JP6723188B2 - Dialog rule selection device, dialog device, dialog rule selection method, dialog method, and dialog rule selection program - Google Patents

Description

The present invention relates to a dialogue rule selection device, a dialogue device, a dialogue rule selection method, a dialogue method, a dialogue rule selection program, and a dialogue program.

In a dialogue system, human beings interact with a computer to obtain various kinds of information and satisfy a request. In addition, there are dialog systems that not only accomplish predetermined tasks but also carry out daily conversations, which enable humans to obtain mental stability, satisfy approval, and build trust.

The type of dialogue system is described in detail in Non-Patent Document 1.

The most popular form of dialog system implementation in commercial systems is to use dialog rules. The dialogue rule is composed of a speech pattern and a response speech template. In the dialog system using this dialog rule, when the user's utterance matches the utterance pattern by, for example, matching of word strings, the system responds by using the response utterance corresponding to the utterance pattern. For example, an example of a dialogue rule described in a language for describing the dialogue rule called AIML (Artificial Intelligence Markup Language) is shown below. The asterisk “*” included in the utterance pattern is a symbol that matches any word string.

<category>
<pattern>Alcohol*Drink*ka</pattern>
<template>I like alcohol</template>
</category>

For example, when the user's utterance is "Can I drink alcohol?", the user utterance matches the description content of <pattern> indicating the utterance pattern. Therefore, the system utters "I like alcohol", which is the description content of <template> indicating a response utterance.

The AIML is described in detail in Non-Patent Document 2.

Mikio Nakano, Kazunori Komatani, Kotaro Funakoshi, Yukiko Nakano, Manabu Okumura (Supervised) Dialogue System, Corona Publishing, 2016. WALLACE, Richard S. The anatomy of ALICE. In: Parsing the Turing Test. Springer Netherlands, 2009.p. 181-210.

The creator of a dialogue rule cannot know in advance all the situations in which the dialogue rule is used. Therefore, a response rule including a content that the business operator of the dialogue system does not want the system to speak may be included in the dialogue rule. In this case, for example, utter a content that is favorable to a competitor's product, utter a content that does not match the individuality of the system assumed by the operator, or utter a content that includes slander. There is a possibility that it will be lost.

The present invention provides a dialogue rule selection device, a dialogue device, a dialogue rule selection method, a dialogue method, and a dialogue rule selection that can realize a speech response using only a dialogue rule that matches a desired condition among a plurality of dialogue rules. The purpose is to provide programs and dialogue programs.

In order to achieve the above object, the dialogue rule selection device of the present invention comprises a set of a utterance pattern and a response utterance to the utterance of the utterance pattern, genre information, and flag information indicating whether or not the response utterance is inappropriate. Of the plurality of dialogue rules as an input, based on a setting file that defines conditions for selecting the dialogue rule regarding at least one of the genre information and the flag information, the plurality of dialogue rules And a dialogue rule selection unit for selecting the dialogue rule satisfying the above condition.

The flag information includes flag information indicating whether or not a predetermined proper noun or product name is included, flag information indicating whether or not it depends on a specific character setting, and slander or NG word. At least one of flag information indicating whether or not it may be included.

In order to achieve the above object, the dialogue device of the present invention comprises a set of a utterance pattern and a response utterance to the utterance of the utterance pattern, and at least genre information and flag information indicating whether or not the response utterance is inappropriate. One of the plurality of dialogue rules is input from the plurality of dialogue rules, and based on a setting file that defines a condition for selecting the dialogue rule regarding at least one of the genre information and the flag information. A dialogue rule selection unit that selects the dialogue rule that satisfies the condition, an input unit that inputs a utterance, and one of the dialogue rules selected by the dialogue rule selection unit that corresponds to the utterance that is input by the input unit. An utterance rule matching unit that extracts the dialogue rule having a utterance pattern, and an output unit that outputs a response utterance of the dialogue rule extracted by the utterance rule matching unit are included.

The flag information includes flag information indicating whether or not a predetermined proper noun or product name is included, flag information indicating whether or not it depends on a specific character setting, and slander or NG word. At least one of flag information indicating whether or not it may be included.

