topic module set in an embodiment. FIG. FIG. 1 is a sequence diagram (part 1) showing a process flow of the dialogue system in the embodiment. FIG. 11 is a diagram illustrating an example of updating state information in the embodiment. FIG. 13 is a diagram showing an example (part 2) of a topic module set in the embodiment. FIG. 1 is a sequence diagram (part 1) showing a process flow of the dialogue system in the embodiment. FIG. 11 is a diagram illustrating an example of updating state information in the embodiment. FIG. 11 is a diagram illustrating an example of updating state information in the embodiment. FIG. 11 is a diagram showing an example (part 3) of a topic module set in the embodiment. FIG. 3 is a sequence diagram (part 3) showing the processing flow of the dialogue system in the embodiment. FIG. 11 is a diagram illustrating an example of updating state information in the embodiment. FIG. 11 is a diagram showing an example (part 4) of a topic module set in the embodiment. FIG. 4 is a sequence diagram showing the processing flow of the dialogue system in the embodiment (part 4). FIG. 11 is a diagram illustrating an example of updating state information in the embodiment. FIG. 13 is a diagram illustrating an example of a configuration of a dialogue system according to a modified example. FIG. 5 is a diagram showing an example (part 5) of a topic module set in a modified example.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
(summary)
In the dialogue device of the present invention, when a predetermined condition is satisfied, a topic to be provided to a user is determined based on a topic module set including multiple topic modules indicating a topic to be provided to the user, who is the dialogue partner, and status information indicating the user's status according to the user's actions, the user's speech content, or the content input by text, and content based on the determined topic is output to the user (audio output or text output). The topic to be provided to the user, who is the dialogue partner, is material for dialogue with the user, and may be any matter. The topic module set is composed of multiple topic modules arranged in a hierarchical structure according to priority order.

The dialogue in this invention includes a dialogue by voice between a user and an interactive device, a dialogue by text between a user and an interactive device, and a dialogue by voice in which the user or the interactive device dialogues and the other dialogues by text. In the explanation of one embodiment, a dialogue by voice between a user and an interactive device will be used as an example.

Furthermore, the dialogue device of the present invention creates a topic module set and state information based on one or more scenarios in which a one-way dialogue flow is described. Here, the one-way dialogue flow refers to a flow of topics that is expected to be exchanged until a certain subject reaches a goal that the subject wishes to achieve in a dialogue with the dialogue partner. For example, assume that the goal that the subject wishes to achieve in a dialogue with the dialogue partner is "recommend a ramen shop." In this case, possible one-way dialogue topics that are expected to be exchanged to reach the goal include "Are you hungry?", "Want to eat ramen?", "Can I recommend a ramen shop?", etc. A one-way dialogue flow could be asking "Are you hungry?" followed by "Want to eat ramen?", followed by asking "Can I recommend a ramen shop?", followed by "I will recommend a ramen shop."

The dialogue device of the present invention creates a topic module set and state information using one or more scenarios that describe the one-way dialogue flow as described above. The dialogue device then uses the created topic module set and state information to determine a topic to be provided to the user, and outputs content based on the determined topic to the user (audio output or text output).

The topic module set is created based on the subsumption architecture technology used in the field of mobile robots. The topic module set in the present invention differs from module sets created using conventional subsumption architecture technology in that the topic module set can be created automatically based on a scenario and that the user's state is taken into account. For example, in the subsumption architecture used in the field of mobile robots, a module set is constructed by arranging each pre-created module according to a predetermined priority. In contrast, the dialogue system in the present invention can automatically create a topic module set based on a scenario that describes a one-way dialogue flow.

Furthermore, in the subsumption architecture used in the field of mobile robots, whether or not the activation condition is satisfied is determined based only on the information (e.g., sensor output) input to each module constituting a module set, and information on behavior when the condition is satisfied is output. In contrast, in the dialogue system of the present invention, whether or not a predetermined condition is satisfied is determined based on a plurality of candidate states for representing the user's state, such as information input to each topic module constituting a topic module set (e.g., the user's state according to the content of the user's utterance = current state) and history information of the dialogue with the user conducted up to that point (e.g., the user's state in the past dialogue = past state), and information on behavior when the predetermined condition is satisfied is output. In this way, in the dialogue system of the present invention, the next topic to be provided to the user is determined from the flow of the past dialogue, so that it is possible to continue the dialogue even when complex topic transition patterns cannot be predicted.
Each embodiment will be described in detail below.

(Embodiment)
[Configuration of dialogue system 100]
1 is a diagram showing an example of a configuration of a dialogue system 100 according to an embodiment. The dialogue system 100 includes a dialogue device 10, a camera 20, a microphone 30, a speaker 40, and a display device 50. The camera 20, the microphone 30, the speaker 40, and the display device 50 are connected to the dialogue device 10 by wire or wirelessly.

The dialogue device 10 determines a topic to be provided to the user who is the subject of the dialogue, and realizes a dialogue with the user by outputting content corresponding to the determined topic. For example, the dialogue device 10 realizes a dialogue with the user by outputting a voice having content corresponding to the determined topic. The dialogue device 10 is configured using an information processing device such as a personal computer.

The camera 20 captures moving images of the surroundings of the dialogue device 10. The camera 20 generates a video signal corresponding to the captured moving images, and inputs image information based on the video signal to the dialogue device 10.

The microphone 30 picks up sounds around the dialogue device 10. For example, the microphone 30 acquires the voice of a user approaching the dialogue device 10. The microphone 30 generates a voice signal based on the acquired voice. The microphone 30 outputs the generated voice signal to the dialogue device 10. Note that the microphone 30 may be provided inside the dialogue device 10.

The speaker 40 outputs an audio signal generated by the dialogue device 10. For example, the speaker 40 outputs audio content based on a determined topic. The speaker 40 is provided near the display device 50 (for example, to the side of the display device 50 or behind the display device 50).

The display device 50 is an image display device such as a liquid crystal display, an organic EL (Electro Luminescence) display, or an electrophoretic display. The display device 50 displays an agent represented in two dimensions. The agent represented in two dimensions is, for example, a character displayed on the screen of the display device 50. The speaker 40 and the display device 50 may be integrated.

Next, the functional configuration of the dialogue device 10 will be described. The dialogue device 10 includes a memory unit 11 and a control unit 12. The memory unit 11 stores a dictionary 111, state information 112, a topic module set 113, output language information 114, and operation control information 115. The memory unit 11 is configured using a storage device such as a magnetic storage device or a semiconductor storage device.

Dictionary 111 is a dictionary used for semantic analysis in natural language processing.

The state information 112 is information that represents the state of the user. The user here is a user who is a target of dialogue with the dialogue device 10. FIG. 2 is a diagram showing an example of the state information 112. As shown in FIG. 2, the state information 112 is composed of a plurality of candidate states for representing the state of the user, and variables associated with each of the plurality of candidate states. Note that in the example shown in FIG. 2, the candidate states show the state of the user determined based on the user's utterance, but the candidate states may also include states related to the user's actions (e.g., a person approaching, a person stopping, a person waving, etc.) detected from an image obtained by the camera 20. The candidate states represent, for example, states that the user can take in terms of topics that are expected to be exchanged before moving toward a goal to be achieved in a dialogue with the user.

For example, suppose the goal to be achieved in a dialogue with a user is "recommend a ramen shop" or "recommend a pasta restaurant." In this case, possible topics that may be exchanged to reach the goal include "Are you tired?", "Are you hungry?", "Do you want to eat ramen?", "Do you want to eat pasta?", "Can I recommend a ramen shop?", "Can I recommend a pasta restaurant?", etc. Taking these topics into consideration, the following states of the user are possible: being tired, being hungry, wanting to eat ramen, wanting to eat pasta, wanting a ramen shop recommendation, wanting a pasta restaurant recommendation, having heard a ramen shop recommendation, and having heard a pasta restaurant recommendation.

The variables used are Y (Yes), which indicates that the user's state is a corresponding candidate state, N (No), which indicates that the user's state is not a corresponding candidate state, and U (Unknown), which indicates that it is not specified whether the user's state is a corresponding candidate state or not. Y is one aspect of the first variable, N is one aspect of the second variable, and U is one aspect of the third variable. When the dialogue device 10 starts processing, all of the candidate state variables are set to initial values (e.g., U).

The above-mentioned topics and candidate states are just examples, and designers can freely set the topics and candidate states. In the following explanation, we will use the above-mentioned topics and candidate states as examples to explain the configuration.

Returning to Figure 1, the explanation continues. The topic module set 113 is composed of multiple topic modules that indicate the topics to be provided to the user who is the conversation partner when certain conditions are met. Figure 3 is a diagram showing one example (part 1) of a topic module set 113. The topic module set 113 is composed, for example, of topic modules 113-1 to 113-9 arranged in a hierarchical structure with a predetermined priority as shown in Figure 3. The priority of the hierarchical structure can be freely changed depending on the purpose of the designer.

In the example shown in FIG. 3, all topic modules 113-n (n is an integer equal to or greater than 1) are arranged in order of distance from the goal (the number of topics passed through to reach the goal). For example, suppose that the goals to be achieved in a dialogue with a user are "recommend a ramen restaurant" and "recommend a pasta restaurant", and the topics leading to the goal "recommend a ramen restaurant" are "Are you hungry?" ⇒ "Want to eat ramen?" ⇒ "Can we recommend a ramen restaurant?" ⇒ "Recommend a ramen restaurant", and the topics leading to the goal "recommend a pasta restaurant" are "Are you tired?" ⇒ "Are you hungry?" ⇒ "Want to eat pasta?" ⇒ "Can we recommend a pasta restaurant?" ⇒ "Recommend a pasta restaurant".

In this case, the topic "Are you tired?" is the furthest from the goal (distance = 4), followed by "Are you hungry?" (distance = 3), followed by "Do you want to eat ramen?" and "Do you want to eat pasta?" (distance = 2), and then the topics "Can I introduce you to a ramen shop?" and "Can I introduce you to a pasta shop?" (distance = 1). Then, by arranging each topic in a predetermined order of priority (for example, in order of distance from the goal, and giving priority to ramen over pasta), and setting the corresponding activation conditions, the configuration shown in Figure 3 is obtained.

The start-up condition is a condition for outputting the contents of the associated topic to the user, and is composed of, for example, a combination of two or more of the states indicated in the state information 112. The start-up condition needs to be changed according to the arrangement of the topics as described above, but can also be a condition that takes into account the contents of the user's speech.

Returning to FIG. 1 for further explanation, the output language information 114 is information on a character string for dialogue corresponding to a topic determined based on the topic module set 113. FIG. 4 is a diagram showing an example of the output language information 114. As shown in FIG. 4, the output language information 114 associates a topic with an output speech character string. For example, when the topic determined based on the topic module set 113 is "ask if you are tired", it is shown that "Aren't you tired?" is selected as the character string for dialogue.

Returning to FIG. 1 for further explanation, the action control information 115 includes information for controlling the action of the agent to be displayed on the display device 50. For example, the action control information 115 may be a table in which a topic or an output voice string is associated with a control content. The control content is content for controlling the action of the agent (e.g., facial expressions, gestures, etc.).

Returning to FIG. 1, the explanation will continue. The control unit 12 controls the entire dialogue device 10. The control unit 12 is configured using a processor such as a CPU (Central Processing Unit) and a memory. By executing a program, the control unit 12 realizes the functions of a topic module set creation unit 120, a state information creation unit 121, a detection unit 122, a voice recognition unit 123, an analysis unit 124, a state update unit 125, a topic determination unit 126, a language generation unit 127, a voice synthesis unit 128, and an operation control unit 129.

Part or all of the topic module set creation unit 120, state information creation unit 121, detection unit 122, voice recognition unit 123, analysis unit 124, state update unit 125, topic determination unit 126, language generation unit 127, voice synthesis unit 128, and operation control unit 129 may be realized by hardware (including circuitry) such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array), or may be realized by a combination of software and hardware. The program may be recorded on a computer-readable recording medium. A computer-readable recording medium is, for example, a non-transitory storage medium such as a portable medium such as a flexible disk, a magneto-optical disk, a ROM, or a CD-ROM, or a storage device such as a hard disk built into a computer system. The program may be transmitted via a telecommunications line.

