JP7560627B2 - Information processing system, information processing method, and program - Google Patents

Description

The present invention relates to an information processing system, an information processing method, and a program.

2. Description of the Related Art Conventionally, there is known a technique for recognizing a user's speech and automatically responding to an inquiry.
For example, in the automatic voice question answering device described in Patent Document 1, a voice recognition engine is used to recognize questions in the form of free speech, and the voice recognition results are corrected by discourse analysis or the like to appropriately recognize the content of the question.

JP 2003-263190 A

However, in conventional technology, even if an attempt was made to identify an answer based on an inquiry input by voice, there were cases in which an appropriate answer could not be derived due to insufficient information contained in the inquiry, etc.
Furthermore, when a freely spoken inquiry is subjected to speech recognition using a general-purpose speech recognition engine, there is a possibility that the content of the inquiry cannot be properly recognized.
Furthermore, the voice recognition process and the process of identifying an answer may take a long time, which may make it impossible to respond to an inquiry promptly.
Furthermore, since speakers making inquiries use a wide variety of expressions, the system may not be able to adequately handle such a wide variety of expressions.
If any one of these reasons or a combination of multiple reasons occurs, it becomes difficult to provide an appropriate automatic response to a voice inquiry.
It should be noted that such a problem may arise not only in the case of voice recognition of a voice inquiry, but also in the case of an inquiry by text input such as a chat.

The objective of the present invention is to provide a technology that provides a more appropriate automatic response to inquiries.

In order to achieve the above object, an information processing system according to one aspect of the present invention comprises:
An information processing system including one or more information processing devices,
a first attribute information storage means for storing first attribute information relating to an anticipated inquiry;
a second attribute information storage means for storing second attribute information relating to a response to the anticipated inquiry;
An inquiry content acquisition means for acquiring the content of an inquiry;
a response specifying means for specifying a response including an answer to the inquiry based on attribute information included in the content of the inquiry, the first attribute information, and the second attribute information;
The present invention is characterized by comprising:

In order to achieve the above object, an information processing system according to another aspect of the present invention comprises:
An information processing system including one or more information processing devices,
A speech recognition means for acquiring a speech recognition result obtained by subjecting a speech for a query to speech recognition processing;
a language processing means for specifying answer candidates matching the inquiry by language processing based on the meaning of the speech recognition result by the speech recognition means;
an attribute processing means for extracting words contained in the query from a speech recognition result by the speech recognition means, and for identifying answer candidates matching the query based on attribute information based on the words contained in the query and attribute information based on an answer to the query;
a response identifying means for identifying a response including an answer to the inquiry based on the answer candidates identified by the language processing means and the answer candidates identified by the attribute processing means;
The present invention is characterized by comprising:

In order to achieve the above object, an information processing system according to another aspect of the present invention comprises:
An information processing system including one or more information processing devices,
a search data generating means for generating search data composed of one or more attributes associated with the answer to the anticipated inquiry based on first attribute information related to the anticipated inquiry and second attribute information related to the answer to the anticipated inquiry;
An inquiry content acquisition means for acquiring the content of an inquiry;
a response specifying means for specifying a response including an answer to the content of the inquiry based on attribute information of the content of the inquiry and the search data;
The present invention is characterized by comprising:

In order to achieve the above object, an information processing system according to another aspect of the present invention comprises:
An information processing system including one or more information processing devices,
A speech recognition means for acquiring a speech recognition result obtained by subjecting a speech for a query to speech recognition processing;
a filter processing means for filtering attribute information included in a speech recognition result by said speech recognition means using a filter including, as elements, first attribute information related to an expected inquiry, second attribute information related to a response to said expected inquiry, and expanded attribute information related to said first attribute information and said second attribute information;
a response identifying means for identifying a response including a response to the utterance for inquiry based on a filtering result by the filtering means;
The present invention is characterized by comprising:

The present invention provides a technology that provides a more appropriate automatic response to inquiries.

1 is a schematic diagram showing a configuration of an information processing system 1 according to an embodiment of the present invention. FIG. 8 is a schematic diagram showing a hardware configuration of an information processing device 800 constituting the terminal device 10 or the server 20. FIG. 2 is a block diagram showing the functional configuration of a terminal device 10. FIG. 2 is a block diagram showing the functional configuration of the server 20. FIG. 13 is a schematic diagram illustrating an example of a response content table. FIG. 13 is a schematic diagram showing an example of a connector table. FIG. 13 is a schematic diagram showing an example of generating an extraction filter. FIG. 13 is a schematic diagram illustrating an example process for identifying answer candidates based on attribute relevance using an extraction filter. 1 is a schematic diagram showing an example of a specific functional configuration of an information processing system 1 according to a first embodiment; 13 is a flowchart showing the flow of an automatic response process executed by the information processing system 1. 13 is a flowchart showing the flow of an extraction filter generation process executed by the information processing system 1. FIG. 2 is a schematic diagram illustrating a concept of processing in an embodiment. FIG. 11 is a schematic diagram showing an example of a connector table in the second embodiment. FIG. 2 is a schematic diagram showing an example of a narrowing-down table 1; FIG. 2 is a schematic diagram showing an example of a narrowing-down table 2; FIG. 4 is a schematic diagram showing an example of a narrowing-down table 3. FIG. 13 is a schematic diagram showing an example of generating an extraction filter. FIG. 13 is a schematic diagram illustrating an example process for identifying answer candidates based on attribute relevance using an extraction filter. 11 is a schematic diagram showing an example of changes in narrowing-down tables 1 to 3 when an inquiry utterance is input multiple times; FIG. 11 is a flowchart showing the flow of an automatic search process executed by the information processing system 1. FIG. 13 is a block diagram showing a functional configuration of a server 20 according to a third embodiment. FIG. 13 is a schematic diagram showing an example of a specific functional configuration of an information processing system 1 according to a third embodiment. FIG. 13 is a schematic diagram illustrating an example of a response content table in the third embodiment. FIG. 13 is a schematic diagram showing an example of a first connector table. FIG. 13 is a schematic diagram showing an example of a second connector table. FIG. 13 is a schematic diagram showing an example of generating an extraction filter for a query; FIG. 13 is a schematic diagram showing an example of generating an answer extraction filter. FIG. 13 is a schematic diagram showing an example of a process for identifying answer candidates based on attribute relevance using a query extraction filter. FIG. 13 is a schematic diagram showing an example of a process for identifying answer candidates based on attribute relevance using an answer extraction filter. 13 is a flowchart showing the flow of an automatic response process executed by the information processing system 1.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
[First embodiment]
The information processing system according to the present embodiment performs a process of identifying answer candidates that match the inquiry (a process of identifying using language processing) by performing speech recognition processing on an utterance for making a query, and a process of determining the relevance between attributes represented by words included in the speech recognition result of the utterance for making a query and attributes based on an expected query and answers prepared for the expected query, and performs a process of identifying answer candidates that match the uttered query based on the relevance of the attributes (a process of identifying based on the relevance of the attributes).The information processing system according to the present embodiment then identifies a final answer to the query based on the answer candidates identified by the process of identifying using language processing and the answer candidates identified by the process of identifying based on the relevance of the attributes, and outputs a response (a dialogue including the answer) to the query.
Therefore, according to the information processing system according to the present embodiment, when an automatic response is made to a speech, it is possible to specify more appropriate response content while shortening the processing time.

Furthermore, in the present invention, the contents of the queries and answers are implemented to be specialized for a specific purpose (i.e., a specific target), and as a result, an appropriate combination of queries and answers can be prepared according to this purpose. Then, as described below, processing such as attribute classification, attribute expansion, and extraction filter generation can be performed on the prepared queries and answers before the query is spoken. Therefore, when a query is spoken, it is possible to quickly identify the answer.