In order to achieve the above-mentioned object, a dialogue rule selection method of the present invention is a dialogue rule selection method in a dialogue rule selection device including a dialogue rule selection unit, wherein the dialogue rule selection unit selects a speech pattern and a speech pattern. It consists of a set of response utterances to utterances, and inputs a plurality of dialogue rules to which at least one of genre information and flag information indicating whether or not the response utterances are inappropriate is input, and at least the genre information and the flag information. The method includes the step of selecting the interaction rule satisfying the condition from the plurality of interaction rules based on a setting file that defines a condition for selecting the interaction rule regarding one.

In order to achieve the above object, a dialogue method of the present invention is a dialogue method in a dialogue device including a dialogue rule selection unit, an input unit, an utterance rule matching unit, and an output unit, wherein the dialogue rule selection unit is a utterance. The genre information includes a pattern and a set of response utterances to the utterance of the utterance pattern, and inputs a plurality of dialogue rules provided with at least one of genre information and flag information indicating whether the response utterance is inappropriate or not. And a step of selecting the dialogue rule satisfying the condition from the plurality of dialogue rules based on a setting file defining conditions for selecting the dialogue rule regarding at least one of the flag information, and the input unit. A step of inputting an utterance, the utterance rule matching unit having the utterance pattern having the utterance pattern corresponding to the utterance input by the input unit among the dialogue rules selected by the dialogue rule selection unit. And a step in which the output unit outputs the response utterance of the dialogue rule extracted by the utterance rule matching unit.

In order to achieve the above object, the dialogue rule selection program of the present invention is a program for causing a computer to function as each unit of the dialogue rule selection device.

In order to achieve the above object, the dialogue program of the present invention is a program for causing a computer to function as each unit of the dialogue device.

According to the present invention, it is possible to realize an utterance response using only a dialogue rule that matches a desired condition among a plurality of dialogue rules.

It is a block diagram showing functional composition of a dialogue device concerning an embodiment. It is a flow chart which shows a flow of conversation rule selection processing concerning an embodiment. It is a flow chart which shows a flow of interaction processing concerning an embodiment.

Hereinafter, the present embodiment will be described with reference to the drawings.

FIG. 1 is a block diagram showing a functional configuration of a dialogue device 10 according to the present embodiment. As shown in FIG. 1, the dialogue device 10 according to the present embodiment includes a dialogue rule storage unit 12, a setting file storage unit 14, a dialogue rule selection unit 16, a dialogue rule subset storage unit 18, an input unit 20, and an utterance rule matching unit. 22 and an output unit 24.

The dialogue rule storage unit 12 stores data indicating a plurality of dialogue rules. The dialogue rule is a set of dialogue rules described in the AIML format described above. In the present embodiment, tens of thousands of dialogue rules are created and stored in the dialogue rule storage unit 12.

Each dialogue rule is described by a category tag, and the description content between the <category> tag and the </category> tag is regarded as one dialogue rule. Further, regarding the response utterance described by the template tag (utterance by the system), the genre information shown in (A) below and the response utterance shown in (B) to (D) below indicating whether or not the response utterance is inappropriate 3 At least one of the type flags (BRAND, JINKAKU, NG) is added as meta information.

(A) Genre information (genre label)
The genre label is a label that represents a genre of a conversation topic. In this embodiment, about 20 kinds of genres such as “hobby”, “TV/radio”, “sports”, “reading”, and “movie” are used.

(B) BRAND Flag The BRAND flag is a flag indicating whether or not the response utterance includes a predetermined proper noun or a product name, and is added when the response utterance relates to a portrait right, a trademark, or the like. The BRAND flag is mainly given to "personal name", "product name", "organization name", etc., but not to "place name".

(C) JINKAKU flag The JINKAKU flag is a flag indicating whether or not the response utterance depends on a specific character setting, and when the response utterance depends on the unchanged profile (name, age, gender, etc.) of the character. Granted to. The JINKAKU flag is not given when the response utterance depends on a variable profile (hobbies, preferences, habits, experiences, thoughts, knowledge, possession, etc.) of the character.