Some of the functions of the topic module set creation unit 120, state information creation unit 121, detection unit 122, voice recognition unit 123, analysis unit 124, state update unit 125, topic determination unit 126, language generation unit 127, voice synthesis unit 128 and operation control unit 129 do not need to be pre-installed in the dialogue device 10, and may be realized by installing additional application programs in the dialogue device 10.

The topic module set creation unit 120 creates a topic module set based on one or more scenarios that describe a one-way dialogue flow. The topic module set creation unit 120 stores the created topic module set in the memory unit 11. The topic module set created by the topic module set creation unit 120 is the topic module set 113 stored in the memory unit 11. The scenario is created in advance by the user. The scenario may be created by an external device and input to the dialogue device 10, or may be created by the user operating the dialogue device 10.

The state information creation unit 121 creates state information by treating each topic constituting the scenario as a candidate state and associating variables with each candidate state. The state information creation unit 121 stores the created state information in the storage unit 11. The state information created by the state information creation unit 121 is the state information 112 stored in the storage unit 11.

The detection unit 122 detects the behavior of a person based on video images captured by the camera 20. Examples of the behavior of a person include the movement or behavior of a person approaching the dialogue device 10, stopping, waving, etc. Note that the method of detecting the behavior of a person is not limited to this, and any method capable of detecting a person may be used. For example, the detection unit 122 may detect the behavior of a person based on information detected by a sensor (not shown). For example, the detection unit 122 may detect the behavior of a person in response to information regarding the user's behavior being input from another device.

The voice recognition unit 123 executes a voice recognition process. The voice recognition process is a process for generating a character string based on a voice signal. The voice recognition unit 123 executes the voice recognition process to generate a character string based on a voice signal output from the microphone 30. The voice recognition unit 123 may generate a character string using a known method.

The analysis unit 124 analyzes the content of the user's speech by performing natural language processing using the character string generated by the voice recognition unit 123 and the dictionary 111 stored in the memory unit 11.

The state update unit 125 updates the state information 112 based on the speech content (user's speech content) analyzed by the analysis unit 124 or the behavior of the person detected by the detection unit 122. Specifically, the state update unit 125 updates the variables of the candidate states according to the speech content based on the speech content analyzed by the analysis unit 124. Furthermore, the state update unit 125 updates the variables of the candidate states according to the speech content based on the behavior of the person detected by the detection unit 122. For example, if the behavior of the person detected by the detection unit 122 is the behavior of "a person waving his/her hand", the state update unit 125 updates the variables associated with "the behavior of a person waving his/her hand" in the state information 112. The state update unit 125 notifies the topic determination unit 126 that the state information 112 has been updated.

In response to the state information 112 being updated by the state update unit 125, the topic determination unit 126 determines a topic according to the user's state based on the state information 112 and the topic module set 113. Specifically, the topic determination unit 126 determines whether or not any of the activation conditions in the topic modules 113-n constituting the topic module set 113 is satisfied based on the user's state indicated by the state information 112, and determines the topic with the highest priority among the topics associated with the satisfied activation conditions as the topic according to the user's state.

The language generation unit 127 generates a string of characters to be output as speech based on the topic determined by the topic determination unit 126 and the output language information 114.

The voice synthesis unit 128 generates a voice signal corresponding to the character string generated by the language generation unit 127. The voice signal generated by the voice synthesis unit 128 is output from the speaker 40. In other words, the voice synthesis unit 128 causes the content according to the determined topic to be output from the speaker 40. The voice synthesis unit 128 is a positional aspect of the output unit.

The operation control unit 129 controls the operation of the agent to be displayed on the display device 50 based on the topic or information based on the topic determined by the topic determination unit 126 and the operation control information 115. Specifically, the operation control unit 129 refers to the operation control information 115 and acquires the control content associated with the topic determined by the topic determination unit 126. The operation control unit 129 controls the operation of the agent by generating video information of the agent performing an operation according to the acquired control content and displaying the generated video information on the display device 50.

When a table in which output voice character strings and control contents are associated is used as the operation control information 115, the operation control unit 129 refers to the output language information 114 and acquires the output voice character string associated with the topic determined by the topic determination unit 126. The operation control unit 129 refers to the operation control information 115 and acquires the control contents associated with the acquired output voice character string. The operation control unit 129 generates video information of an agent performing an operation according to the acquired control contents, and controls the operation of the agent by displaying the generated video information on the display device 50.

[Creating topic module sets and state information]
Next, a method for creating a topic module set and state information will be specifically described. A scenario used by the topic module set creation unit 120 to create a topic module set and state information is shown in FIG. 5. FIG. 5 is a diagram showing an example of a scenario in the embodiment. FIG. 5 shows three scenarios SC1, SC2, and SC3. Scenario SC1 is a scenario that assumes that the goal to be achieved in a dialogue with a user is "recommending a ramen shop". In scenario SC1, the topics are set in the following order as an example of a one-way dialogue flow that is assumed to be exchanged toward the goal: "Are you hungry?"->"Want to eat ramen?"->"Can I recommend a ramen shop?"->"Recommend a ramen shop". In this way, one scenario SC is composed of multiple topics.

Scenario SC2 is a scenario that assumes that the goal to be achieved in a dialogue with a user is to "recommend a pasta restaurant." In scenario SC2, therefore, as an example of a one-way dialogue flow that is expected to take place toward the goal, the topics are set in the following order: "Are you tired?" ⇒ "Are you hungry?" ⇒ "Want to eat pasta?" ⇒ "Can I recommend a pasta restaurant?" ⇒ "Recommend a pasta restaurant."

Scenario SC3 is a scenario that assumes that the goal to be achieved in a dialogue with a user is to "recommend a massage parlor." In scenario SC3, therefore, as an example of a one-way dialogue flow that is expected to take place toward the goal, the topics are set in the following order: "Are you tired?" ⇒ "Would you like to know how to relax?" ⇒ "Can I recommend a massage parlor?" ⇒ "Recommend a massage parlor."

As shown in FIG. 5, each scenario SC does not have a branch such as changing the topic. If you want to branch, such as changing the topic, just create a new scenario SC that starts with a different flow halfway through. Each scenario SC shown in FIG. 5 is created assuming only the case where the user who is the conversation partner will respond positively to content related to one topic. For example, in scenario SC1, it is assumed that the user who is the conversation partner will respond positively to content related to the topic "Are you hungry?" (for example, "Are you hungry?"), such as "Do you want to eat ramen?" is set after the topic "Are you hungry?".

It is also possible to create a scenario SC assuming a case where a negative opinion (for example, "No" or "That's not right") is responded to, but for the sake of concise and clear explanation, the scenario SC described below assumes only a case where the user who is the other party in the conversation responds with a positive opinion. Each scenario SC shown in FIG. 5 is an example, and the topic within the scenario SC may be changed as appropriate.

First, the method of creating state information will be described with reference to FIG. 6 and FIG. 7. The state information creation unit 121 receives one or more scenarios SC created as described above. Based on the input scenario SC, the state information creation unit 121 creates state information by setting each topic constituting the scenario SC as a candidate state and associating a variable with each candidate state. Specifically, as shown in FIG. 6, the state information creation unit 121 first divides each input scenario SC into topics. Next, the state information creation unit 121 sets each divided topic as a candidate state and associates a variable (for example, "U" as an initial value) with each candidate state. As described above, the variable U is a variable that indicates that it is not specified whether the user's state is the associated candidate state or not. Next, the state information creation unit 121 searches for a candidate state that has the same meaning among multiple candidate states, as shown in FIG. 7.

The state information creation unit 121 is not particularly limited in the method of searching for a candidate state having the same meaning among a plurality of candidate states. For example, a pre-trained language model may be used, or a trained model in which a teacher label indicating that two variables have the same meaning may be used. When a pre-trained language model is used, the state information creation unit 121 searches for a candidate state having the same meaning among a plurality of candidate states based on the similarity of vectors. The language model may be a transformer-based model. The state information creation unit 121 searches for a candidate state having the same meaning by any of the above-mentioned methods. In the example shown in FIG. 7, the state information creation unit 121 searches for two candidate states, "Are you hungry?", as candidate states having the same meaning, and searches for two candidate states, "Are you tired?" and "Are you not tired?", as candidate states having the same meaning.

The state information creation unit 121 selects one candidate state from among multiple candidate states with the same meaning obtained as a search result, and deletes the remaining candidate states. In the example shown in FIG. 7, the state information creation unit 121 selects one candidate state indicating "Are you hungry?" from two candidate states, for example, "Are you hungry?", and deletes the remaining candidate state indicating "Are you hungry?". Furthermore, in the example shown in FIG. 7, the state information creation unit 121 selects one candidate state indicating "Are you tired?" from two candidate states, for example, "Are you tired?" and "Are you not tired?", and deletes the remaining candidate state indicating "Are you not tired?". In this way, the state information creation unit 121 makes adjustments so that multiple candidate states with the same meaning do not exist. Then, the state information creation unit 121 creates the state information 112 shown in FIG. 2 by combining the combinations of candidate states and variables obtained as a result of the above-mentioned processing.

Next, a method for creating a topic module set will be described with reference to Figs. 8 to 19. The topic module set creation unit 120 receives one or more scenarios SC created as described above. The topic module set creation unit 120 creates each topic constituting the one or more input scenarios SC as a topic module. In this case, the topic module set creation unit 120 may unify the terms for topics that have the same meaning and create them as a topic module. For example, "Are you tired?", which is one of the topics constituting scenario SC2, and "Are you tired?", which is one of the topics constituting scenario SC3, have the same meaning. Therefore, the topic module set creation unit 120 unifies the terms to either "Are you tired?" or "Are you tired?" and creates a topic module. Fig. 8 shows a case where "Are you tired?", which is one of the topics constituting SC3, is changed to "Are you tired?" and created as a topic module.

Next, the topic module set creation unit 120 sets activation conditions for each topic module based on topic connections. The topic module set creation unit 120 performs three procedures (activation condition setting 1 to setting 3) as a procedure for setting activation conditions. First, as activation condition setting 1, the topic module set creation unit 120 associates an activation condition (IF: variable corresponding to topic == Unknown) with each topic module as shown in Figures 9 and 10. The purpose of activation condition setting 1 is to prevent the user from repeating a topic that has already been answered.

For example, in order to avoid asking "Are you hungry?" again after asking "Are you hungry?", the topic module set creation unit 120 sets IF: hungry == Unknown for the topic "Are you hungry?" as a start-up condition for outputting content corresponding to the topic "Are you hungry?". This means that the module is started only when it is not specified whether the user is hungry or not. For example, if the user is once asked whether he is hungry or not, the user's state changes according to the user's answer (the variable associated with the candidate state changes), so the condition IF: hungry == Unknown is no longer satisfied. As a result, the topic "Are you hungry?" is not selected. This achieves the purpose of setting the start-up condition 1, which is to not repeat a topic that the user has already answered. Here, the topic "Are you hungry?" has been used as an example, but start-up conditions are set for other topic modules in the same way as shown in Figures 9 and 10 to prevent the user from repeating a topic that the user has already answered.

Next, as shown in Figs. 11 to 13, the topic module set creation unit 120 adds a start condition (IF: future topic of each topic == Unknown) related to the topic toward the goal to be achieved, based on each topic constituting each scenario SC, as the start condition setting 2. For example, in the case of scenario SC1, where the topics are set in the order of "Are you hungry?" ⇒ "Want to eat ramen?" ⇒ "Can I recommend a ramen shop?" ⇒ "Recommend a ramen shop", the topics assumed to be used to reach the goal to be achieved ("Recommend a ramen shop") based on "Are you hungry?" are "Want to eat ramen?", "Can I recommend a ramen shop?", and "Recommend a ramen shop". Therefore, the topic module set creation unit 120 sets additional conditions of "Want to eat ramen == Unknown", "Can I recommend a ramen shop == Unknown", and "Have I heard a ramen shop recommended == Unknown" to the start condition associated with the topic of "Are you hungry?".

Similarly, using "Do you want to eat ramen?" as a criterion, the topics expected to lead to the goal to be achieved ("recommend a ramen shop") are "Can I introduce a ramen shop?" and "Recommend a ramen shop." Therefore, the topic module set creation unit 120 sets additional conditions of "Can I introduce a ramen shop?" == Unknown and "Have I heard a ramen shop recommendation?" == Unknown to the start-up conditions associated with the topic of "Do you want to eat ramen?". Similarly, using "Can I introduce a ramen shop?" as a criterion, the topic expected to lead to the goal to be achieved ("Recommend a ramen shop") is "Recommend a ramen shop." Therefore, the topic module set creation unit 120 sets additional conditions of "Have I heard a ramen shop recommendation?" == Unknown to the start-up conditions associated with the topic of "Can I introduce a ramen shop?".