In this embodiment, the item identified as the answer to the spoken inquiry (i.e., the result of the automatic response) is a person capable of providing an answer or content representing the answer itself (e.g., text, video, audio content, or a web page, etc.), and the person identified as the answer to the spoken inquiry will answer the inquiry based on their own knowledge, and the content identified as the answer will be provided to the user (the speaker who made the inquiry).
The information processing system according to this embodiment will be specifically described below.

FIG. 1 is a schematic diagram showing the configuration of an information processing system 1 according to the present embodiment.
As shown in FIG. 1, the information processing system 1 includes a plurality of terminal devices 10 and a server 20 , and the plurality of terminal devices 10 and the server 20 are configured to be able to communicate with each other via a network 30 .
An information processing system 1 shown in FIG. 1 shows an example in which an information processing system according to the present invention is applied to a business system such as a call center that accepts voice inquiries via a network, and automatically responds to the voice inquiries.

The terminal device 10 is configured as an information processing device such as a smartphone or a PC (Personal Computer). The terminal device 10 is used by a user making an inquiry, accepts voice input representing the content of the user's inquiry, and transmits the input voice data to the server 20. The terminal device 10 also outputs to the user a response to the inquiry transmitted from the server 20 (a person capable of providing a response or content representing the response itself (e.g., text, video, audio content, or a web page, etc.)).

The server 20 is composed of an information processing device such as a PC or a server computer. The server 20 receives voice data representing the inquiry content transmitted from the terminal device 10, executes voice recognition processing, and acquires text data (voice recognition result) representing the inquiry content. In this embodiment, the server 20 acquires multiple voice recognition results by executing different types of voice recognition processing in parallel. Specifically, the server 20 executes in parallel a voice recognition processing for the purpose of dictation, which converts the entire input voice data into text data, and a voice recognition processing for word extraction, which extracts specific words included in the voice data and converts them into text data. However, the server 20 may execute one voice recognition processing (e.g., a voice recognition processing for dictation) and process one voice recognition result (e.g., a word extraction processing, etc.), thereby acquiring multiple voice recognition results.

Furthermore, the server 20 identifies answer candidates using language processing based on the acquired voice recognition results. That is, the server 20 extracts the meaning of the text data obtained by voice recognition of the user's speech using natural language processing, and identifies answer candidates that are likely to match the content of the spoken inquiry from among pre-prepared answers (here, content representing the person who will provide the answer or the answer itself). Note that hereinafter, answer candidates identified using language processing will be referred to as "answer candidates by language processing" as appropriate.

The server 20 also extracts predetermined words from the inquiry utterance by speech recognition processing, and identifies answer candidates based on the relevance between the attributes represented by the words included in the utterance and the attributes of the expected inquiry and the answers prepared for the expected inquiry. That is, the server 20 expands the attributes of the expected inquiry and the answers prepared for the expected inquiry by acquiring synonyms of the words representing the attributes of the expected inquiry and the answers prepared for the expected inquiry, and determines whether the words included in the inquiry utterance match any of the expanded attributes using the expanded attributes (i.e., the words representing the original attributes and their synonyms) as a filter. The server 20 then identifies answer candidates that correspond to the expanded attributes that match the words included in the inquiry utterance. In the following, answer candidates identified based on the relevance of attributes will be referred to as "answer candidates based on the relevance of attributes".

In this manner, in this embodiment, answer candidates are identified based on the relevance between the attributes represented by the words included in the inquiry utterance and the attributes of the expected inquiry and the answers prepared for the expected inquiry. Therefore, it is more likely that an answer that matches the uttered inquiry can be identified than when a prepared answer is identified based only on the relevance between the attributes represented by the words included in the inquiry utterance and the attributes of the expected inquiry (i.e., the consistency between the content of the uttered inquiry and the content of the expected inquiry).
In addition, in this embodiment, the attributes of the anticipated inquiry and the answers prepared for the anticipated inquiry are expanded, and the expanded attributes (i.e., words expressing the original attributes and their similar words) are used as a filter to determine whether the expanded attributes match the words included in the inquiry utterance. Therefore, it is possible to expand the scope of the identified relevance to the range of contents close to the inquiry utterance and contents close to the anticipated inquiry and the answers prepared for the anticipated inquiry, and it is possible to identify an answer that matches the uttered inquiry while taking into account the intention of the inquiry in a broad sense.

Furthermore, the server 20 executes a process of identifying an answer by integrating the identified language processing answer candidates and the attribute relevance-based answer candidates to identify a final answer to the spoken inquiry. In this embodiment, if a language processing answer candidate with a certain degree of accuracy (e.g., a score of the identification result by language processing is equal to or higher than a threshold) is identified as an answer candidate by language processing, the language processing answer candidate is prioritized, and if the accuracy of the language processing answer candidate is low, the attribute relevance-based answer candidate is selected. The server 20 then transmits the identified final answer to the terminal device 10 as voice or text data.
This makes it possible to obtain highly accurate speech recognition results quickly when a query is spoken, and to identify an appropriate response.

In this embodiment, when setting attributes to an anticipated inquiry and a response prepared for the anticipated inquiry, it is possible to set a tag or a hashtag to the data of the anticipated inquiry and the response prepared for the anticipated inquiry.
In addition, in order to enable a person capable of responding to an inquiry to be identified, a telephone network can be appropriately connected to the server 20 for making calls to the person capable of responding.

[Hardware configuration]
FIG. 2 is a schematic diagram showing a hardware configuration of an information processing device 800 constituting the terminal device 10 or the server 20. As shown in FIG.
As shown in FIG. 2, the information processing device 800 includes a CPU (Central Processing Unit) 811, a ROM (Read Only Memory) 812, a RAM (Random Access Memory) 813, a bus 814, an input unit 815, an output unit 816, a storage unit 817, a communication unit 818, a drive 819, and an imaging unit 820.

The CPU 811 executes various processes according to a program recorded in the ROM 812 or a program loaded from the storage unit 817 to the RAM 813 .
The RAM 813 also stores data and the like necessary for the CPU 811 to execute various processes.

The CPU 811, ROM 812, and RAM 813 are connected to each other via a bus 814. An input unit 815, an output unit 816, a memory unit 817, a communication unit 818, and a drive 819 are connected to the bus 814.

The input unit 815 accepts input of various information to the information processing device 800. In this embodiment, the input unit 815 includes an operation input unit 815a that accepts operation input via a display screen using a mouse, a keyboard, or the like, and an audio input unit 815b that accepts audio input using a microphone or the like.
The output unit 816 is composed of a display, a speaker, etc., and outputs images and sounds.
The storage unit 817 is composed of a hard disk or a dynamic random access memory (DRAM) or the like, and stores various data managed by each server.
The communication unit 818 controls communications with other devices via the network.

Removable media 831, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is appropriately loaded into the drive 819. A program read from the removable media 831 by the drive 819 is installed in the storage unit 817 as necessary.
The imaging section 820 is configured with an imaging device including a lens, an imaging element, etc., and captures a digital image of a subject.
The above hardware configuration is the basic configuration of the information processing device 800, and it is possible to configure the information processing device 800 without some of the hardware, to include additional hardware, or to change the implementation form of the hardware. For example, the information processing device 800 can be configured without the imaging unit 820, or to include a DSP (Digital Signal Processor) for performing voice recognition processing at high speed. In addition, the information processing device 800 can be configured to include a touch panel by configuring the input unit 815 with a touch sensor and arranging it on the display of the output unit 816.