(D) NG Flag The NG flag is a flag indicating whether or not the response utterance includes slander and NG words. The NG flag is, in response utterances, abusive (abusive speech, discriminatory expression), vulgar/obscene (adult type), negative evaluation (negative evaluation), negative action (reference to negative action), negative event (Refer to negative events), Moral violations (Positive expressions that are generally regarded as undesirable), Anti-social behaviors (expressions that suggest criminal or anti-social behavior) , It is granted when it includes self-defense (statement to escape responsibility).

The utterance pattern for matching with the user utterance is described by a pattern tag, and only the NG flag is added to the utterance pattern. The NG flag added to the utterance pattern is used to recognize that the user has made an inappropriate statement when the user makes an inappropriate statement.

The following is a specific example of a dialogue rule to which meta information is added.

In the first specific example below, the genre information indicating “shopping” and the dialogue rule in which the BRAND flag is added to the response utterance are shown.

<category>
<pattern>100 yen shop *use*?</pattern>
<template genre="Shopping"Flag="BRAND"> For stationery, we use Daiso for storing ceria. Do you have a favorite 100 yen shop?</template>
</category>

In the second specific example below, the genre information indicating “others” and the dialogue rule in which the JINKAKU flag is added to the response utterance are shown.

<category>
<pattern>Friend *call*?</pattern>
<template genre="Other"Flag="JINKAKU">Please call me Aipon</template>
</category>

In the third specific example below, the dialogue rule in which the NG flag is added to the utterance pattern is shown.

<category>
<pattern Flag="NG">Fool* or </pattern>
<template genre="Other"Flag="">Don't say that</template>
</category>

The setting file storage unit 14 stores a setting file for selecting a dialogue rule regarding at least one of genre information and flag information. The setting file selects which interaction rule among the plurality of interaction rules stored in the interaction rule storage unit 12 is selected as the interaction rule used for the interaction, or which interaction rule is selected as the interaction rule used for the interaction. It is a file that describes the conditions for deciding whether or not to do it.

Specifically, the setting file is an XML (Extensible Markup Language) file in which, as shown below, it is described whether a dialogue rule is selected or a dialogue rule is not selected for the combination of the meta information described above. Implemented as.

<config>
<deny>
<genre>Sports</genre>
<genre>Movies</genre>
<flag>NG</flag>
<flag>BRAND</flag>
</deny>
</config>

According to this description, since the deny tag indicating that the interaction rule is not selected is described, the interaction rule including “sport” or “movie” as the genre label is not selected. Moreover, the dialogue rule to which the NG flag or the BRAND flag is added is not selected.

If you want to use only some conversation rules, you can describe the configuration file as follows.

<config>
<allow>
<genre>Movies</genre>
</allow>
</config>

According to this description, since the allow tag indicating selection of the conversation rule is described, only the conversation rule to which the genre label indicating "movie" is attached is selected.

In the present embodiment, the case where the setting file described only by the deny tag and the allow tag is used has been described, but the present invention is not limited to this, and a setting file in which a condition in which genre information and flag information are arbitrarily combined is described is used. Is also good. Further, in the present embodiment, the case of using the setting file described in XML has been described, but the present invention is not limited to this, and a setting file described by the IF-THEN rule shown below may be used.

IF genre="Sports" and !flag="NG" and !flag="BRAND" THEN
ALLOW
IF genre="Movie" or genre="Beauty" THEN
DENY

According to this description, when the genre information indicating “sports” is added, or when the NG flag and the BRAND flag are not added, the dialogue rule is selected and “movie” or “beauty” is indicated. When the genre information is given, the dialogue rule is not selected.

The IF-THEN rule is applied from the upper level. Further, the description "!" in the IF-THEN rule is an operator that represents negation.