The above explanation is about scenario SC1, but the topic module set creation unit 120 performs similar processing for other scenarios SC (for example, scenario SC2 and scenario SC3). As a result, additional activation conditions are set as shown in Figures 11 to 13. The purpose of setting 2 is to not select a topic module in which the topic to be reached has already been reached. For example, the topic "Are you hungry?" is thought to make the user recall that the system has the next topic related to "are you hungry = YES or NO". Therefore, when responding to "Are you hungry?", if the next topic that is not related to "are you hungry" follows, the user is likely to wonder why the topic "are you hungry?" was asked. To prevent such a situation, the activation condition of setting 2 is added, and topic modules in which the topic to be reached no longer exists (has already been reached) are identified.

Next, the topic module set creation unit 120 additionally sets conditions defined in the start conditions associated with the second and subsequent topics among the topics constituting each scenario SC as start condition setting 3, as shown in Figures 14 to 16. The purpose of start condition setting 3 is to set a definition that serves as the basis for talking about the second and subsequent topics associated with the start conditions. In order to output content corresponding to the second and subsequent topics, a condition is that a positive response (for example, "yes" or "free") is obtained from the conversation partner to the previous topic.

For example, in order to output content corresponding to the second topic "Do you want to eat ramen?" in scenario SC1, a condition is that a positive response (e.g., "yes" or "it's open") is obtained from the conversation partner to the first topic "Are you hungry?" in scenario SC1. Furthermore, in order to output content corresponding to the third topic "Can I introduce a ramen shop?" in scenario SC1, a condition is that a positive response (e.g., "yes" or "it's open") is obtained from the conversation partner to the second topic "Do you want to eat ramen?" in scenario SC1. Furthermore, in order to output content corresponding to the fourth topic "Recommend a ramen shop" in scenario SC1, a condition is that a positive response (e.g., "yes" or "it's open") is obtained from the conversation partner to the third topic "Can I introduce a ramen shop?" in scenario SC1. Therefore, the topic module set creation unit 120 additionally sets the definition content related to the previous topic among the contents defined in the start conditions associated with the second and subsequent topics to "YES" (previous topic == YES).

For example, the topic module set creation unit 120 additionally sets the content of the previous topic "Are you hungry?" from "hungry == Unknown" to "hungry == Yes" in the content of the definition of the activation condition associated with the topic "Want to eat ramen?" shown in FIG. 14. This makes it possible to set the activation condition for outputting the content corresponding to the topic "Want to eat ramen?" to be satisfied when a positive response is obtained from the conversation partner to the previous topic "Are you hungry?".

Similarly, the topic module set creation unit 120 additionally sets the content of the previous topic "Want to eat ramen?" from "Want to eat ramen == Unknown" to "Want to eat ramen == Yes" in the content of the definition of the activation condition associated with the topic "Can I recommend a ramen shop?" shown in FIG. 14. This makes it possible to set the activation condition to be satisfied for outputting the content corresponding to the topic "Can I recommend a ramen shop?" when a positive response is obtained from the conversation partner for the previous topic "Want to eat ramen?".

Similarly, the topic module set creation unit 120 additionally sets the content of the previous topic "Ramen shop introduction OK?" in the content of the definition of the activation condition associated with the topic "recommending a ramen shop" shown in FIG. 14, from "Ramen shop introduction OK==Unknown" to "Ramen shop introduction OK==Yes". This makes it possible to set the activation condition to be satisfied for outputting the content corresponding to the topic "recommending a ramen shop" when a positive response is obtained from the conversation partner for the previous topic "Ramen shop introduction OK?".

The above explanation is about scenario SC1, but the topic module set creation unit 120 also performs similar processing for other scenarios SC (e.g., scenario SC2 and scenario SC3). This makes it possible to set additional content defined in the start-up conditions as shown in Figures 14 to 16. As described above, the topic module set creation unit 120 sets start-up conditions for each topic module by performing processes for start-up condition settings 1 to 3.

Next, the topic module set creation unit 120 sets an output voice string for each topic module, as shown in Figures 17 to 19. This is a string that the dialogue device 10 outputs by voice when the activation condition of the topic module is satisfied. The topic module set creation unit 120 may automatically generate an output voice string according to the context of the dialogue, or may set it using a string created by a designer or the like together with the scenario SC. When automatically generating an output voice string, a language model or a generative model may be used. When using a string created by a designer or the like together with the scenario SC, the designer creates a string in association with each topic that constitutes the scenario SC. The topic module set creation unit 120 sets the string directly as an output voice string for each topic module in association with each topic.

When the process described in FIG. 8 to FIG. 19 is completed, the topic module set creation unit 120 generates a list in which the activation conditions, topic modules, and output voice strings are associated with each other by arranging multiple topic modules in a hierarchical structure in a predetermined priority order. The topic module set creation unit 120 creates a combination of the content of the topic module and the output voice string from the generated list as output language information. The topic module set creation unit 120 stores the created output language information in the storage unit 11. Furthermore, the topic module set creation unit 120 creates a combination of the activation conditions and topic modules from the generated list as a topic module set. The topic module set creation unit 120 stores the created topic module set in the storage unit 11. In the following explanation, for simplicity, a topic module set created only from the content of scenarios SC1 and SC2 will be used. In addition, a topic module set created using scenario SC3 may be used as necessary.

[Processing of the dialogue system 100 (part 1)]
Fig. 20 is a sequence diagram (part 1) showing the flow of processing of the dialogue system 100 in the embodiment. In the explanation of the processing of Fig. 20, it is assumed that the topic modules 113-n are arranged in the hierarchical structure shown in Fig. 3. In Fig. 20, the character string on the arrow pointing from the user to the dialogue device 10 is the voice output by the dialogue device 10, and the character string on the arrow pointing from the dialogue device 10 to the user is the content of the user's utterance analyzed by the dialogue device 10. Furthermore, it is assumed that the variables of each state indicated in the state information 112 at the start of the processing of Fig. 20 are initial values (e.g., U).

The topic determination unit 126 of the dialogue device 10 determines a topic based on the state information 112 and the topic module set 113 when the detection unit 122 detects the user's behavior at the start of processing and the state update unit 125 updates the information of the state information 112. The variables of each state shown in the state information 112 at the start of processing in FIG. 20 are "U" as shown in FIG. 21 (A). The topic determination unit 126 refers to the topic module set 113 and selects a topic module 113-n that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112 and has a high priority. For example, the topic determination unit 126 selects the topic module 113-7 with the highest priority among the topic modules 113-n that include a combination of variables (for example, each candidate state is "U") associated with each candidate state in the state information 112 shown in FIG. 21 (A) and that satisfies the combination of variables defined as the start condition. The topic determination unit 126 determines the topic in the selected topic module 113-7 (e.g., "ask if you're hungry") as the topic to be output. The topic determination unit 126 outputs information about the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if you're hungry") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if you're hungry"). In the example shown in FIG. 4, the output language information 114 selects "Are you hungry?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "Are you hungry?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "Are you hungry?" is output from the speaker 40 (step S101).

The user responds to the voice output from the speaker 40 by uttering "it's free" (step S102). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user has uttered "it's free". The status update unit 125 updates the status information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the utterance content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "asking if you're hungry," and the utterance content analyzed by the analysis unit 124 is "empty." Therefore, the state update unit 125 selects "hungry" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

As shown in FIG. 21(B), the state update unit 125 updates the variable "U" associated with the state "hungry" shown in the state information 112 to "Y". When the state update unit 125 updates the state information 112, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 21(B), the variables of each state shown in the state information 112 at this point are "Y" for the state "hungry" shown in the state information 112, and "U" for other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112 and has the highest priority.

For example, the topic determination unit 126 selects the topic module 113-5 with the highest priority among the topic modules 113-n that contain a condition that satisfies the combination of variables defined as the activation condition and is a combination of variables associated with each candidate state in the state information 112 shown in FIG. 21 (B) (for example, the candidate state "hungry" is "Y" and other states are "U"). The topic determination unit 126 determines the topic in the selected topic module 113-5 (for example, "asking if the person wants to eat ramen") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "ask if you want to eat ramen") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "ask if you want to eat ramen"). In the example shown in FIG. 4, the output language information 114 selects "If you're hungry, don't you want to eat ramen?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you're hungry, don't you want to eat ramen?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "If you're hungry, don't you want to eat ramen?" is output from the speaker 40 (step S103).

The user utters "I want to eat" in response to the voice output from the speaker 40 (step S104). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user uttered "I want to eat". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "asking if you want to eat ramen", and the speech content analyzed by the analysis unit 124 is "want to eat". Then, the state update unit 125 selects "want to eat ramen" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "I want to eat ramen" shown in the state information 112 to "Y" as shown in FIG. 21 (C). When the state update unit 125 updates the state information 112, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 21 (C), the variables of each state shown in the state information 112 at this point are "Y" for the states "I'm hungry" and "I want to eat ramen" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112 and has the highest priority.

For example, the topic determination unit 126 selects the topic module 113-3 with the highest priority among the topic modules 113-n that contain a condition that satisfies the combination of variables defined as the activation condition and is a combination of variables associated with each candidate state in the state information 112 shown in FIG. 21 (C) (for example, the candidate states "I'm hungry" and "I want to eat ramen" are "Y", and other states are "U"). The topic determination unit 126 determines the topic in the selected topic module 113-3 (for example, "Ask if it's OK to introduce a ramen shop") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a ramen shop") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a ramen shop"). In the example shown in FIG. 4, the output language information 114 selects "If you want to eat ramen, can I introduce you to a ramen shop?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you want to eat ramen, can I introduce you to a ramen shop?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "If you want to eat ramen, can I introduce you to a ramen shop?" is output from the speaker 40 (step S105).

The user responds to the voice output from the speaker 40 by saying "good" (step S106). The voice spoken by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice spoken by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed through natural language processing by the analysis unit 124. As a result, it is analyzed that the user has said "good". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "Ask if it's OK to introduce a ramen shop," and the speech content analyzed by the analysis unit 124 is "Sure." Therefore, the state update unit 125 selects "Ramen shop introduction OK" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "Ramen restaurant introduction OK" shown in the state information 112 to "Y" as shown in FIG. 21 (D). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 21 (D), the variables of each state shown in the state information 112 at this point are "Y" for the states "I'm hungry", "I want to eat ramen", and "Ramen restaurant introduction OK" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 selects the topic module 113-1 with the highest priority among the topic modules 113-n that contain a condition that satisfies the combination of variables defined as the activation condition and is a combination of variables associated with each candidate state in the state information 112 shown in FIG. 21 (D) (for example, the candidate states "hungry", "want to eat ramen", and "ramen restaurant recommendation OK" are "Y", and other states are "U"). The topic determination unit 126 determines the topic in the selected topic module 113-1 (for example, "recommend a ramen restaurant") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "recommend a ramen shop") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "recommend a ramen shop"). In the example shown in FIG. 4, the output language information 114 selects "If you are OK with introducing a ramen shop, I recommend a ramen shop called XX" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you are OK with introducing a ramen shop, I recommend a ramen shop called XX" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, a voice saying "If you are OK with introducing a ramen shop, I recommend a ramen shop called XX" is output from the speaker 40 (step S107).

The user responds to the voice output from the speaker 40 by uttering "I understand" (step S108). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed through natural language processing by the analysis unit 124. As a result, it is analyzed that the user has uttered "I understand". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "recommend a ramen shop", and the speech content analyzed by the analysis unit 124 is "I understand". Then, the state update unit 125 selects "I heard a ramen shop recommended" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "I heard a ramen shop recommendation" shown in the state information 112 to "Y" as shown in FIG. 21 (E). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 21 (E), the variables of each state shown in the state information 112 at this point are "Y" for the states "I'm hungry", "I want to eat ramen", "I'm OK with a ramen shop recommendation" and "I heard a ramen shop recommendation" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112.

However, at the current time, there are no selectable topics. In this case, the dialogue device 10 ends the dialogue with the user. Note that, when the dialogue ends, the state update unit 125 initializes all variables in each state indicated by the state information 112. This makes it possible to respond when a dialogue with a new user occurs.