[Functional configuration]
Next, the functional configuration of the terminal device 10 will be described.
FIG. 3 is a block diagram showing the functional configuration of the terminal device 10. As shown in FIG.
As shown in FIG. 3, in the CPU 811 of the terminal device 10, an utterance receiving unit 51, an utterance data transmitting unit 52, a response receiving unit 53, and a response output unit 54 function.

The speech receiving unit 51 receives speech made as an inquiry by a user who uses the information processing system 1. The speech received by the speech receiving unit 51 is stored as voice data.
The speech data transmitting unit 52 transmits the inquiry speech (voice data) accepted by the speech accepting unit 51 to the server 20 .
The response receiving unit 53 receives a response (web page data, voice data, video data or text data) sent from the server 20 in response to the inquiry utterance sent by the utterance data sending unit 52 .
The response output unit 54 outputs a response to the inquiry received by the response receiving unit 53 by voice or screen display.

Next, the functional configuration of the server 20 will be described.
FIG. 4 is a block diagram showing the functional configuration of the server 20. As shown in FIG.
4, in a CPU 811 of the information processing device 800, an utterance data receiving unit 151, a preprocessing unit 152, a voice recognition processing unit 153, a language processing unit 154, a narrowing-down processing unit 155, and an integration processing unit 156 function. In addition, in a storage unit 817 of the information processing device 800, a voice recognition dictionary database (voice recognition dictionary DB) 171, a word dictionary database (word dictionary DB) 172, a response content database (response content DB) 173, and a connector database (connector DB) 174 are formed.
In this embodiment, the voice recognition engines used to perform voice recognition processing include one for the purpose of dictation, which converts the entire input voice data into text data, and one for the purpose of word extraction voice recognition processing, which extracts specific words contained in the voice data and converts them into text data.

The voice recognition dictionary DB171 stores data on various dictionaries and models (such as acoustic models, language models, and pronunciation dictionaries) used by each of the multiple voice recognition engines. For example, the voice recognition dictionary DB171 stores data on an acoustic model in which voice features and phoneme models are stored in association with each other, a language model in which the results of statistical processing of patterns in which character strings or word strings are used as a language are stored, and a pronunciation dictionary in which words in a language model and phonemes in an acoustic model are stored in association with each other.

The word dictionary DB 172 stores a dictionary for word extraction in which words to be extracted are registered in advance. Each word in the word dictionary DB 172 is assigned a reading (pronunciation), and when a speech recognition process for extracting words is executed, the pre-registered words are extracted based on the reading (pronunciation) information. This makes it possible to improve the accuracy and speed of speech recognition.
The response content DB173 stores response content table data in which text data representing the anticipated inquiry content, words contained in the anticipated inquiry content (words representing attributes of the inquiry content), and information identifying an answer prepared for the anticipated inquiry (here, the name of the operator handling the inquiry or the name of the content representing the answer) are associated with each other.

FIG. 5 is a schematic diagram illustrating an example of a response content table.
In FIG. 5, the cells marked with "..." indicate that some data is present or that no data is present (this also applies to the other drawings below).
As shown in FIG. 5, the response content table stores a sentence (text data) representing the expected inquiry content, an attribute (word) of the expected inquiry content, information identifying a response prepared for the expected inquiry (an operator name or a content name representing the response), and a number assigned to the response prepared for the expected inquiry, in association with each other. For example, in the second row of the response content table in FIG. 5, a sentence "I would like to register for automatic transfer" is stored as the expected inquiry content, and the words "automatic transfer" and "registration" are stored as the attributes of the inquiry content. In addition, "operator A" is stored as information identifying the response prepared for this inquiry. Note that "operator A" is assigned the number "#1". In addition, for example, in the tenth row of the response content table in FIG. 5, a sentence "I would like to open a new account" is stored as the expected inquiry content, and the words "new", "account", and "open" are stored as the attributes of the inquiry content. In addition, "opening guidance" (movie file name) is stored as information identifying the response prepared for this inquiry. The "Opening Guidance" is given the number "#101."

The connector DB 174 stores data of a connector table in which words representing attributes of the expected inquiry content are associated with words representing attributes of the answer prepared for the expected inquiry (here, the operator's specialty field or answer record, or the classification to which the content representing the answer belongs, etc.). In the following, the attributes of the expected inquiry content are appropriately referred to as "type A attributes", and the attributes of the answer prepared for the expected inquiry are appropriately referred to as "type B" attributes. In this embodiment, the connector DB 174 stores, as attributes of the expected inquiry content (type A attributes), words that are explicitly included in the expected inquiry content (explicit attributes), as well as words that are not explicitly stated (implicit attributes), such as the meaning of the expected inquiry content itself and the name of the answer. These words are matched and synonymous words are unified, and each is made into a unique word (original word or original attribute). When other tables are created, the same rules as those used when generating the connector table are used when matching and synonymous words are unified. In addition, various data generated based on the connector table (the "response specific attribute table," "response group specific attribute table," "expanded attribute table," and "extraction filter" described below) are stored in one area of RAM 813 and, as appropriate, in connector DB 174.

FIG. 6 is a schematic diagram showing an example of a connector table.
As shown in FIG. 6, the connector table stores, for each pair of expected inquiries and responses prepared for the expected inquiries, words representing the attributes of the expected inquiry (type A attributes) and words representing the attributes of the responses prepared for the expected inquiries (type B attributes) in association with each other. For example, in the second row of the connector table shown in FIG. 6, the words "automatic transfer" and "registration" are stored as attributes of the expected inquiry (type A attributes), and the words "credit card", "debit card", "cash card", "automatic withdrawal", "transfer", etc. are stored as attributes of responses prepared for the expected inquiries (type B attributes). In the connector table, the implicit attributes of type A are stored on the left side of each entry. Data in one row of the connector table indicates original attributes that are preset as attributes of a pair of expected inquiries and responses prepared for the expected inquiries.

In this embodiment, in order to specify an appropriate answer from the contents of the inquiry, "attribute expansion" and "narrowing down" answers based on the expanded attributes are performed. In order to realize "attribute expansion", the following method is used: original words (original attributes) are classified into two types: words representing attributes that can uniquely specify answers prepared for the expected inquiry (response specific attributes), and words representing attributes that can specify a group of answers prepared for the expected inquiry (response group specific attributes), and each word (attribute) is expanded to a similar word, and if necessary, all words (attributes) including the expanded word (attribute) are expanded to a similar word acquired via a network (generating an "expanded attribute table" to be described later). Then, the attributes expanded by these methods are referred to, and words included in the utterance for the inquiry are filtered to specify an appropriate answer.
By preparing the connector table, the response specific attribute and the response group specific attribute (specifically, the "response specific attribute table" and the "response group specific attribute table" described later), the extended attribute table, and the filter for extracting words (specifically, the "extraction filter" described later) before an inquiry is made, it is possible to quickly identify the answer when the inquiry is uttered. Note that data used to identify "answers prepared for expected inquiries" in the response content table, such as the response specific attribute table, the response group specific attribute table, the extended attribute table, and the filter for extracting words, are appropriately referred to as "search data".

The speech data receiving unit 151 receives an inquiry speech (voice data) transmitted from the terminal device 10 .
The pre-processing unit 152 performs pre-processing such as noise removal on the voice data (utterance for inquiry) received by the utterance data receiving unit 151 .