The dialogue rule selection unit 16 receives a plurality of dialogue rules stored in the dialogue rule storage unit 12 as input, and based on the conditions described in the above-mentioned setting file, a plurality of dialogue rule storage units 12 are stored. From the dialog rules, select the dialog rule used for the dialog. At this time, the dialogue rule selection unit 16 extracts only the dialogue rules that satisfy the conditions described in the setting file from the plurality of input dialogue rules. Then, the dialogue rule selection unit 16 stores the extracted dialogue rules in the dialogue rule subset storage unit 18 as a subset (subset) of the dialogue rules.

The dialogue rule subset storage unit 18 stores a subset of the dialogue rules selected by the dialogue rule selection unit 16.

The input unit 20 receives a user utterance. The user utterance may be an utterance input as text by the user, or may be an utterance in which the voice of the user utterance is converted into text using voice recognition.

The utterance rule matching unit 22 matches the user utterance input by the input unit 20 with the utterance pattern of the dialogue rule included in the subset stored in the dialogue rule subset storage unit 18. As a matching method, for example, a matching method by forward matching, which is disclosed in Non-Patent Document 2 above, can be mentioned.

In the present embodiment, for each utterance pattern of the dialogue rules included in the subset stored in the dialogue rule subset storage unit 18, a word set other than the asterisk “*” described in the utterance pattern, and a user utterance. And a word set obtained by morphological decomposition are expressed by word vectors, weighted by TF-IDF (Term Frequency-Inverse Document Frequency), and then the cosine similarity between the word vectors is calculated to obtain the cosine. The response utterance corresponding to the utterance pattern having the highest similarity is output.

The weight of TF-IDF is calculated by using a large amount of dialogue data, newspaper articles and the like. The method of calculating the similarity is not limited to the method of using the cosine similarity, but a method of using a general Jaccard coefficient as the similarity between sets, such as a word set included in an utterance pattern and words included in a user utterance. Any method may be used as long as it can calculate the similarity to the set.

The output unit 24 outputs the response utterance output by the utterance rule matching unit 22 as a system utterance. As an output method, there is a method of displaying data indicating a response utterance on a display unit such as a display, transmitting data indicating a response utterance to an external device, and outputting a voice indicating the response utterance by a voice output unit. Can be mentioned.

The interactive apparatus 10 according to the present embodiment is configured by, for example, a computer apparatus including a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Only Memory) that stores various programs. Further, the computer configuring the dialogue device 10 may include a storage unit such as a hard disk drive or a non-volatile memory. In the present embodiment, the CPU reads and executes a program stored in a storage unit such as a ROM or a hard disk, whereby the above hardware resources and the program cooperate with each other to realize the above-described functions.

Next, the flow of the dialogue rule selection processing by the dialogue apparatus 10 according to the present embodiment will be described using the flowchart shown in FIG. In the present embodiment, the dialogue rule selection process is started at the timing when predetermined data for starting the execution of the dialogue rule selection process is input to the dialogue device 10, but the dialogue rule selection process is started. Not limited to this, for example, the dialogue rule selection processing may be started at the timing when the data indicating the dialogue rule is input.

In step S101, the dialogue rule selection unit 16 inputs a plurality of dialogue rules stored in the dialogue rule storage unit 12.

In step S103, the dialogue rule selection unit 16 inputs the setting file stored in the setting file storage unit 14.

In step S105, the dialogue rule selection unit 16 selects one dialogue rule from the plurality of input dialogue rules.

In step S107, the dialogue rule selection unit 16 determines whether or not the selected dialogue rule satisfies the condition described in the setting file. When it is determined in step S107 that the selected dialogue rule satisfies the condition described in the setting file (S107, Y), the process proceeds to step S109. If it is determined in step S107 that the selected dialogue rule does not satisfy the condition described in the setting file (S107, N), the process proceeds to step S111.

In step S109, the dialogue rule selection unit 16 recognizes the selected dialogue rule as a dialogue rule to be included in the subset, and proceeds to step S113.

In step S111, the dialogue rule selection unit 16 recognizes the selected dialogue rule as a dialogue rule that is not included in the subset, and proceeds to step S113.