Figure 22 is a diagram showing an example (part 2) of a topic module set 113. In the example shown in Figure 22, all topic modules 113-n are sorted in order of distance from the goal (the number of topics to be passed through on the way to the goal). By sorting in this way, it is possible to obtain as much information as possible about the user's state.

[Processing of the dialogue system 100 (part 2)]
Fig. 23 is a sequence diagram (part 2) showing the flow of processing of the dialogue system 100 in the embodiment. In the explanation of the processing of Fig. 23, it is assumed that the topic modules 113-n are arranged in the hierarchical structure shown in Fig. 22. In Fig. 23, the character string on the arrow pointing from the user to the dialogue device 10 is the voice output by the dialogue device 10, and the character string on the arrow pointing from the dialogue device 10 to the user is the content of the user's utterance analyzed by the dialogue device 10. Furthermore, it is assumed that the variables of each state indicated in the state information 112 at the start of processing of Fig. 23 are initial values.

The topic determination unit 126 of the dialogue device 10 determines a topic based on the state information 112 and the topic module set 113 when the detection unit 122 detects the user's behavior and the state update unit 125 updates the state information 112. The variables of each state shown in the state information 112 at the start of processing in FIG. 23 are "U" as shown in FIG. 24(A). The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112 and has the highest priority. For example, the topic determination unit 126 refers to the combination of variables (e.g., each candidate state is "U") associated with each candidate state in the state information 112 shown in FIG. 24(A) and selects the topic module 113-9 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-9 (e.g., "Ask if you are tired") as the topic to be output. The topic determination unit 126 outputs information about the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if you are tired") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if you are tired"). In the example shown in FIG. 4, the output language information 114 selects "Aren't you tired?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "Aren't you tired?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "Aren't you tired?" is output from the speaker 40 (step S201).

The user utters "I'm tired" in response to the voice output from the speaker 40 (step S202). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user has uttered "I'm tired". The status update unit 125 updates the status information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "ask if you are tired", and the speech content analyzed by the analysis unit 124 is "tired". Then, the state update unit 125 selects "tired" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

As shown in FIG. 24(B), the state update unit 125 updates the variable "U" associated with the state "tired" shown in the state information 112 to "Y". When the state update unit 125 updates the state information 112, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 24(B), the variables of each state shown in the state information 112 at this point are "Y" for the state "tired" shown in the state information 112, and "U" for other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 refers to the combination of variables associated with each candidate state in the state information 112 shown in FIG. 24(B) (for example, the candidate state "tired" is "Y" and other states are "U"), and selects the topic module 113-8 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-8 (for example, "asking if you're hungry") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if you're hungry") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if you're hungry"). In the example shown in FIG. 4, the output language information 114 selects "Are you hungry?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "Are you hungry?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "Are you hungry?" is output from the speaker 40 (step S203).

The user responds to the voice output from the speaker 40 by uttering "it's free" (step S204). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user has uttered "it's free". The status update unit 125 updates the status information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the utterance content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "asking if you're hungry," and the utterance content analyzed by the analysis unit 124 is "empty." Therefore, the state update unit 125 selects "hungry" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "hungry" shown in the state information 112 to "Y" as shown in FIG. 24(C). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 24(C), the variables of each state shown in the state information 112 at this point are "Y" for the states "tired" and "hungry" shown in the state information 112, and "U" for other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112 and has the highest priority.

For example, the topic determination unit 126 refers to the combination of variables associated with each candidate state in the state information 112 shown in FIG. 24(C) (for example, the candidate states "tired" and "hungry" are "Y", and other states are "U"), and selects the topic module 113-5 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-5 (for example, "asking if the person wants to eat ramen") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "ask if you want to eat ramen") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "ask if you want to eat ramen"). In the example shown in FIG. 4, the output language information 114 selects "If you're hungry, don't you want to eat ramen?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you're hungry, don't you want to eat ramen?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "If you're hungry, don't you want to eat ramen?" is output from the speaker 40 (step S205).

The user utters "I want to eat" in response to the voice output from the speaker 40 (step S206). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user uttered "I want to eat". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "asking if you want to eat ramen", and the speech content analyzed by the analysis unit 124 is "want to eat". Then, the state update unit 125 selects "want to eat ramen" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "I want to eat ramen" shown in the state information 112 to "Y" as shown in FIG. 24 (D). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 24 (D), the variables of each state shown in the state information 112 at this point are "Y" for the states "Tired", "Hungry" and "I want to eat ramen" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 refers to the combination of variables associated with each candidate state in the state information 112 shown in FIG. 24 (D) (for example, the candidate states "tired", "hungry", and "want to eat ramen" are "Y", and other states are "U"), and selects the topic module 113-6 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-6 (for example, "asking if the person wants to eat pasta") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "ask if you want to eat pasta") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "ask if you want to eat pasta"). In the example shown in FIG. 4, the output language information 114 selects "If you're hungry, don't you want to eat pasta?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you're hungry, don't you want to eat pasta?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "If you're hungry, don't you want to eat pasta?" is output from the speaker 40 (step S207).

The user utters "I want to eat" in response to the voice output from the speaker 40 (step S208). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user uttered "I want to eat". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "asking if you want to eat pasta", and the speech content analyzed by the analysis unit 124 is "want to eat". Then, the state update unit 125 selects "want to eat pasta" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "I want to eat pasta" shown in the state information 112 to "Y" as shown in FIG. 24 (E). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 24 (E), the variables of each state shown in the state information 112 at this point are "Y" for the states "Tired", "Hungry", "I want to eat ramen", and "I want to eat pasta" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 refers to the combination of variables associated with each candidate state in the state information 112 shown in FIG. 24(E) (for example, the candidate states "Tired", "Hungry", "Want to eat ramen", and "Want to eat pasta" are "Y", and other states are "U"), and selects the topic module 113-3 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-3 (for example, "Ask if it's OK to introduce a ramen shop") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a ramen shop") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a ramen shop"). In the example shown in FIG. 4, the output language information 114 selects "If you want to eat ramen, can I introduce you to a ramen shop?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you want to eat ramen, can I introduce you to a ramen shop?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "If you want to eat ramen, can I introduce you to a ramen shop?" is output from the speaker 40 (step S209).

It is assumed that the user utters "good" in response to the voice output from the speaker 40 (step S210). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed through natural language processing by the analysis unit 124. As a result, it is analyzed that the user uttered "good". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "Ask if it's OK to introduce a ramen shop," and the speech content analyzed by the analysis unit 124 is "Sure." Therefore, the state update unit 125 selects "Ramen shop introduction OK" as the corresponding candidate state. Note that, if the topic in the topic module 113-n and each candidate state indicated in the state information 112 are associated in advance, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "Ramen restaurant introduction OK" shown in the state information 112 to "Y" as shown in FIG. 24 (F). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 24 (F), the variables of each state shown in the state information 112 at this time are "Y" for the states "Tired", "Hungry", "Want to eat ramen", "Want to eat pasta", and "Ramen restaurant introduction OK" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 refers to the combination of variables associated with each candidate state in the state information 112 shown in FIG. 24 (F) (for example, the candidate states "Tired", "Hungry", "Want to eat ramen", "Want to eat pasta", and "Ramen restaurant introduction OK" are "Y", and other states are "U"), and selects the topic module 113-4 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-4 (for example, "Ask if pasta restaurant introduction OK") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a pasta restaurant") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a pasta restaurant"). In the example shown in FIG. 4, the output language information 114 selects "If you want to eat pasta, can I introduce you to a pasta restaurant?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you want to eat pasta, can I introduce you to a pasta restaurant?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the voice "If you want to eat pasta, can I introduce you to a pasta restaurant?" is output from the speaker 40 (step S211).

The user responds to the voice output from the speaker 40 by saying "good" (step S212). The voice spoken by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice spoken by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user has said "good". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "Ask if it's OK to introduce a pasta restaurant," and the speech content analyzed by the analysis unit 124 is "Sure." Therefore, the state update unit 125 selects "Pasta restaurant introduction OK" as the corresponding candidate state. Note that, if the topic in the topic module 113-n and each candidate state indicated in the state information 112 are associated in advance, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

25(A), the state update unit 125 updates the variable "U" associated with the state "pasta restaurant introduction OK" shown in the state information 112 to "Y". When the state update unit 125 updates the state information 112, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 25(A), the variables of each state shown in the state information 112 at this time are "Y" for the states "Tired", "Hungry", "Want to eat ramen", "Want to eat pasta", "Ramen restaurant introduction OK" and "Pasta restaurant introduction OK" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 refers to the combination of variables associated with each candidate state in the state information 112 shown in FIG. 25(A) (for example, the candidate states "Tired", "Hungry", "Want to eat ramen", "Want to eat pasta", "Ramen restaurant introduction OK", and "Pasta restaurant introduction OK" are "Y", and other states are "U"), and selects the topic module 113-1 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-1 (for example, "Recommend a ramen restaurant") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "recommend a ramen shop") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "recommend a ramen shop"). In the example shown in FIG. 4, the output language information 114 selects "If you are OK with introducing a ramen shop, I recommend a ramen shop called XX" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you are OK with introducing a ramen shop, I recommend a ramen shop called XX" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, a voice saying "If you are OK with introducing a ramen shop, I recommend a ramen shop called XX" is output from the speaker 40 (step S213).

The user responds to the voice output from the speaker 40 by uttering "I understand" (step S214). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed through natural language processing by the analysis unit 124. This analyzes that the user has uttered "I understand." The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "recommend a ramen shop", and the speech content analyzed by the analysis unit 124 is "I understand". Then, the state update unit 125 selects "I heard a ramen shop recommended" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

25(B), the state update unit 125 updates the variable "U" associated with the state "ramen shop recommendation" shown in the state information 112 to "Y". When the state update unit 125 updates the state information 112, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 25(B), the variables of each state shown in the state information 112 at this time are "Y" for the states "Tired", "Hungry", "Want to eat ramen", "Want to eat pasta", "Ramen shop introduction OK", "Pasta shop introduction OK" and "Ramen shop recommendation" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 refers to combinations of variables associated with each candidate state in the state information 112 shown in FIG. 25(B) (for example, the candidate states "Tired", "Hungry", "Want to eat ramen", "Want to eat pasta", "Ramen restaurant introduction OK", "Pasta restaurant introduction OK", and "Ramen restaurant recommendation" are "Y", and other states are "U"), and selects the topic module 113-2 with the highest priority among the topic modules 113-n that include conditions that satisfy the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-2 (for example, "Recommend a pasta restaurant") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "recommend a pasta restaurant") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "recommend a pasta restaurant"). In the example shown in FIG. 4, the output language information 114 selects "If you're OK with a pasta restaurant introduction, I recommend a pasta restaurant called XX" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you're OK with a pasta restaurant introduction, I recommend a pasta restaurant called XX" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, a voice saying "If you're OK with a pasta restaurant introduction, I recommend a pasta restaurant called XX" is output from the speaker 40 (step S215).

It is assumed that the user utters "I understand" in response to the voice output from the speaker 40 (step S216). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed through natural language processing by the analysis unit 124. As a result, it is analyzed that the user uttered "I understand". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "recommend a pasta restaurant", and the speech content analyzed by the analysis unit 124 is "I understand". Then, the state update unit 125 selects "I heard a pasta restaurant recommended" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "I heard a pasta restaurant recommendation" shown in the state information 112 to "Y" as shown in FIG. 25(C). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 25(C), the variables of each state shown in the state information 112 at this point are "Y" for the states "I'm hungry", "I want to eat ramen", "I'm OK with a ramen restaurant recommendation", "I heard a ramen restaurant recommendation" and "I heard a pasta restaurant recommendation" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112.

However, at the current time, there are no selectable topics. In this case, the dialogue device 10 ends the dialogue with the user. Note that, when the dialogue ends, the state update unit 125 initializes all variables in each state indicated by the state information 112. This makes it possible to respond when a dialogue with a new user occurs.

Figure 26 is a diagram showing an example (part 3) of a topic module set 113. In the example shown in Figure 26, all topic modules 113-n are arranged in order of priority to topics related to goals that are to be prioritized. In Figure 26, the modules are arranged so that the goal "recommend a ramen shop" is given priority over the goal "recommend a pasta shop." By arranging the modules in this order, it is possible to give top priority to a goal with a high priority, regardless of how far the goal may be.