The speech recognition processing unit 153 executes speech recognition processing in parallel using a plurality of speech recognition engines on the speech data after the preprocessing has been executed. Each speech recognition result acquired by the speech recognition processing is output to the language processing unit 154 and the narrowing down processing unit 155. As shown in FIG. 4, the speech recognition processing unit 153 can include a first speech recognition unit 153-1 to an n-th speech recognition unit 153-n (n is an integer of 2 or more), and the first speech recognition unit 153-1 to the n-th speech recognition unit 153-n can execute the speech recognition processing using different speech recognition engines. In the first speech recognition unit 153-1 to the n-th speech recognition unit 153-n, which speech recognition engine is used to execute the speech recognition processing and how many of the first speech recognition unit 153-1 to the n-th speech recognition unit 153-n are to be included can be determined according to specific implementation conditions such as the processing purpose and information processing capacity of the server 20. However, in this embodiment, the server 20 executes in parallel a voice recognition process for the purpose of dictation, which converts the entire input voice data into text data, and a word extraction voice recognition process for extracting specific words contained in the voice data and converting them into text data.
The voice recognition processor 153 may also perform auxiliary processing such as filler removal as appropriate.

The language processing unit 154 extracts semantic content by language processing from the text data of the dictation result acquired by the voice recognition processing unit 153, and identifies a candidate answer (a person who will give the answer or content representing the answer itself) that is likely to match the spoken content from among pre-prepared answers (here, content representing the person who will give the answer or the answer itself). At this time, the language processing unit 154 scores the degree of suitability (certainty) of the candidate answers identified from the semantic content as an answer to the inquiry, and the one with the highest score can be selected as the candidate answer by language processing.

The narrowing down processing unit 155 determines whether the voice recognition result of the word extraction obtained by the voice recognition processing unit 153 (i.e., the words extracted from the utterance) matches the words included in the extraction filter, and if there is a match, narrows down the answers (content representing the person who will give the answer or the answer itself) prepared in advance based on the matching words included in the extraction filter. This allows answer candidates based on the relevance of attributes to be identified. Note that the narrowing down processing unit 155 generates the connector table, response specific attribute table, response group specific attribute table, expanded attribute table, and extraction filter used in this embodiment before the user inputs the inquiry utterance.

FIG. 7 is a schematic diagram showing an example of generating an extraction filter.
For ease of explanation, attention will be paid only to the columns in FIG. 7 in which specific data is shown, and data in columns indicated with "..." will not be taken into consideration (the same applies to FIG. 8).
As shown in FIG. 7, when generating an extraction filter, first, a connector table is referred to, words of response specific attributes are extracted, and table-format data (hereinafter referred to as a "response specific attribute table") is generated in which the words are associated with the numbers assigned to each answer. Also, the connector table is referred to, words of response group specific attributes are extracted, and table-format data (hereinafter referred to as a "response group specific attribute table") is generated in which the numbers assigned to the answers are associated with each attribute. Furthermore, the words included in the response specific attribute table and the words included in the response group specific attribute table are expanded with synonyms. For example, the words are expanded with data of a group of synonyms prepared in advance for the words, or data of words that can be used as synonyms obtained from the outside via the Internet, etc. (for example, data of words representing similar concepts obtained by referring to an ontology). Furthermore, table-format data (hereinafter referred to as an "expanded attribute table") is generated in which original words included in the response specific attribute table and the response group specific attribute table are associated with synonyms of the words. Then, an extraction filter is generated consisting of a string of words included in the expanded attribute table.

When adding a new answer candidate, the words representing the attributes of the answer candidate are matched with the words representing the existing original attributes and synonyms are unified, and the words representing the original attributes are updated. Then, if the words representing the original attributes, including the words representing the existing original attributes, are not included in the response specific attribute table, the words representing those attributes and the answer candidate are added to the response specific attribute table, and if they are already included in the response specific attribute table, the words representing those attributes and the answer candidate are deleted from the response specific attribute table and added to the "response group specific attribute table."

FIG. 8 is a schematic diagram showing an example of a process for identifying answer candidates based on attribute relevance using an extraction filter.
For example, assume that the following utterances (1) to (3) are made.
Example utterance (1): "Can I manage my assets online?"
Example utterance (2): "Can I use my fixed-term deposit for a card loan?"
Example utterance (3): "How can I check the remaining balance of my mortgage?"

At this time, it is assumed that words expressing attributes are extracted from each example utterance as follows:
Example utterance (1): "Internet" (similar word), "asset management" (original word)
Example utterance (2): "Fixed term deposit" (original word), "Card loan" (original word)
Example utterance (3): "Mortgage" (original word), "Balance" (original word)

In the case of utterance example (1), "net" and "asset management" match the words included in the extraction filter, and the original word "internet" that corresponds to the synonym "net" is included in the response specific attribute table, so the answer prepared for the anticipated inquiry can be uniquely identified. That is, in the case of utterance example (1), operator B (#2) is identified from the response specific attribute table and becomes a candidate for the answer based on the relevance of attributes.

In the case of the utterance example (2), "fixed term deposit" and "card loan" match the words included in the extraction filter, and both of the original words "fixed term deposit" and "card loan" are included in the response group specific attribute table. At this time, the answers of #3, #4, and #5 (i.e., operators B, C, and D) are identified by "fixed term deposit". Also, the answers of #2 and #4 (i.e., operators B and D) are identified by "card loan". Operator D (#4) is identified by the logical product of the group of identified answers, and is set as a candidate answer based on the relevance of attributes. In this example, the answers are narrowed down to one, but if only words included in the response group specific attribute table are extracted, multiple answers may be candidates. In such a case, a response may be made that presents multiple answer candidates to the user. For example, a response such as "Operator B and Operator D can answer" may be output, and a unique answer may be selected by the user's utterance. Also, if multiple answers are candidates, a response may be output that requests further utterances to uniquely identify the answer.

In the case of example utterance (3), "mortgage" and "balance" match the words included in the extraction filter, and the original words "mortgage" and "balance" are both included in the response group identification attribute table. In this case, "mortgage" identifies the answers in #2 and #4 (i.e., operators B and D). Also, "balance" identifies the answers in #4 and #5 (i.e., operators D and E). Operator D (#4) is identified by the logical product of the groups of identified answers, and is made a candidate answer based on the relevance of attributes.

Returning to FIG. 4, the integration processing unit 156 executes a process of integrating and identifying the identified answer candidates by language processing and answer candidates based on attribute relevance, and identifies a final answer to the spoken inquiry. At this time, as described above, if an answer candidate by language processing is identified that has a certain degree of accuracy (for example, an answer candidate whose score of the identification result by language processing is equal to or higher than a threshold value, etc.), the integration processing unit 156 prioritizes the answer candidate by language processing, and if the accuracy of the answer candidate by language processing is low, the integration processing unit 156 selects the answer candidate based on attribute relevance. Then, the integration processing unit 156 transmits the identified final answer to the terminal device 10 as voice or text data. Note that if the answer candidate by language processing does not have a certain degree of accuracy and an answer candidate based on attribute relevance is not identified, the integration processing unit 156 decides to output a response indicating that an answer cannot be identified.
The response content table, connector table, and extraction filter used in the server 20 are prepared and installed in the server 20 before accepting an utterance from a user (i.e., before the automatic response process is executed).

[Specific functional configuration example]
FIG. 9 is a schematic diagram showing an example of a specific functional configuration of the information processing system 1 in this embodiment.
As shown in FIG. 9, when the voice recognition processing unit 153 performs voice recognition processing such as dictation and word extraction, the voice recognition processing unit 153 includes a first voice recognition unit 153-1 and a second voice recognition unit 153-2. and is formed.