In step S113, the dialogue rule selection unit 16 determines whether or not there is an unprocessed dialogue rule that has not been subjected to the processing of steps S107 to S111 among the plurality of input dialogue rules. When it is determined in step S113 that there is an unprocessed dialogue rule (S113, Y), the process returns to step S105, and the processes in steps S107 to S111 are performed on the unprocessed dialogue rule. If it is determined in step S113 that there is no unprocessed dialogue rule (S113, N), the process proceeds to step S115.

In step S115, the dialogue rule selection unit 16 outputs the dialogue rules recognized as the dialogue rules to be included in the subset in step S109 to the dialogue rule subset storage unit 18, and the execution of the program for the dialogue rule selection processing ends.

As described above, in the present embodiment, a plurality of dialogues each including an utterance pattern and a response utterance for the utterance of the utterance pattern, to which at least one of genre information and flag information indicating whether or not the response utterance is inappropriate is attached. A dialog rule that satisfies the conditions is selected from a plurality of dialog rules based on a setting file that has a rule as an input and that defines a condition for selecting the dialog rule regarding at least one of genre information and flag information.

Next, the flow of the dialogue processing by the dialogue device 10 according to the present embodiment will be described with reference to the flowchart shown in FIG. In the present embodiment, the dialogue process is started at the timing when the predetermined data for starting the execution of the dialogue process is input to the dialogue device 10, but the timing at which the dialogue process is started is not limited to this. For example, the dialogue process may be started at the timing when the utterance is input.

In step S201, the speech rule matching unit 22 reads the subset of the dialogue rules stored in the dialogue rule subset storage unit 18. This subset is a subset of the dialogue rules selected from the plurality of dialogue rules stored in the dialogue rule storage unit 12 and output to the dialogue rule subset storage unit 18 in the dialogue rule selection processing shown in FIG. 2 described above. ..

In step S203, the input unit 20 inputs a user utterance.

In step S205, the utterance rule matching unit 22 extracts a dialogue rule having a utterance pattern corresponding to the user utterance from among the dialogue rules included in the subset.

In step S207, the utterance rule matching unit 22 determines whether or not there are a plurality of extracted dialogue rules. When it is determined in step S207 that there are a plurality of extracted dialogue rules (S207, Y), the process proceeds to step S209. If it is determined in step S207 that there is one extracted dialog rule (S207, N), the process proceeds to step S211.

In step S209, the utterance rule matching unit 22 selects one dialog rule having the most similar user utterance and the utterance pattern included in the dialog rule from the extracted dialog rules.

In step S211, the output unit 24 outputs the extracted dialogue rule or the response utterance included in the selected dialogue rule, and ends the execution of the program for the dialogue rule selection processing.

As described above, in the present embodiment, among the dialogue rules included in the subset, the dialogue rule having the utterance pattern corresponding to the input utterance is extracted, and the response utterance of the extracted dialogue rule is output.

As a result, it is possible to realize the utterance response using only the conversation rule useful for the business operator. In particular, avoid references to product names that are important to businesses, remarks that are incompatible with the personality of the dialogue system that businesses already hold, and slandering such as abusive language that may damage the reputation of the business. It becomes possible. Then, the feasibility of the business using the dialogue system is enhanced, and the human and the computer can collaborate efficiently.

In the present embodiment, the operations of the components of the functions shown in FIG. 1 are constructed as a program and installed in a computer used as the dialogue device 10 to be executed. However, the present invention is not limited to this, and is distributed via a network. May be.

Further, the constructed program may be stored in a portable storage medium such as a hard disk or a CD-ROM and installed or distributed in a computer.