[Processing of the dialogue system 100 (part 3)]
Fig. 27 is a sequence diagram (part 3) showing the flow of processing of the dialogue system 100 in the embodiment. In the explanation of the processing of Fig. 27, it is assumed that the topic modules 113-n are arranged in the hierarchical structure shown in Fig. 26. In Fig. 27, the character string on the arrow from the user to the dialogue device 10 is the voice output by the dialogue device 10, and the character string on the arrow from the dialogue device 10 to the user is the content of the user's utterance analyzed by the dialogue device 10. Furthermore, it is assumed that the variables of each state shown in the state information 112 at the start of the processing of Fig. 27 are initial values.

The topic determination unit 126 of the dialogue device 10 determines a topic based on the state information 112 and the topic module set 113 when the detection unit 122 detects the user's behavior and the state update unit 125 updates the state information 112. The variables of each state shown in the state information 112 at the start of processing in FIG. 27 are "U" as shown in FIG. 28(A). The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112 and has the highest priority. For example, the topic determination unit 126 refers to the combination of variables (e.g., each candidate state is "U") associated with each candidate state in the state information 112 shown in FIG. 28(A) and selects the topic module 113-7 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-7 (e.g., "ask if you're hungry") as the topic to be output. The topic determination unit 126 outputs information about the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if you're hungry") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if you're hungry"). In the example shown in FIG. 4, the output language information 114 selects "Are you hungry?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "Are you hungry?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "Are you hungry?" is output from the speaker 40 (step S301).

The user utters "It's free" in response to the voice output from the speaker 40 (step S302). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user uttered "It's free". The status update unit 125 updates the status information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the utterance content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "asking if you're hungry," and the utterance content analyzed by the analysis unit 124 is "empty." Therefore, the state update unit 125 selects "hungry" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

As shown in FIG. 28(B), the state update unit 125 updates the variable "U" associated with the state "hungry" shown in the state information 112 to "Y". When the state update unit 125 updates the state information 112, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 28(B), the variables of each state shown in the state information 112 at this point are "Y" for the state "hungry" shown in the state information 112, and "U" for other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112 and has the highest priority.

For example, the topic determination unit 126 refers to the combination of variables associated with each candidate state in the state information 112 shown in FIG. 28 (B) (for example, the candidate state "hungry" is "Y" and other states are "U"), and selects the topic module 113-5 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-5 (for example, "asking if the person wants to eat ramen") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "ask if you want to eat ramen") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "ask if you want to eat ramen"). In the example shown in FIG. 4, the output language information 114 selects "If you're hungry, don't you want to eat ramen?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you're hungry, don't you want to eat ramen?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "If you're hungry, don't you want to eat ramen?" is output from the speaker 40 (step S303).

The user utters "I don't want to eat" in response to the voice output from the speaker 40 (step S304). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user has uttered "I don't want to eat". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "asking if you want to eat ramen", and the speech content analyzed by the analysis unit 124 is "don't want to eat". Then, the state update unit 125 selects "want to eat ramen" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

As shown in FIG. 28(C), the state update unit 125 updates the variable "U" associated with the state "I want to eat ramen" shown in the state information 112 to "N". When the state update unit 125 updates the state information 112, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 28(C), the variables of each state shown in the state information 112 at this point are "Y" for the state "I'm hungry" shown in the state information 112, "N" for the state "I want to eat ramen" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 refers to the combination of variables associated with each candidate state in the state information 112 shown in FIG. 28(C) (for example, the candidate state "hungry" is "Y", the state "want to eat ramen" shown in the state information 112 is "N", and other states are "U"), and selects the topic module 113-6 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-6 (for example, "ask if they want to eat pasta") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "ask if you want to eat pasta") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "ask if you want to eat pasta"). In the example shown in FIG. 4, the output language information 114 selects "If you're hungry, don't you want to eat pasta?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you're hungry, don't you want to eat pasta?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "If you're hungry, don't you want to eat pasta?" is output from the speaker 40 (step S305).

The user utters "I want to eat" in response to the voice output from the speaker 40 (step S206). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user uttered "I want to eat". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "asking if you want to eat pasta", and the speech content analyzed by the analysis unit 124 is "want to eat". Then, the state update unit 125 selects "want to eat pasta" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "I want to eat pasta" shown in the state information 112 to "Y" as shown in FIG. 28 (D). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 28 (D), the variables of each state shown in the state information 112 at this point are "Y" for the states "I'm hungry" and "I want to eat pasta" shown in the state information 112, "N" for the state "I want to eat ramen" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 refers to the combination of variables associated with each candidate state in the state information 112 shown in FIG. 28 (D) (for example, the candidate states "hungry" and "want to eat pasta" are "Y", the state "want to eat ramen" shown in the state information 112 is "N", and other states are "U"), and selects the topic module 113-4 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-4 (for example, "ask if it's OK to introduce a pasta restaurant") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a pasta restaurant") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a pasta restaurant"). In the example shown in FIG. 4, the output language information 114 selects "If you want to eat pasta, can I introduce you to a pasta restaurant?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you want to eat pasta, can I introduce you to a pasta restaurant?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the voice "If you want to eat pasta, can I introduce you to a pasta restaurant?" is output from the speaker 40 (step S307).

The user responds to the voice output from the speaker 40 by saying "good" (step S208). The voice spoken by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice spoken by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user has said "good". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "Ask if it's OK to introduce a pasta restaurant," and the speech content analyzed by the analysis unit 124 is "Sure." Therefore, the state update unit 125 selects "Pasta restaurant introduction OK" as the corresponding candidate state. Note that, if the topic in the topic module 113-n and each candidate state indicated in the state information 112 are associated in advance, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "pasta restaurant introduction OK" shown in the state information 112 to "Y" as shown in FIG. 28 (E). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 28 (E), the variables of each state shown in the state information 112 at this time are "Y" for the states "hungry", "want to eat pasta", and "pasta restaurant introduction OK" shown in the state information 112, "N" for the state "want to eat ramen" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 refers to the combination of variables associated with each candidate state in the state information 112 shown in FIG. 28(E) (for example, the candidate states "hungry", "want to eat pasta", and "pasta restaurant recommendation OK" are "Y", the state "want to eat ramen" shown in the state information 112 is "N", and other states are "U"), and selects the topic module 113-2 with the highest priority among the topic modules 113-n that include a condition that satisfies the combination of variables defined as the activation condition. The topic determination unit 126 determines the topic in the selected topic module 113-2 (for example, "recommend a pasta restaurant") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "recommend a pasta restaurant") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "recommend a pasta restaurant"). In the example shown in FIG. 4, the output language information 114 selects "If you're OK with a pasta restaurant introduction, I recommend a pasta restaurant called XX" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you're OK with a pasta restaurant introduction, I recommend a pasta restaurant called XX" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, a voice saying "If you're OK with a pasta restaurant introduction, I recommend a pasta restaurant called XX" is output from the speaker 40 (step S309).

It is assumed that the user utters "I understand" in response to the voice output from the speaker 40 (step S310). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user has uttered "I understand". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "recommend a pasta restaurant", and the speech content analyzed by the analysis unit 124 is "I understand". Then, the state update unit 125 selects "I heard a pasta restaurant recommended" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

As shown in FIG. 28 (F), the state update unit 125 updates the variable "U" associated with the state "I heard a pasta restaurant recommendation" shown in the state information 112 to "Y". When the state update unit 125 updates the state information 112, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 28 (F), the variables of each state shown in the state information 112 at this time are "Y" for the states "I'm hungry", "I want to eat pasta", "Pasta restaurant recommendation OK" and "I heard a pasta restaurant recommendation" shown in the state information 112, "N" for the state "I want to eat ramen" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112 and has the highest priority.

However, at the current time, there are no selectable topics. In this case, the dialogue device 10 ends the dialogue with the user. Note that, when the dialogue ends, the state update unit 125 initializes all variables in each state indicated by the state information 112. This makes it possible to respond when a dialogue with a new user occurs.

The above explanation shows the process when the user only gives positive responses to the output from the dialogue device 10. In a typical dialogue, it is also assumed that a negative response will be given. In the dialogue device 10 of the present invention, a natural change of topic is possible even when a negative response is given, by using a combination of topic modules that constitute a topic module set. A detailed explanation will be given below. First, a topic module set 113 as shown in FIG. 29 is used for the explanation. FIG. 29 is a diagram showing an example (part 4) of a topic module set 113 in the embodiment. The topic module set 113 is configured, for example, by arranging each topic module 113-11 to 113-18 in a hierarchical structure in a predetermined priority order as shown in FIG. 29.

In the example shown in FIG. 29, all topic modules 113-11 to 113-18 are arranged in order of distance from the goal (the number of topics passed through to reach the goal). For example, suppose that the goals to be achieved in a dialogue with a user are "recommend a ramen shop" and "recommend a massage shop", and the topics leading to the goal of "recommend a ramen shop" are "Are you hungry?" ⇒ "Want to eat ramen?" ⇒ "Can we recommend a ramen shop?" ⇒ "Recommend a ramen shop", and the topics leading to the goal of "recommend a massage shop" are "Are you tired?" ⇒ "Want to know how to relax?" ⇒ "Can we recommend a massage shop?" ⇒ "Recommend a massage shop".

In this case, the topics "Are you tired?" and "Are you hungry?" are the furthest from the goal (distance = 3), followed by "Do you want to eat ramen?" and "Want to know how to relax?" (distance = 2), and then "Can I recommend a ramen shop?" and "Can I recommend a massage shop?" which are the closest to the goal (distance = 1). Then, by arranging each topic in a predetermined order of priority (for example, in order of distance from the goal, and giving ramen priority over massage), and setting the corresponding activation conditions, the configuration shown in FIG. 29 is obtained.

[Processing of the dialogue system 100 (part 4)]
Fig. 30 is a sequence diagram (part 4) showing the flow of processing of the dialogue system 100 in the embodiment. In the explanation of the processing of Fig. 30, it is assumed that the topic modules 113-n are arranged in the hierarchical structure shown in Fig. 29. In Fig. 30, the character string on the arrow from the user to the dialogue device 10 is the voice output by the dialogue device 10, and the character string on the arrow from the dialogue device 10 to the user is the content of the user's utterance analyzed by the dialogue device 10. Furthermore, it is assumed that the variables of each state shown in the state information 112 at the start of the processing of Fig. 30 are initial values (e.g., U).

The topic determination unit 126 of the dialogue device 10 determines a topic based on the state information 112 and the topic module set 113 when the detection unit 122 detects the user's behavior at the start of processing and the state update unit 125 updates the information of the state information 112. The variables of each state shown in the state information 112 at the start of processing in FIG. 30 are "U" as shown in FIG. 31 (A). The topic determination unit 126 refers to the topic module set 113 and selects a topic module 113-n that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112 and has a high priority. For example, the topic determination unit 126 selects the topic module 113-17 with the highest priority among the topic modules 113-n that include a combination of variables (for example, each candidate state is "U") associated with each candidate state in the state information 112 shown in FIG. 31 (A) and that satisfies the combination of variables defined as the start condition. The topic determination unit 126 determines the topic in the selected topic module 113-17 (e.g., "ask if you're hungry") as the topic to be output. The topic determination unit 126 outputs information about the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if you're hungry") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if you're hungry"). In the example shown in FIG. 4, the output language information 114 selects "Are you hungry?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "Are you hungry?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "Are you hungry?" is output from the speaker 40 (step S401).

The user responds to the voice output from the speaker 40 by saying "It's not free" (step S402). The voice spoken by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice spoken by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. This analyzes that the user has said "It's not free". The status update unit 125 updates the status information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the utterance content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "asking if you're hungry," and the utterance content analyzed by the analysis unit 124 is "not hungry." Therefore, the state update unit 125 selects "hungry" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

As shown in FIG. 31(B), the state update unit 125 updates the variable "U" associated with the state "hungry" shown in the state information 112 to "N". When the state update unit 125 updates the state information 112, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 31(B), the variables of each state shown in the state information 112 at this point are "N" for the state "hungry" shown in the state information 112, and "U" for other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 selects the topic module 113-18 with the highest priority from among the topic modules 113-n that contain a condition that satisfies the combination of variables defined as the activation condition and is a combination of variables associated with each candidate state in the state information 112 shown in FIG. 31 (B) (for example, the candidate state "hungry" is "N" and other states are "U"). The topic determination unit 126 determines the topic in the selected topic module 113-18 (for example, "asking if the person is tired") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if you're tired") and the output language information 114. Specifically, the language generation unit 127 refers to the output language information 114 shown in FIG. 4 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if you're tired"). In the example shown in FIG. 4, the output language information 114 selects "Aren't you tired?" as the output speech character string. The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "Aren't you tired?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "Aren't you tired?" is output from the speaker 40 (step S403).