The first speech recognition unit 153 - 1 includes a dictation speech recognition unit 501 , a first recognition result specification unit 502 , and a first recognition result output unit 503 .
The dictation voice recognition unit 501 processes voice data using a voice recognition engine for dictation, and converts the entire input voice data into text data.
The first recognition result identification unit 502 uniquely identifies a speech recognition result from the processing result of the dictation speech recognition unit 501. For example, the first recognition result identification unit 502 identifies, from among the candidates of the speech recognition result acquired by the dictation speech recognition unit 501, the one with the highest accuracy as the speech recognition result based on a score assigned during the speech recognition process.
The first recognition result output unit 503 outputs the speech recognition result identified by the first recognition result identifying unit 502 to the language processing unit 154 .

The second speech recognition unit 153 - 2 includes a word extraction speech recognition unit 511 , a second recognition result specification unit 512 , and a second recognition result output unit 513 .
The word extraction speech recognition unit 511 processes the speech data using a speech recognition engine for performing speech recognition processing for word extraction (here, this is assumed to be rule grammar speech recognition processing or speech recognition processing using DNN (Deep Neural Network)), extracts specific words contained in the speech data, and converts them into text data.
The second recognition result identification unit 512 uniquely identifies a speech recognition result from the processing result of the word extraction speech recognition unit 511. For example, the second recognition result identification unit 512 identifies, as a speech recognition result, a word registered in the word dictionary DB 172 from among homonyms with different letters in the candidates of the speech recognition result acquired by the word extraction speech recognition unit 511.
The second recognition result output unit 513 outputs the speech recognition result identified by the second recognition result identifying unit 512 to the narrowing-down processing unit 155 .
The speech recognition processes in the first speech recognition unit 153-1 and the second speech recognition unit 153-2 may be executed as one dictation speech recognition process (for example, speech recognition process by the dictation speech recognition unit 501), and the subsequent processes may be executed in parallel as described above. In this case, the word dictionary DB 172 is used to extract words from the text data that is the result of the dictation speech recognition process.

[Action]
Next, the operation of the information processing system 1 will be described.
[Auto-Response Processing]
FIG. 10A is a flowchart showing the flow of the automatic response process executed by the information processing system 1.
The automatic response process is started in response to an operation of the server 20 instructing execution of the automatic response process.
In addition, before the automatic response process is executed (i.e., the user inputs an inquiry), the connector table, the response specific attribute table, the response group specific attribute table, the extended attribute table, and the extraction filter (search data) are generated in advance (see Figure 10B).

When the automatic response process is started, in step S1, the speech receiving unit 51 of the terminal device 10 determines whether or not a speech has been input by the user.
If no speech has been input by the user, the result of the determination in step S1 is NO, and the process of step S1 is repeated.
On the other hand, if a speech by the user has been input, a YES determination is made in step S1, and the process proceeds to step S2.

In step S2, the speech data transmitting unit 52 of the terminal device 10 transmits the voice data of the input utterance by the user to the server 20. The transmitted voice data is received by the speech data receiving unit 151 of the server 20.
In step S3, the preprocessing unit 152 of the server 20 performs preprocessing such as noise removal on the received voice data.

In step S4, the speech recognition processing unit 153 of the server 20 executes speech recognition processes (S41-1 to S43-1 and S41-2 to S43-2) in parallel on the pre-processed speech data using a plurality of speech recognition engines.
Specifically, in step S41-1, the dictation speech recognition unit 501 of the server 20 processes the speech data using a speech recognition engine for performing dictation, and converts the entire input speech data into text data.

In step S42-1, the first recognition result identification unit 502 uniquely identifies the speech recognition result from the processing result of the dictation speech recognition unit 501.
In step S 43 - 1 , the first recognition result output unit 503 outputs the speech recognition result identified by the first recognition result identifying unit 502 to the language processing unit 154 .
After step S43-1, the process proceeds to step S5.

Furthermore, in step S41-2, the word extraction speech recognition unit 511 processes the speech data using a speech recognition engine for performing word extraction speech recognition processing (here, this is rule grammar speech recognition processing or speech recognition processing using DNN), extracts specific words contained in the speech data, and converts them into text data.
In step S42-2, the second recognition result identification unit 512 uniquely identifies the speech recognition result from the processing result of the word extraction speech recognition unit 511.
In step S 43 - 2 , the second recognition result output unit 513 outputs the speech recognition result identified by the second recognition result identifying unit 512 to the narrowing-down processing unit 155 .
After step S43-2, the process proceeds to step S6.

In step S5, the language processing unit 154 performs language processing to extract semantic content from the text data of the dictation result acquired by the voice recognition processing unit 153, and identifies, from among pre-prepared answers (content representing the person who will give the answer or the answer itself), a candidate answer (a person who will give the answer or content representing the answer itself) that is likely to match the answer to the spoken content. The result of identifying the candidate answers by language processing in step S5 is output to the integration processing unit 156.

In step S6, the narrowing down processing unit 155 determines whether the voice recognition result of the word extraction acquired by the voice recognition processing unit 153 (i.e., the word extracted from the utterance) matches the word included in the extraction filter, and if they match, narrows down the answers prepared in advance based on the matching word included in the extraction filter. That is, the narrowing down processing unit 155 identifies words representing attributes in the expanded attribute table from the words included in the extraction filter, and further identifies the answer numbers (i.e., the answer numbers included in the connector table) included in the response specific attribute table and the response group specific attribute table from the words representing the identified attributes. Then, the narrowing down processing unit 155 narrows down specific answers by referring to the response content table based on the identified answer numbers. As a result, answer candidates based on the attribute relevance are identified. The identification result of answer candidates based on the attribute relevance in step S6 is output to the integration processing unit 156.
After steps S5 and S6, the process proceeds to step S7.

In step S7, the integration processing unit 156 determines whether or not answer candidates have been identified through language processing (for example, whether or not the score of the identification result through language processing is equal to or greater than a threshold).
If answer candidates have been identified through language processing, a YES determination is made in step S7, and the process proceeds to step S9.
On the other hand, if answer candidates have not been identified through language processing, a NO determination is made in step S7, and the process proceeds to step S8.
In step S8, the integration processing unit 156 executes a process of identifying answers based on the identified attribute relevance, and determines a response to the spoken inquiry (either an answer candidate based on the attribute relevance or a response indicating that no answer can be identified).

In step S9, the integration processing unit 156 outputs (sends to the terminal device 10) a response to the inquiry as voice or text data. At this time, if a candidate answer by language processing has been identified as a response to the inquiry, the integration processing unit 156 outputs the identified candidate answer by language processing, and if a candidate answer by language processing has not been identified, the integration processing unit 156 outputs a candidate answer based on attribute relevance or a response indicating that no answer can be identified. The response data sent in step S9 is received by the response receiving unit 53 of the terminal device 10, and output by the response output unit 54 as voice or image display. Note that the response to the inquiry output in step S9 may be a response that encourages speech to uniquely identify multiple answer candidates, or a response that outputs multiple answer candidates, in addition to a uniquely identified answer or a response indicating that no answer can be identified.

In step S10, the integrated processing unit 156 determines whether or not a termination condition for the automatic response process is met. The termination condition for the automatic response process can be defined as, for example, a unique response to an uttered inquiry being identified and outputted, or an operation for terminating the inquiry being performed by the user.
If the end condition for the inquiry process is not met, a NO result is obtained in step S10, and the process proceeds to step S1.
On the other hand, if the end condition for the automatic response process is met, the answer in step S10 is YES, and the automatic response process ends.