10 Dialogue Device 12 Dialogue Rule Storage Unit 14 Setting File Storage Unit 16 Dialogue Rule Selection Unit 18 Dialogue Rule Subset Storage Unit 20 Input Unit 22 Utterance Rule Matching Unit 24 Output Unit

Claims (5)

An utterance pattern and a set of response utterances to the utterance of the utterance pattern, genre information, and a plurality of dialogue rules provided with meta information including flag information indicating whether or not the response utterance is inappropriate, as input, Among the plurality of dialogue rules , the dialogue among the plurality of dialogue rules is described based on a setting file that defines a condition for selecting the dialogue rule or a condition for not selecting the dialogue rule, which is described using a combination of meta information. A meta-information given to the rule satisfies a condition for selecting the dialog rule described in the setting file, or does not satisfy the condition for not selecting the dialog rule, and selects a dialog rule selecting unit for selecting the dialog rule. seen including,
The flag information includes flag information indicating whether or not a predetermined proper noun or product name is included, flag information indicating whether or not it depends on a specific character setting, and whether or not it includes slander or NG words. A dialogue rule selection device including flag information indicating whether or not . An utterance pattern and a set of response utterances to the utterance of the utterance pattern, genre information, and a plurality of dialogue rules provided with meta information including flag information indicating whether or not the response utterance is inappropriate, as input, Among the plurality of dialogue rules , the dialogue among the plurality of dialogue rules is described based on a setting file that defines a condition for selecting the dialogue rule or a condition for not selecting the dialogue rule, which is described using a combination of meta information. A meta-information given to the rule satisfies a condition for selecting the dialog rule described in the setting file, or does not satisfy the condition for not selecting, a dialog rule selecting unit for selecting the dialog rule, and ,
An input section for inputting utterances,
Among the dialogue rules selected by the dialogue rule selection unit, an utterance rule matching unit that extracts the dialogue rule having a utterance pattern corresponding to the utterance input by the input unit,
An output unit for outputting a response utterance of the dialogue rule extracted by the utterance rule matching unit,
Only including,
The flag information includes flag information indicating whether or not a predetermined proper noun or product name is included, flag information indicating whether or not it depends on a specific character setting, and whether or not it includes slander or NG words. An interactive device including flag information indicating whether or not . A dialogue rule selection method in a dialogue rule selection device including a dialogue rule selection unit, comprising:
The dialogue rule selection unit is composed of a set of an utterance pattern and a response utterance to the utterance of the utterance pattern, and a plurality of genre information and a plurality of meta information including meta information including flag information indicating whether or not the response utterance is inappropriate. Based on a setting file that defines a condition for selecting the dialog rule or a condition for not selecting the dialog rule, which is described using a combination of the meta information, with the dialog rule as an input, Among the conversation rules, the meta information given to the conversation rule satisfies the condition for selecting the conversation rule described in the setting file, or does not satisfy the condition for not selecting the conversation rule. viewing including the selection step,
The flag information includes flag information indicating whether or not a predetermined proper noun or product name is included, flag information indicating whether or not it depends on a specific character setting, and whether or not it includes slander or NG words. A dialogue rule selection method including flag information indicating whether or not . A dialogue method in a dialogue device including a dialogue rule selection unit, an input unit, a speech rule matching unit, and an output unit,
The dialogue rule selection unit is composed of a set of an utterance pattern and a response utterance to the utterance of the utterance pattern, and a plurality of genre information and a plurality of meta information including meta information including flag information indicating whether or not the response utterance is inappropriate. Based on a setting file that defines a condition for selecting the dialog rule or a condition for not selecting the dialog rule, which is described using a combination of the meta information, with the dialog rule as an input, Among the conversation rules, the meta information given to the conversation rule satisfies the condition for selecting the conversation rule described in the setting file, or does not satisfy the condition for not selecting the conversation rule. The steps to select,
The input unit inputs a utterance,
The utterance rule matching unit extracts the dialogue rule having the utterance pattern corresponding to the utterance input by the input unit among the dialogue rules selected by the dialogue rule selection unit;
The output unit outputs a response utterance of the dialogue rule extracted by the utterance rule matching unit;
Only including,
The flag information includes flag information indicating whether or not a predetermined proper noun or product name is included, flag information indicating whether or not it depends on a specific character setting, and whether or not it includes slander or NG words. A method of interaction that includes flag information that indicates The computer according to claim 1 to function as each section of an interactive rule selection device according, or to a computer, interactive rule selection program for executing the steps of the interaction rule selection method of claim 3.