The user utters "I'm tired" in response to the voice output from the speaker 40 (step S404). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user has uttered "I'm tired". The status update unit 125 updates the status information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "ask if you are tired", and the speech content analyzed by the analysis unit 124 is "tired". Then, the state update unit 125 selects "tired" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "tired" shown in the state information 112 to "Y" as shown in FIG. 31 (C). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 31 (C), the variables of each state shown in the state information 112 at this point are "N" for the state "hungry" shown in the state information 112, "Y" for "tired", and "U" for other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 selects the topic module 113-16 with the highest priority from among the topic modules 113-n that contain conditions that satisfy the combination of variables defined as the activation conditions and are a combination of variables associated with each candidate state in the state information 112 shown in FIG. 31 (C) (for example, the candidate state "hungry" is "N", "tired" is "Y", and other states are "U"). The topic determination unit 126 determines the topic in the selected topic module 113-16 (for example, "Ask if you want to know how to heal") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if you want to know how to heal") and the output language information 114. Here, the output speech character string corresponding to scenario SC3 is not shown in the output language information 114 shown in FIG. 4 (this has been omitted for convenience of explanation), but the output speech character string corresponding to scenario SC3 is shown in FIG. 19. When the content corresponding to scenario SC3 is included in the topic module set as in FIG. 29, the topic module set creation unit 120 generates the output speech character string corresponding to scenario SC3 in FIG. 19 so that it is included in the output language information 114. Here, the output speech character string corresponding to scenario SC3 will be explained with reference to FIG. 19.

Specifically, the language generation unit 127 refers to the output speech character string shown in FIG. 19 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "asking if you want to know how to heal"). In the example shown in FIG. 19, "If you're tired, don't you want to know how to heal?" is selected as the output speech character string corresponding to the determined topic (e.g., "asking if you want to know how to heal"). The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you're tired, don't you want to know how to heal?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the speech "If you're tired, don't you want to know how to heal?" is output from the speaker 40 (step S405).

The user utters "I want to know" in response to the voice output from the speaker 40 (step S406). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user has uttered "I want to know". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "Ask if you want to know how to heal," and the speech content analyzed by the analysis unit 124 is "I want to know." Therefore, the state update unit 125 selects "I want to know how to heal" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "I want to know how to heal" shown in the state information 112 to "Y" as shown in FIG. 31 (D). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 31 (D), the variables of each state shown in the state information 112 at this point are "N" for the state "hungry", "Y" for "tired" and "I want to know how to heal", and "U" for other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 selects the topic module 113-14 with the highest priority from among the topic modules 113-n that contain a condition that satisfies the combination of variables defined as the activation condition and is a combination of variables associated with each candidate state in the state information 112 shown in FIG. 31 (D) (for example, the candidate state "hungry" is "N", "tired" and "want to know how to relax" are "Y", and other states are "U"). The topic determination unit 126 determines the topic in the selected topic module 113-14 (for example, "ask if it's OK to introduce a massage shop") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a massage shop") and the output language information 114. Specifically, the language generation unit 127 refers to the output speech character string shown in FIG. 19 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a massage shop"). In the example shown in FIG. 19, "If you want to know how to heal, can I introduce you to a massage shop?" is selected as the output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "Ask if it's OK to introduce a massage shop"). The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you want to know how to heal, can I introduce you to a massage shop?" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, the voice "If you want to know how to heal, can I introduce you to a massage shop?" is output from the speaker 40 (step S407).

The user responds to the voice output from the speaker 40 by saying "good" (step S408). The voice spoken by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice spoken by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed through natural language processing by the analysis unit 124. As a result, it is analyzed that the user has said "good". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "Ask if it's OK to introduce a massage shop," and the speech content analyzed by the analysis unit 124 is "Yes." Therefore, the state update unit 125 selects "Massage shop introduction OK" as the corresponding candidate state. Note that, if the topic in the topic module 113-n and each candidate state indicated in the state information 112 are associated in advance, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "Massage shop introduction OK" shown in the state information 112 to "Y" as shown in FIG. 31 (E). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 31 (E), the variables of each state shown in the state information 112 at this point are "N" for the state "Hungry", "Tired", "Want to know how to heal", and "Massage shop introduction OK" shown in the state information 112, and "U" for the other states. The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112.

For example, the topic determination unit 126 selects the topic module 113-12 with the highest priority from among the topic modules 113-n that contain a condition that satisfies the combination of variables defined as the activation condition, which is a combination of variables associated with each candidate state in the state information 112 shown in FIG. 31 (E) (for example, the candidate state "hungry" is "N", "tired", "want to know how to relax", and "massage shop introduction OK" are "Y", and other states are "U"). The topic determination unit 126 determines the topic in the selected topic module 113-12 (for example, "recommend a massage shop") as the topic to be output. The topic determination unit 126 outputs information related to the determined topic to the state update unit 125.

The language generation unit 127 generates a character string to be output as speech based on the topic determined by the topic determination unit 126 (e.g., "recommend a massage parlor") and the output language information 114. Specifically, the language generation unit 127 refers to the output speech character string shown in FIG. 19 and selects an output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "recommend a massage parlor"). In the example shown in FIG. 19, "If you are OK with a massage parlor introduction, I recommend a massage parlor called XX" is selected as the output speech character string corresponding to the topic determined by the topic determination unit 126 (e.g., "recommend a massage parlor"). The speech synthesis unit 128 generates a speech signal corresponding to the output speech character string "If you are OK with a massage parlor introduction, I recommend a massage parlor called XX" selected by the language generation unit 127, and outputs the generated speech signal via the speaker 40. As a result, a speech "If you are OK with a massage parlor introduction, I recommend a massage parlor called XX" is output from the speaker 40 (step S409).

It is assumed that the user utters "I understand" in response to the voice output from the speaker 40 (step S410). The voice uttered by the user is input to the dialogue device 10 via the microphone 30. The voice recognition unit 123 generates a character string corresponding to the voice uttered by the user through voice recognition processing. The character string generated by the voice recognition unit 123 is analyzed by the analysis unit 124 through natural language processing. As a result, it is analyzed that the user has uttered "I understand". The state update unit 125 updates the state information 112 based on the content of the utterance analyzed by the analysis unit 124.

Specifically, the state update unit 125 selects a corresponding candidate state from the candidate states indicated in the state information 112, based on the information on the topic output from the topic determination unit 126 and the speech content analyzed by the analysis unit 124. For example, the information on the topic output from the topic determination unit 126 is "recommend a massage parlor", and the speech content analyzed by the analysis unit 124 is "understand". Therefore, the state update unit 125 selects "heard a massage parlor recommended" as the corresponding candidate state. Note that, if the topic in the topic module 113-n is associated in advance with each candidate state indicated in the state information 112, the state update unit 125 may select the candidate state associated with the information on the topic output from the topic determination unit 126.

The state update unit 125 updates the variable "U" associated with the state "I heard about a massage shop recommendation" shown in the state information 112 to "Y" as shown in FIG. 31 (F). When the state information 112 is updated by the state update unit 125, the topic determination unit 126 determines the next topic to talk about based on the updated state information 112 and the topic module set 113. As shown in FIG. 31 (F), the variables of each state shown in the state information 112 at this point are "N" for the state "I'm hungry", "I want to know how to heal", "Massage shop introduction OK", and "I heard about a massage shop recommendation" shown in the state information 112, and "Y" for the other states, and "U". The topic determination unit 126 refers to the topic module set 113 and selects the topic module 113-n with the highest priority that includes a condition that is satisfied by a combination of variables associated with each candidate state in the state information 112.

However, at the current time, there is no topic that can be selected. In this case, the dialogue device 10 ends the dialogue with the user. The state update unit 125 may initialize some or all of the variables in response to an instruction from another device or functional unit. In this way, in the process shown in FIG. 30, even if a negative response is obtained from the user, the dialogue can be continued by naturally switching to another topic.

(Configuration in which a user and the dialogue device 10 have a text-based dialogue)
The above is an explanation of one embodiment of a configuration in which the dialogue device 10 determines a topic according to the content of a user's utterance and outputs the content of the determined topic by voice. Next, a configuration in which the user and the dialogue device 10 have a dialogue through text will be explained. When configured in this way, the dialogue system 100 does not need to include the microphone 30 and the speaker 40. Furthermore, the dialogue device 10 does not need to include the voice recognition unit 123 and the voice synthesis unit 128. A text output means such as a chatbot is used as a means for outputting the content related to the topic by text.

When a user and the dialogue device 10 have a text dialogue, the user inputs a character string related to the dialogue content to the dialogue device 10 via a communication device such as a smartphone owned by the user or an input device such as a keyboard connected to the dialogue device 10. The analysis unit 124 of the dialogue device 10 analyzes the content input by the user by performing natural language processing using the input character string and the dictionary 111 stored in the storage unit 11. The process from the analysis of the content by the analysis unit 124 to the generation of the character string by the language generation unit 127 is the same as the process described above, except that voice and speech are replaced with text. After that, the dialogue device 10 causes the display control unit (not shown) to display the character string generated by the language generation unit 127 on the display device 50. The display control unit is a function realized by the control unit 12.

When configured in this way, even users with hearing impairments can use the dialogue system 100. In this way, it is possible to improve the convenience of the dialogue system 100.

(Configuration in which a user conducts a dialogue by voice and the dialogue device 10 conducts a dialogue by text)
A configuration will be described in which a user conducts a dialogue by voice and the dialogue device 10 conducts a dialogue by text. When configured in this manner, the dialogue system 100 does not need to include a speaker 40. Furthermore, the dialogue device 10 does not need to include a voice synthesis unit 128. When a user conducts a dialogue by voice and the dialogue device 10 conducts a dialogue by text, the process from voice input by the microphone 30 to generation of a character string by the language generation unit 127 is the same as the process in the above-mentioned embodiment. Thereafter, the dialogue device 10 causes a display control unit (not shown) to display the character string generated by the language generation unit 127 on the display device 50. The display control unit is a function realized by the control unit 12.

When configured in this way, even users with hearing impairments can use the dialogue system 100. In this way, it is possible to improve the convenience of the dialogue system 100.

(Configuration in which a user interacts with a text and the interaction device 10 interacts with the text by voice output)
A configuration will be described in which a user has a dialogue using text and the dialogue device 10 has a dialogue using voice output. In this configuration, the dialogue system 100 does not need to include a microphone 30. Furthermore, the dialogue device 10 does not need to include a voice recognition unit 123. In the case where a user has a dialogue using text and the dialogue device 10 has a dialogue using voice output, the user inputs a character string related to the dialogue content to the dialogue device 10 via a communication device such as a smartphone held by the user or an input device such as a keyboard connected to the dialogue device 10. The analysis unit 124 of the dialogue device 10 analyzes the content input by the user by performing natural language processing using the input character string and the dictionary 111 stored in the storage unit 11. The subsequent processing is the same as the processing described in the above-mentioned embodiment.

When configured in this manner, even users who have difficulty speaking can use the dialogue system 100. In this way, it is possible to improve the convenience of the dialogue system 100.

According to the dialogue system 100 configured as above, the topic module set creation unit 120 of the dialogue device 10 creates a topic module set including multiple topic modules indicating topics to be provided to the user who is the dialogue partner when a predetermined condition is satisfied, based on one or more scenarios describing a one-way dialogue flow. The dialogue device 10 determines a topic according to the user's state based on the state information 112 and the topic module set 113, and outputs content based on the determined topic. In this way, the dialogue device 10 can refer to the state information 112 and output content based on a topic that satisfies the activation condition from among multiple topic modules 113-n. Therefore, content based on a topic according to the user's state can be output. Furthermore, since the dialogue device 10 can output content based on any topic that satisfies the activation condition, it becomes possible to continue the dialogue even when complex topic transition patterns cannot be fully anticipated.

As described above, the dialogue device 10 uses one or more scenarios SC in which a one-way dialogue flow is described. This eliminates the need for the designer to create a scenario SC that includes complex branches, and simply creates one or more scenarios SC in which a one-way dialogue flow is described. The dialogue device 10 can easily create a topic module set by using one or more scenarios SC in which a one-way dialogue flow is described.