[Extraction filter generation process]
FIG. 10B is a flowchart showing the flow of the extraction filter generation process executed by the information processing system 1.
The extraction filter generation process is started in response to an operation for instructing execution of the extraction filter generation process being performed in the server 20. A series of search data is generated by the extraction filter generation process.
When the extraction filter generation process is started, in step S21, the narrowing-down processing unit 155 acquires the expected inquiry content and attributes of the answers prepared for the expected inquiry. These attributes can be acquired from words contained in the expected inquiry content and the answers prepared for the expected inquiry, or from preset attributes, etc.

In step S22, the narrowing-down processing unit 155 performs name matching and synonym unification of words expressing attributes.
In step S23, the narrowing-down processing unit 155 generates a connector table by associating words representing attributes of the anticipated inquiry content (type A attributes) with words representing attributes of the answers prepared for the anticipated inquiry (type B attributes).

In step S24, the narrowing-down processing unit 155 classifies the original words (original attributes) in the connector table into two categories: words that represent attributes that can uniquely identify a response prepared for an anticipated inquiry (response-specific attributes), and words that represent attributes that can identify a group of responses prepared for an anticipated inquiry (response-group-specific attributes).

In step S25, the narrowing-down processing unit 155 generates a response specific attribute table and a response group specific attribute table based on the classified response specific attributes and response group specific attributes.
In step S26, the narrowing-down processing unit 155 expands the words included in the response specific attribute table and the words included in the response group specific attribute table into synonyms.

In step S27, the narrowing-down processing unit 155 generates an expanded attribute table in which the original words included in the response specific attribute table and the response group specific attribute table are associated with synonyms of the original words.
In step S28, the narrowing-down processing unit 155 generates an extraction filter made up of a word string included in the expanded attribute table.
After step S28, the extraction filter generation process ends.

As described above, in the information processing system 1 of this embodiment, answer candidates are identified based on the relevance between the attributes represented by the words included in the inquiry utterance and the attributes of the expected inquiry and the answer prepared for the expected inquiry. In addition, a connector table, a response specific attribute table, a response group specific attribute table, an expanded attribute table, and an extraction filter (search data) are prepared before an inquiry is made. That is, as attributes of the expected inquiry content, words explicitly included in the expected inquiry content (explicit attributes) and words not explicitly stated (implicit attributes) such as the meaning of the expected inquiry content itself and the name of the answer are used, and a connector table is generated together with the attributes of the answer prepared for the expected inquiry. In addition, the attributes of the expected inquiry and the answer prepared for the expected inquiry are expanded to synonyms, and an extraction filter consisting of words representing the original attributes and their synonyms is generated. In addition, a predetermined word is extracted from the utterance for the inquiry by voice recognition processing, and it is determined whether the words extracted from the utterance match the words included in the extraction filter. Then, when the words extracted from the utterance match the words included in the extraction filter, the information processing system 1 narrows down the answers (the person who will answer or the content representing the answer itself) prepared in advance based on the matching words included in the extraction filter. This identifies answer candidates based on the relevance of attributes. The information processing system 1 also performs semantic content extraction using language processing on text data of the dictation result obtained by processing the utterance with voice recognition, and identifies answer candidates (the person who will answer or the content representing the answer itself) that are likely to match the answer to the spoken content from among the answers (the person who will answer or the content representing the answer itself) prepared in advance. Furthermore, the information processing system 1 performs processing to integrate and identify the identified answer candidates based on the language processing and the answer candidates based on the relevance of attributes, and identifies the final answer to the uttered inquiry.

Through this processing, the information processing system 1 can identify an answer to a spoken inquiry using not only an identification method based on language processing, but also an identification method based on the consistency of words expressing attributes.
Therefore, when an automatic response to a speech is made, it is possible to specify more appropriate response content while shortening the processing time.
In addition, it is more likely that an answer that matches the spoken inquiry can be identified than when a prepared answer is identified based only on the correlation between the attributes expressed by the words contained in the utterance of the inquiry and the attributes of the expected inquiry (i.e., the consistency between the content of the uttered inquiry and the content of the expected inquiry).
In addition, the connector table, response specific attribute table, response group specific attribute table, extended attribute table, and extraction filter (search data) used in the process of identifying an answer to an inquiry are prepared before an inquiry is made, making it possible to identify an answer quickly when an inquiry is spoken.

In addition, when determining the consistency of words expressing attributes, an extraction filter is generated by expanding words expressing attributes based on an expected inquiry and answers prepared for the expected inquiry, and a consistency with the attributes expressed by words included in the speech recognition results of the utterance for the inquiry is determined.
Therefore, even if an unexpected expression is used, such as when there is a variation in expression or when words with similar meanings are used, it is possible to easily identify an answer that matches the uttered inquiry. That is, since the scope of the identified relevance can be expanded to include contents close to the utterance of the inquiry, and contents close to anticipated inquiries and answers prepared for the anticipated inquiries, it is possible to identify an answer that matches the uttered inquiry while taking into account the broad intent of the inquiry.

[Second embodiment]
Next, a second embodiment of the present invention will be described.
In the first embodiment, the matter identified as the answer is a person capable of answering or content representing the answer itself (for example, text, video, audio content, or a web page, etc.). In addition, attributes of the anticipated inquiry and the answers prepared for the anticipated inquiry are classified into response-specific attributes that uniquely identify the answer and response group-specific attributes that identify the answers as a group, and these attributes are associated with each other and converted into data. Furthermore, by expanding the attributes, the words being spoken are filtered to identify the answer to the inquiry.

In contrast, in this embodiment, the answer of the person in charge in the first embodiment or a specific part of the entire content representing the answer (for example, a specific part of text, video, audio content, or a web page, for example, a paragraph, a section, a time period of a video, or a specific part of a web page) is also targeted for identification as an answer. In addition, in this embodiment, as an identification process based on the relevance of attributes, multiple processes having different functions (hereinafter referred to as "narrowing processes") are executed in parallel, and the identification results of these parallel processes and the identification results of the identification process using language processing are integrated to identify the content of the response (the content of the dialogue that induces the answer or the utterance of an attribute effective for identifying the answer). For example, as one of the narrowing processes, a narrowing process 1 is executed in which the answers are narrowed down by the words included in the latest utterance, targeting all the words representing the attributes included in the original connector table. In the narrowing process 1, the result of narrowing down all the words representing the attributes included in the original connector table by the latest utterance is output. As another type of narrowing down process, a narrowing down process 2 is executed, in which the results of narrowing down the words representing the attributes included in the original connector table by a predetermined number of past utterances are narrowed down by the words included in the latest utterance. In the narrowing down process 2, the results of narrowing down the words representing the attributes included in the original connector table by a predetermined number of past utterances different from the narrowing down process 2 are output. As another type of narrowing down process, a narrowing down process 3 is executed, in which the results of narrowing down the words representing the attributes included in the original connector table by a predetermined number of past utterances different from the narrowing down process 2 (for example, all past utterances since the inquiry was made) are narrowed down by the words included in the latest utterance. In the narrowing down process 3, the results of narrowing down the words representing the attributes included in the original connector table by a predetermined number of past utterances different from the narrowing down process 2 (for example, all past utterances since the inquiry was made) are output. Then, the identification results by the narrowing down processes 1 to 3 and the identification results by the identification process using language processing are integrated by selecting them based on a selection condition set in advance, and the content of the response to the inquiry (the content of the dialogue that induces the answer or the utterance of an attribute effective for identifying the answer) is identified.

In this embodiment, narrowing-down processes 1 to 3 are used in conjunction with the identification process using language processing, and further, narrowing-down processes 1 to 3 are structured to enable autonomous dialogue with the user (automatic dialogue), thereby increasing the likelihood of identifying an appropriate answer.
The main parts of the system configuration of the information processing system 1, the hardware configuration of the information processing device, and the functional configurations of the terminal device 10 and the server 20 in this embodiment are similar to those in the first embodiment, so the differences will be mainly described.

FIG. 11 is a schematic diagram showing the concept of processing in this embodiment.
As shown in FIG. 11, in the information processing system 1 according to the present embodiment, the content of the response is specified by integrating the results of the narrowing down processes 1 to 3 executed in parallel and the results of the narrowing down process using language processing. In this case, in the narrowing down processes 1 to 3, the original connector table, the response specific attribute table generated from the original connector table, and the response group specific attribute table (hereinafter referred to as "connector tables") are used as starting points to generate narrowing down tables 1, which are the results of the connector tables being narrowed down by the narrowing down process 1, narrowing down tables 2, which are the results of the narrowing down tables 1 being narrowed down by the narrowing down process 2, and narrowing down tables 3, which are the results of the narrowing down tables 2 being narrowed down by the narrowing down process 3, thereby specifying answer candidates based on the relevance of attributes. However, with regard to the narrowing down tables 3, if a series of inquiries by the user reaches four or more times, the output result of the narrowing down tables 3 becomes the input of the narrowing down process 3 for the next inquiry, and narrowing down by the narrowing down process 3 is performed recursively. The narrowing down tables 1 to 3, which are the results of the narrowing down processes 1 to 3, are included in the narrowing down tables 1 to 3, respectively. The narrowing down tables 1 to 3 are data in the form of a table showing which answer (specifically, the answer number) each word narrowed down by the narrowing down processes 1 to 3 is related to. In the narrowing down tables 1 to 3, for example, "1" indicating that the word is related to the answer number and "0" indicating that the word is not related are stored for each word narrowed down by the narrowing down processes 1 to 3. In addition, when an answer candidate cannot be uniquely identified, the attributes included in the narrowing down tables 1 to 3 are updated to attributes that can narrow down the answers more than the current ones, and then handed over as input for the next stage of the narrowing down process (the input is replaced).

In FIG. 11, the response specific attribute table included in the connector tables is called "response specific attribute table ATS0", the response group specific attribute table is called "response group specific attribute table ATG0", the response specific attribute table included in the narrowing down tables 1 is called "response specific attribute table ATS1", the response group specific attribute table is called "response group specific attribute table ATG1", the response specific attribute table included in the narrowing down tables 2 is called "response specific attribute table ATS2", the response group specific attribute table is called "response group specific attribute table ATG2", the response specific attribute table included in the narrowing down tables 3 is called "response specific attribute table ATS3", and the response group specific attribute table is called "response group specific attribute table ATG3". Similar notations will be used as appropriate in the following explanation.

[Functional configuration]
In this embodiment, the functional configuration of the terminal device 10 is similar to the functional configuration of the first embodiment shown in FIG.
Furthermore, the functional configuration of the server 20 differs from the functional configuration of the first embodiment shown in FIG. 4 in the contents stored in the connector DB 174 and the contents of the processes executed by the narrowing-down processing unit 155 and the integration processing unit 156.
Connector DB174 stores data of a connector table in which words representing attributes of anticipated inquiry content are associated with words representing attributes of answers prepared for the anticipated inquiries (here, answers from personnel capable of providing an answer, or attributes of the entire content representing the answer and parts of the content representing the answer).

FIG. 12A is a schematic diagram showing an example of a connector table in this embodiment.
As shown in FIG. 12A, in the connector table, as in the first embodiment, for each pair of anticipated inquiries and answers prepared for the anticipated inquiries, words representing the attributes of the anticipated inquiry content (type A attributes) and words representing the attributes of the answers prepared for the anticipated inquiries (type B attributes) are stored in association with each other.

In this embodiment, connector DB174 stores, as attributes of the expected inquiry content (type A attributes), not only words explicitly included in the expected inquiry content (explicit attributes), but also words that are not explicitly stated (implicit attributes), such as the meaning of the expected inquiry content itself and the name of the answer. These words are matched and synonymized, and each is made into a unique word (original word, original attribute).

In this embodiment, a category to which the entire content representing the answer belongs and a category to which the part constituting the content representing the answer belongs are set as attributes of the content representing the answer (attributes of type B), and words representing attributes for the entire content and the part are stored in the connector table. The part constituting the content representing the answer can be any part of the entire content representing the answer that can be divided and set with attributes, and can be, for example, a part divided by paragraph, section, or time period of a video.
In FIG. 12A, "answer address" indicates the location of data on a network, such as the URL (Uniform Resource Locator) of the content representing the answer or an address within a server.

In this embodiment, as in the first embodiment, in order to identify an appropriate answer from the content of the inquiry, the connector table is referred to to "expand the attributes" and the answers are "narrowed down" based on the expanded attributes. In order to realize "expanding the attributes", as in the first embodiment, the following techniques are used: (1) in addition to words (explicit attributes) explicitly included in the expected content of the inquiry, words (implicit attributes) that are not explicitly included, such as the meaning of the expected content of the inquiry itself and the name of the answer, are stored in the connector table; and when generating a response specific attribute table from the connector table, (2) the words representing the attributes are expanded to similar words set in advance, and (3) the words representing the attributes are expanded to similar words acquired via a network. Then, the attributes expanded by these techniques are referred to, and the words included in the utterance for the inquiry are filtered, and if necessary, the information processing system 1 identifies original attributes that are effective for narrowing down, thereby identifying an appropriate answer while carrying out a dialogue to guide the utterance.

The narrowing-down processing unit 155 acquires synonyms of attributes of the anticipated inquiry and the answers prepared for the anticipated inquiry. For example, the narrowing-down processing unit 155 acquires data of a group of synonyms prepared in advance for words expressing attributes, and data of words that can be used as synonyms acquired from outside via the Internet, etc. (for example, data of words expressing similar concepts acquired by referring to an ontology, etc.).

Then, the narrowing-down processing unit 155 collects words (words representing the original attributes stored in the connector table) that represent the attributes (type A and type B attributes) of the anticipated inquiry and the answers prepared for the anticipated inquiry, and words acquired as synonyms (synonyms of the original attributes), and generates an extraction filter for extracting the association between the anticipated inquiry and the answers prepared for the anticipated inquiry.
At this time, the narrowing-down processing unit 155 generates a response specific attribute table ATS0 and a response group specific attribute table ATG0 from the attributes obtained by expanding the words included in the connector table in order to generate the extraction filter. The response specific attribute table ATS0 and the response group specific attribute table ATG0, together with the connector table, constitute the connector tables.

In addition, in the narrowing down process 1, the narrowing down processing unit 155 narrows down the words expressing attributes contained in the connector tables based on the words contained in the latest utterance, and generates a narrowing down table 1 whose elements are information indicating the relevance between the words expressing the narrowed down attributes and the answer number (for example, "1" indicates relevance and "0" indicates no relevance). Specifically, the narrowing down processing unit 155 refers to the connector tables and generates the narrowing down table 1 by words that are related to any of the answer candidates related to the words contained in the latest utterance. Note that the connector table serves as a narrowing down table before being narrowed down by the utterance.

FIG. 12B is a schematic diagram showing an example of the narrowing-down table 1. As shown in FIG.
In FIG. 12B, narrowing-down table 1 stores, in correspondence with each other, words representing attributes contained in the connector table that are related to one of the candidate answers to which the words contained in the most recent utterance relate, along with information indicating whether each word is related to each of the answer numbers ("1" indicating related or "0" indicating not related).
12B, for the word "minutes product" representing an attribute in the narrowing-down table 1, "1" is stored, indicating that it is related to answers #1, #2, and #X. Also, for the word "XX system" representing another attribute, "0" is stored, indicating that it is not related to answer #1.