The dialogue device 10 updates the state information 112 according to the content of the user's utterance or the content input by text. This allows the dialogue device 10 to determine a topic that takes into account the dialogue history. Therefore, the dialogue device 10 can reduce the probability of selecting a topic that is unrelated to past conversations. This makes it possible to continue the dialogue.

The topic module set is arranged in a hierarchical structure with each topic module being prioritized according to a set priority, and the dialogue device 10 determines, based on the state information, the topic with the highest priority among the topics associated with the satisfied activation conditions as the topic that corresponds to the user's state. This allows the dialogue to proceed in line with the designer's intentions.

<Modification 1>
The order of the topic modules 113-n constituting the topic module set 113 may be changed before, after, or during processing. For example, assuming that the order of the topic modules 113-n is the order of FIG. 3 at the start of processing, the dialogue device 10 may change the order of the topic modules 113-n to the order of FIG. 22 during or after the end of processing. When configured in this way, the dialogue device 10 may hold a plurality of pieces of information regarding the order of the topic modules 113-n, and change the order of the topic modules 113-n at the timing when a sorting condition is satisfied. The sorting condition may be, for example, the end of one process (for example, the process of FIG. 20, FIG. 23, FIG. 27, etc.), the arrival of a predetermined time, or an instruction to change from the outside.
Such a configuration allows for greater freedom in how the dialogue proceeds, thereby improving convenience.

<Modification 2>
In the dialogue system 100 described above, a two-dimensional agent is displayed on the display device 50, and it appears as if the agent is speaking. In contrast, instead of the display device 50, a robot may be installed near the dialogue device 10, and it may be configured so that the robot is speaking. Fig. 30 is a diagram showing an example of the configuration of a dialogue system 100a in a modified example. The dialogue system 100a includes a dialogue device 10a, a camera 20, a microphone 30, a speaker 40, and a robot 60. The camera 20, the microphone 30, the speaker 40, and the robot 60 are connected to the dialogue device 10a by wire or wirelessly.

The robot 60 performs a predetermined operation by controlling the functions provided in the robot 60, such as each drive mechanism, light emitting unit, speaker, or camera, according to the control information transmitted by the dialogue device 10a. For example, the robot 60 operates by activating a drive mechanism provided in each joint of the neck, shoulder, or arm. The robot 60 may be in the form of an animal that walks by activating a drive mechanism provided in each joint of the shoulder or leg. The robot 60 may be a bipedal robot (humanoid) that walks independently by activating a drive mechanism provided in each joint of the shoulder or leg. The robot 60 may be a mobile robot (agentized robot) that can move on wheels or tracks. The robot 60 is placed on a plate-shaped platform, such as a table or a reception desk.

The dialogue device 10a includes a memory unit 11a and a control unit 12a. The memory unit 11a stores a dictionary 111, state information 112, a topic module set 113, output language information 114, and operation control information 115a. The memory unit 11a is configured using a storage device such as a magnetic storage device or a semiconductor storage device. The operation control information 115a includes information for controlling the robot 60. For example, the operation control information 115a may be a table in which a topic or an output voice character string is associated with a control content. The control content is content for controlling the functions provided in the robot 60, such as each drive mechanism, a light-emitting unit, a speaker, or a camera.

The control unit 12a controls the entire dialogue device 10a. The control unit 12a is configured using a processor such as a CPU and a memory. By executing a program, the control unit 12a realizes the functions of the topic module set creation unit 120, the state information creation unit 121, the detection unit 122, the voice recognition unit 123, the analysis unit 124, the state update unit 125, the topic determination unit 126, the language generation unit 127, the voice synthesis unit 128, and the operation control unit 129a.

Some or all of the topic module set creation unit 120, state information creation unit 121, detection unit 122, voice recognition unit 123, analysis unit 124, state update unit 125, topic determination unit 126, language generation unit 127, voice synthesis unit 128, and operation control unit 129a may be realized by hardware (including circuitry) such as an ASIC, PLD, or FPGA, or by a combination of software and hardware. The program may be recorded on a computer-readable recording medium. A computer-readable recording medium is, for example, a non-transitory storage medium such as a portable medium such as a flexible disk, optical magnetic disk, ROM, or CD-ROM, or a storage device such as a hard disk built into a computer system. The program may be transmitted via a telecommunications line.

Some of the functions of the topic module set creation unit 120, state information creation unit 121, detection unit 122, voice recognition unit 123, analysis unit 124, state update unit 125, topic determination unit 126, language generation unit 127, voice synthesis unit 128, and operation control unit 129a do not need to be pre-installed in the dialogue device 10a, and may be realized by installing additional application programs in the dialogue device 10a.

The operation control unit 129a controls the operation of the robot 60 based on the topic or information based on the topic determined by the topic determination unit 126 and the operation control information 115a. Specifically, the operation control unit 129 refers to the operation control information 115a and acquires control content associated with the topic determined by the topic determination unit 126. The operation control unit 129a generates control information for executing the acquired control content. The operation control unit 129a controls the operation of the robot 60 by outputting the generated control information to the robot 60.

When a table in which output voice character strings and control contents are associated is used as the operation control information 115a, the operation control unit 129a refers to the output language information 114 and acquires the output voice character string associated with the topic determined by the topic determination unit 126. The operation control unit 129a refers to the operation control information 115a and acquires the control contents associated with the acquired output voice character string. The operation control unit 129a generates control information for executing the acquired control contents. The operation control unit 129a controls the operation of the robot 60 by outputting the generated control information to the robot 60.

<Modification 3>
In the above embodiment, even if information on the user's state regarding a certain topic has already been obtained, depending on the flow of processing, the content of that topic may be output as voice, resulting in an unnatural dialogue. This phenomenon may occur when a topic is skipped and the content of the previous topic is heard instead of the topic that should be confirmed. For example, assume that the user's answer to the topic "Want to know about pasta restaurants?" in the sequence has already been obtained. In the above configuration, even in such a situation, if the topic modules 113-n of "Want to eat pasta?" and "Are you hungry?", which are assumed before the topic "Want to know about pasta restaurants?", satisfy the activation conditions, those topics may be selected and output. Therefore, even if the topic development is such that an introduction of a pasta restaurant can be predicted, such as "Are you hungry?" → "Want to eat pasta?", the topic "Want to know about pasta restaurants?", which should be selected next, may not be selected, and may become a different topic sequence such as "Want to eat ramen?", which may cause the user to feel that the topic development is disappointing and unnatural.

Therefore, in the above-described embodiment, there is a method of adding, with &, as a start condition for each topic module 113-n, that the answer to the topic that appears after topic module 113-n in the sequence is not filled in. For example, topic module 113-n having the topic "Do you want to eat pasta?" adds, with &, a condition that the state variable that records the answer to the subsequent topic, "I want to know where to find pasta restaurants" = initial value. Similarly, topic module 113-n having the topic "Are you hungry?" adds, with &, a condition that the state variable that records the answer to the subsequent topic, "I want to eat pasta" = initial value, "I want to know where to find pasta restaurants" = initial value. These start conditions make it possible to prevent previous topics from being selected when a user answer to a topic that appears later in each sequence has already been obtained.

<Modification 4>
In the above-mentioned embodiment, there are cases where it is not possible to respond to a change of topic from the user. This is because the above-mentioned topic selection rule does not necessarily allow the dialogue device 10, 10a to select the next topic based on the content of the user's previous utterance. Specifically, when the user's utterance content is obtained such that a topic module 13-n with a lower priority than the currently selected topic module 13-n is activated, the dialogue device 10, 10a repeats the previous topic as it is, and therefore cannot select a topic that reflects the user's previous utterance. For example, consider a case where a topic "Want to know about pasta restaurants?" that corresponds to the last step of a certain sequence is being executed. At this time, even if the user changes the topic and makes a statement to the effect that "I also want to eat ramen", which is the basis for the topic of another sequence, the current topic "Want to know about pasta restaurants?", which has a higher priority in the hierarchical structure, takes precedence in the next turn, so that the topic change that immediately reflects the user's previous utterance "I also want to eat ramen" is not performed. As a result, the dialogue device 10, 10a behaves as if it does not accept the user's intention to change the topic, and presents a topic by prioritizing the dialogue device 10, 10a's own intention. In such a situation, it is highly likely that the user's willingness to engage in dialogue will be reduced. On the other hand, if the user's remarks are constantly taken into account, it may be impossible to achieve topic guidance at all. Pushing through the intentions of the dialogue device 10, 10a to a certain extent may emphasize the intentions and desires of the dialogue system 100, leading to an improved sense of dialogue.

The topic determination unit 126 may determine the topic to be spoken to inform the user that a variable in the state information 112 has been updated (for example, when the variable "I want to eat ramen" is updated to "Y", the topic determination unit 126 may utter "I see, so you want to eat ramen."). This allows the user to understand what the user has said, even if the dialogue device 10, 10a does not respond to a request to change the topic from the user, such as "I see, so you want to eat ramen. Um, on that note, do you want pasta?", and therefore maintains a sense of dialogue.

<Modification 5>
In the above-described embodiment, a topic that allows the user to reach the target topic as quickly as possible is always selected. However, if this continues many times, the user may sense that the dialogue system 100 has an intention to lead the user to some kind of target topic (e.g., a date invitation, advertisement). For example, in the case of advertisement, hiding the intention affects the success of the advertisement, so this behavior may reduce the success rate of the advertisement. To improve this behavior, for example, control may be considered in which the system behaves in a chatty manner on topics with low advertisement intention most of the time, and occasionally develops a topic that leads the user to the advertisement.

Two methods for implementing such a function are described below. One is to set up a topic module group and dynamically change their priority. A topic module group is a unit that groups together multiple topic modules 113-n with similar functions. For example, the hierarchy of topic modules 113-n that are composed of topics at the beginning of each sequence is defined as a chat topic module group, and the hierarchical structure of topics at the end of the sequence is defined as an advertising topic module group. Until a certain number of topics are covered from the start of the dialogue or the end of the advertising, the priority of the chat topic module group is set higher than that of the advertising topic module group. After that, when a certain number of topics are covered, the priority of the advertising topic module group is changed to a relatively high priority. This makes it possible to continue to focus on choosing topics at the beginning of each sequence, which have low advertising intent, for most of the time, and to execute a sequence that leads to the target topic at a relatively low frequency.

The other is to design topic strategy sequences aimed at small talk and set the priority of those sequences high. This makes it possible to realize behavior in which sequences aimed at small talk are executed and consumed early in the conversation, and then sequences aimed at advertising are executed afterwards.

<Modification 6>
In the above-described embodiment, while the topic-guiding dialogue is targeted, topics such as the essential contents of advertisements, such as the location of a store or the features of a product, or answers to questions from users are not targeted. Therefore, the dialogue device 10, 10a may be configured to output such topics by voice. It is desirable that answers to questions from users are always given priority over any other topics presented by the dialogue device 10, 10a. Therefore, in the dialogue device 10, 10a, an answer module set is placed higher than the above-described topic module set 113 (for example, a module set of topic-guiding dialogue). The answer topic module outputs a relevant topic (answer) when a question from a user is detected, similar to the topic module set 113. By placing the answer topic module higher than each module of the topic-guiding dialogue, the output of the answer topic module is always selected with priority over the output of the topic-guiding dialogue, so that questions from users can be answered at any time.

In the above-described embodiment, a topic for which an answer has been given is not selected, but it is assumed that there may be cases where the same topic is desired to be selected again over time or as the conversation flows. For example, questions such as "Are you hungry?" and "Are you tired?" are topics that may be desired to be asked again after a certain amount of time has passed, even if an answer has been given once. Thus, the dialogue device 10, 10a may be configured to initialize some or all of the variables indicated in the state information 112 when a predetermined amount of time has passed.

<Modification 7>
In the above-described embodiment, a configuration in which the robot or the agent displayed on the display device provides the user with content according to the topic has been shown, but other means may be used as a means for providing the user with content according to the topic. As another means, for example, a smart speaker (voice output device) may be used. When a smart speaker is used, the voice synthesis unit 128 outputs content according to the determined topic from the smart speaker. When a smart speaker is used, the dialogue system 100 does not need to include the camera 20, the speaker 40, and the display device 50, and the dialogue system 100a does not need to include the camera 20 and the speaker 40. The dialogue by the dialogue systems 100 and 100a is started when the smart speaker recognizes an utterance from the user. When configured in this way, the operation control unit 129 and 129a control the operation of the smart speaker.

In the above configuration, a configuration has been shown in which a topic is determined based on the user's speech content or the content input by text. The topic determination unit 126 may be configured to determine a topic by taking into account the user's actions in addition to the user's speech content or the content input by text. When configured in this way, a topic module is required in which the user's actions are used as activation conditions and a topic to be provided to the user when the activation conditions related to the user's actions are satisfied is associated with the activation conditions. Furthermore, information regarding the user's actions as a state is set in the state information 112.

For a topic module that uses a user's action as a start condition, the user's action against the robot, such as "a hand is reaching out towards the robot" or "trying to unplug the robot", is set as the start condition. Furthermore, as a topic to be discussed when the start condition is met, a topic for the robot to avoid danger in response to the user's action, such as "be careful". As an output speech string corresponding to the topic "be careful", "don't touch me" or "don't unplug me" is set in the output language information 114.

By placing the topic module for crisis avoidance for the robot created in this way as the top priority in the topic module set 113, when the detection unit 122 detects an action of a hand reaching out towards the robot 60, the state update unit 125 updates the state information 112 based on the user's action detected by the detection unit 122. The topic determination unit 126 refers to the topic module set 113 and determines the topic to be "Caution", which is associated with the activation condition for which "A hand is reaching out towards the robot" is "Y". The language generation unit 127 then outputs the content "Don't touch me", which is associated with the determined topic "Caution", as voice or text.

Similarly, when the detection unit 122 detects an action of trying to unplug the robot 60, the state update unit 125 updates the state information 112 based on the user's action detected by the detection unit 122. The topic determination unit 126 refers to the topic module set 113 and determines the topic to be "Caution", which is associated with the activation condition for which "trying to unplug the robot" is "Y". The language generation unit 127 then outputs the content "Don't unplug" which is associated with the determined topic "Caution" by voice or text. The user's action is detected by the detection unit 122 as described above.

As another example, the following configuration may be used as a topic module with a user's action as a start condition. For example, a user's action toward the robot, such as "waving at the robot 60 or the agent displayed on the display device 50" or "moving close to the robot 60 or the agent displayed on the display device 50" may be set as a start condition, and a topic module may be used in which a topic regarding a user's action, such as "controlling the action of either the robot, the agent, or the audio output device", is set as a topic when the start condition is satisfied. Examples of action control according to the topic "controlling the action of either the robot, the agent, or the audio output device" include "waving back" and "tilting the head". Note that the action control according to the topic is included in the action control information 115, 115a.

By arranging the topic module created in this way in the topic module set 113, when the user waves at the robot 60 or the agent displayed on the display device 50, the topic determination unit 126 refers to the topic module set 113, and when the detection unit 122 detects the action of waving at the robot 60 or the agent displayed on the display device 50, the state update unit 125 updates the state information 112 based on the user's action detected by the detection unit 122. The topic determination unit 126 refers to the topic module set 113 and determines, as the topic, "control the action of either the robot, the agent, or the audio output device" associated with the activation condition in which "the user waves at the robot 60 or the agent displayed on the display device 50" is "Y". Then, the action control unit 129, 129a refers to the action control information 115, 115a and controls the robot 60 or the agent to perform an action based on the content of "wave back" associated with the determined topic "control the action of either the robot, the agent, or the audio output device".

Similarly, when the detection unit 122 detects that the user has brought their face closer to the robot 60 or the agent displayed on the display device 50, the state update unit 125 updates the state information 112 based on the user's action detected by the detection unit 122. The topic determination unit 126 refers to the topic module set 113 and determines, as the topic, "control the action of either the robot, the agent, or the audio output device" associated with the activation condition for which "the user has brought their face closer to the robot 60 or the agent displayed on the display device 50" is "Y". Then, the action control unit 129, 129a refers to the action control information 115, 115a and controls the robot 60 or the agent to perform an action based on the content of "tilt head" associated with the determined topic "control the action of either the robot, the agent, or the audio output device".

By configuring it in this way, it is also possible to determine the topic based on the user's actions. This makes it possible to have conversations that suit a variety of situations.

<Modification 8>
In the above embodiment, the state information 112 is information that represents the state of the user. However, the state information 112 may include information that represents not only the state of the user but also the state of the dialogue system 100. Here, the state of the dialogue system 100 represents a state according to the operation performed by the dialogue system 100. For example, examples of the state of the dialogue system 100 include "uttered XX times" and "performed △△ operation YY times". In actual operation, it is assumed that the dialogue device 10 will utter the same content multiple times to one user or perform the same operation multiple times. By determining the content of the utterance to the user by taking into account the state of the utterances and operations performed by the devices (e.g., the dialogue device 10, etc.) that constitute such a dialogue system 100, the range of utterances can be broadened compared to determining the content of the utterance only based on the state of the user. In this case, the storage unit 11 of the dialogue device 10 stores information that represents the state of the user and the state of the dialogue system 100 as the state information 112. Furthermore, in the topic modules that constitute the topic module set 113, activation conditions including the state of the dialogue system 100 are registered. The state update unit 125 updates variables of the user's state information according to the user's utterance content or the content input by text, and updates variables of the state information of the dialogue system 100 according to the utterance content or the action in the dialogue system 100. The topic determination unit 126 determines a topic according to the state of the user and the state of the dialogue system 100 based on the state information 112 and the topic module set 113.

The topic determination unit 126 may determine a topic according to the state of the dialogue system 100 based on the state information 112 and the topic module set 113. In the dialogue system 100a, the state information 112 may also include information representing not only the state of the user but also the state of the dialogue system 100.

<Modification 9>
In the above embodiment, an example was shown in which the start condition is configured by a combination of any two or more of the states indicated in the state information 112, but the start condition only needs to include at least one candidate state.

<Modification 10>
In the above embodiment, a configuration in which a start condition is associated with each topic module constituting the topic module set 113 has been shown, but a topic module set 113 may include a topic module to which no start condition is associated. FIG. 33 is a diagram showing an example (part 5) of a topic module set in the modified example. The topic module set 113 shown in FIG. 33 is a configuration in which a new topic module 113-20 is added to the topic module set 113 shown in FIG. 2. The topic module 113-20 has no start condition set, and is placed at the lowest priority position. Therefore, the topic module 113-20 is executed when the start condition is not satisfied for all of the topic modules 113-1 to 113-9 that have a higher priority than the topic module 113-20. The topic module 113-n to which no start condition is set may be placed anywhere.

As shown in FIG. 33, a response by AI may be set as a topic module 113-n for which no activation condition is set. For example, an artificial intelligence that automatically creates sentences according to input contents, such as ChatGPT, may be used as a response by AI. In this way, the artificial intelligence automatically creates sentences according to the dialogue contents from the user, and the voice synthesis unit 128 of the dialogue device 10 generates a voice signal corresponding to the generated character string. The voice signal generated by the voice synthesis unit 128 is then output from the speaker 40. With this configuration, it becomes possible to continue the dialogue even with contents that could not be created in the scenario.

<Modification 11>
In the above-described embodiment, the dialogue device 10 is configured to include the topic module set creation unit 120. The topic module set creation unit 120 may be implemented in a device different from the dialogue device 10. In such a configuration, the device including the topic module set creation unit 120 may be referred to as a topic module set creation device. The topic module set creation unit 120 included in the topic module set creation device creates a topic module set by the above-described method. The topic module set created by the topic module set creation unit 120 may be stored in the dialogue device 10 via a recording medium, or may be transmitted from the topic module set creation device to the dialogue device 10 by communication. The recording medium is, for example, a USB (Universal Serial Bus), an SD card, a hard disk, or the like.

By configuring in this way, the topic module set is created in a device different from the dialogue device 10, so there is no need to create the topic module set in the dialogue device 10 that actually dialogues with the user. Therefore, the processing load of the dialogue device 10 can be reduced. Furthermore, in a topic module set creation device equipped with a topic module set, a topic module set including a plurality of topic modules indicating topics to be provided to a user who is the dialogue partner when a predetermined condition is satisfied is created based on one or more scenarios in which a one-way dialogue flow is described. As a result, the designer does not need to create a scenario SC including complex branches, but only needs to create one or more scenarios SC in which a one-way dialogue flow is described. Then, the topic module set creation device can easily create a topic module set by using one or more scenarios SC in which a one-way dialogue flow is described. The topic module set is configured to be able to output the contents of any topic that satisfies the start condition. Therefore, in actual operation, even if complex topic transition patterns cannot be fully anticipated, it becomes possible to continue the dialogue.

The above describes an embodiment of the present invention in detail with reference to the drawings, but the specific configuration is not limited to this embodiment and includes designs that do not deviate from the gist of the present invention.

10, 10a...Dialogue device, 20...Camera, 30...Microphone, 40...Speaker, 50...Display device, 60...Robot, 11, 11a...Memory unit, 12, 12a...Control unit, 120...Topic module set creation unit, 121...State information creation unit, 122...Detection unit, 123...Speech recognition unit, 124...Analysis unit, 125...State update unit, 126...Topic determination unit, 127...Language generation unit, 128...Speech synthesis unit, 129, 129a...Movement control unit

Claims (12)

a topic module set creation unit that creates a topic module set including a plurality of topic modules each indicating a topic to be provided to a user who is a conversation partner when a predetermined condition is satisfied, based on one or more scenarios in which a one-way conversation flow is described;
Equipped with
At least one or more topic modules are set with an activation condition including at least one combination of a candidate state and a variable for expressing a state of the user or the system;
The topic module set creation unit creates multiple topic modules which are sets of the candidate states and the topics by matching content corresponding to the topic indicated in the topic module with each topic module, and creates the topic module set by arranging the multiple topic modules created in a hierarchical structure in a predetermined priority order, in a topic module set creation device. The topic module set creation device according to claim 1 ;
a topic determination unit that determines a topic according to the state of the user or the state of the system based on the topic module set created by the topic module set creation device and state information that represents at least the state of the user or the system;
an output unit that outputs content according to the determined topic;
An interactive device comprising: A scenario consists of multiple topics,
a state information creating unit that creates the state information by treating each topic constituting the scenario as a candidate state and associating a variable with each candidate state,
3. An interactive device according to claim 2 . the state information creating unit, when creating the state information, integrates a plurality of topics having a similar meaning among the topics constituting the scenario into one topic;
4. An interactive device according to claim 3 . A state update unit updates state information about the user in response to at least the contents of the user's speech or the contents input by text, and updates state information about the system in response to the contents of the speech or the actions in the system,
The state update unit updates the state information by updating a variable of a candidate state according to the user's utterance content or the content input by text, or the utterance content or the action in the system.
3. An interactive device according to claim 2 . the topic determination unit determines, based on the state information, a topic with a high priority among the topics associated with the satisfied activation condition as a topic corresponding to the state of the user or the system.
3. An interactive device according to claim 2 . the topic determination unit, when it is necessary to switch to a topic constituting a second scenario during a dialogue based on a first scenario, determines a next topic from among the topics constituting the second scenario based on the state information;
3. An interactive device according to claim 2 . the topic determination unit, when there is no topic module that satisfies the activation condition, selects a dialogue module for which the activation condition is not set.
3. An interactive device according to claim 2 . and an operation control unit that controls an operation of a robot connected to the device, an agent displayed on a display device, or a voice output device based on the topic or information based on the topic determined by the topic determination unit.
3. An interactive device according to claim 2 . The computer
A topic module set is created based on one or more scenarios describing a one-way dialogue flow, the topic module set including a plurality of topic modules each indicating a topic to be provided to a user who is a dialogue partner when a predetermined condition is satisfied;
At least one or more topic modules are set with an activation condition including at least one combination of a candidate state and a variable for expressing a state of the user or the system;
A topic module set creation method, which creates multiple topic modules that are sets of the candidate states and the topics by matching content corresponding to the topic shown in the topic module with each topic module, and creates the topic module set by arranging the multiple topic modules created in a hierarchical structure with a predetermined priority . The computer
determining a topic according to a state of the user or the system based on the topic module set created by the topic module set creation method according to claim 10 and state information representing at least a state of the user or the system;
and outputting content according to the determined topic. Computer,
A computer program for causing the topic module set creation device according to claim 1 to function as the dialogue device according to claim 2 .

Images