The numbers of the answer candidates stored in the narrowing-down table 1 are narrowed down to those related to the words included in the latest utterance, so the narrowing-down table 1 is information showing the answer candidates narrowed down by the latest utterance. In addition, the narrowing-down table 1 also stores words representing attributes related to the answer candidates, so it contains information necessary to narrow down the answer candidates by the next utterance.
At this time, the narrowing-down processing unit 155 generates a response specific attribute table ATS1 and a response group specific attribute table ATG1 from the attributes included in the narrowing-down table 1. The response specific attribute table ATS1 and the response group specific attribute table ATG1, together with the narrowing-down table 1, constitute the narrowing-down tables 1.

The narrowing down processing unit 155 determines whether or not a candidate answer can be uniquely identified by the words contained in the latest utterance based on the narrowing down table 1 generated in the narrowing down process 1. If a candidate answer can be uniquely identified, the candidate answer is used to determine a candidate answer based on attribute relevance.

In addition, when the narrowing-down processing unit 155 cannot uniquely identify answer candidates based on the words included in the latest utterance based on the narrowing-down table 1 generated in the narrowing-down processing 1, the narrowing-down processing unit 155 performs a dialogue to induce the user to speak of attributes that can further narrow down answers (i.e., attributes that are effective in identifying answers). Specifically, the narrowing-down processing unit 155 adopts attributes that can narrow down answers more than at present (i.e., attributes that can uniquely identify answers or attributes that reduce answer candidates more than at present) among attributes whose column elements in the narrowing-down table 1 include "1" as attributes to be presented to the user as information for narrowing down. In addition, the narrowing-down processing unit 155 updates the attributes included in the narrowing-down table 1 to those that can narrow down answers more than at present. That is, the narrowing-down processing unit 155 excludes words that are related to all of the answer candidates related to the words included in the latest utterance (i.e., attributes that are not effective in identifying answers) in the narrowing-down table 1, and updates the narrowing-down table 1 by leaving words that are related to some of the answer candidates related to the words included in the latest utterance (i.e., attributes that are effective in identifying answers). Specifically, the narrowing-down processing unit 155 identifies a group of answers related to the words included in the latest utterance in the narrowing-down table 1, deletes attributes of columns in which the elements of the table corresponding to the identified group of answers are all "1", and leaves attributes of columns in which the elements of other attributes have at least one "1", thereby updating the narrowing-down table 1. In addition, in accordance with the update of the narrowing-down table 1, the response specific attribute table ATS1 and the response group specific attribute table ATG1 of the narrowing-down tables 1 are also updated, and the updated narrowing-down tables 1 are passed as input to the narrowing-down process 2.
By performing such processing, the user can be prompted to speak attributes that are effective for narrowing down answers from the system.

In addition, in the narrowing down process 2, the narrowing down processing unit 155 uses the narrowing down tables 1 as input, narrows down the words expressing attributes included in the updated narrowing down table 1 based on the words included in the latest utterance, and generates the narrowing down table 2 whose elements are information indicating the relevance between the words expressing the narrowed down attributes and the answer number (for example, "1" indicates relevance and "0" indicates no relevance). Specifically, the narrowing down processing unit 155 refers to the narrowing down table 1 and generates the narrowing down table 2 by the words that are related to any of the answer candidates related to the words included in the latest utterance.

FIG. 12C is a schematic diagram showing an example of the narrowing-down table 2. As shown in FIG.
In FIG. 12C , narrowing-down table 2 stores, in correspondence with each other, words representing attributes included in narrowing-down table 1 that are related to one of the candidate answers to which the words included in the most recent utterance relate, along with information indicating whether each word is related to each of the answer numbers ("1" indicating related or "0" indicating not related).
12C, for the word "XX system" representing an attribute in the narrowing-down table 2, "1" is stored, indicating that it is related to the answer of #2, #X. Also, for the word "remote conference" representing another attribute, "0" is stored, indicating that it is not related to the answer of #2.

The numbers of the answer candidates stored in the narrowing-down table 2 are narrowed down to those related to the words included in the latest utterance and the previous utterance, so the narrowing-down table 2 is information showing the answer candidates narrowed down by the latest utterance and the previous utterance. In addition, the narrowing-down table 2 also stores words that represent attributes related to the answer candidates, so it contains information necessary to narrow down the answer candidates by the next utterance.
At this time, the narrowing-down processing unit 155 generates a response specific attribute table ATS2 and a response group specific attribute table ATG2 from the attributes included in the narrowing-down table 2. The response specific attribute table ATS2 and the response group specific attribute table ATG2, together with the narrowing-down table 2, constitute the narrowing-down tables 2.

The narrowing-down processing unit 155 also determines whether or not answer candidates can be uniquely identified by the words contained in the latest utterance, based on the narrowing-down table 2 generated in the narrowing-down process 2. If answer candidates can be uniquely identified, the answer candidates are used to determine answer candidates based on attribute relevance.

In addition, when the narrowing-down processing unit 155 cannot uniquely identify answer candidates based on the words included in the latest utterance based on the narrowing-down table 2 generated in the narrowing-down processing 2, the narrowing-down processing unit 155 performs a dialogue to induce the user to speak of attributes that can further narrow down answers (i.e., attributes that are effective in identifying answers). Specifically, the narrowing-down processing unit 155 adopts attributes that can narrow down answers more than at present (i.e., attributes that can uniquely identify answers or attributes that reduce answer candidates more than at present) among attributes whose column elements in the narrowing-down table 2 include "1" as attributes to be presented to the user as information for narrowing down. In addition, the narrowing-down processing unit 155 updates the attributes included in the narrowing-down table 2 to those that can narrow down answers more than at present. That is, the narrowing-down processing unit 155 excludes words that are related to all of the answer candidates related to the words included in the latest utterance (i.e., attributes that are not effective in identifying answers) in the narrowing-down table 2, and updates the narrowing-down table 2 by leaving words that are related to some of the answer candidates related to the words included in the latest utterance (i.e., attributes that are effective in identifying answers). Specifically, the narrowing-down processing unit 155 updates the narrowing-down table 2 by identifying a group of answers that are related to the words included in the latest utterance, deleting attributes of columns in which all elements of the table corresponding to the identified group of answers are "1", and leaving attributes of columns in which at least one element is "1" among other attributes. At this time, the narrowing-down table 2 is updated by excluding the words included in the latest utterance. In addition, in accordance with the update of the narrowing-down table 2, the response specific attribute table ATS2 and the response group specific attribute table ATG2 of the narrowing-down tables 2 are also updated, and the updated narrowing-down tables 2 are passed as input to the narrowing-down process 3.
By performing such processing, the user can be prompted to speak attributes that are effective for narrowing down answers from the system.

In addition, in the narrowing down process 3, the narrowing down processing unit 155 uses the narrowing down tables 2 as input from the start of the inquiry to the third utterance, narrows down the words expressing attributes included in the updated narrowing down table 2 based on the words included in the latest utterance, and generates the narrowing down table 3 whose elements are information indicating the relevance between the words expressing the narrowed down attributes and the answer number (for example, "1" indicates relevance and "0" indicates no relevance). Specifically, the narrowing down processing unit 155 refers to the narrowing down table 2 and generates the narrowing down table 3 by the words related to any of the answer candidates related to the words included in the latest utterance.

In addition, in the narrowing down process 3, from the fourth utterance onwards from the start of the inquiry, the narrowing down processing unit 155 inputs the narrowing down tables 3 updated in the previous narrowing down process 3, narrows down the words expressing attributes included in the updated narrowing down table 3 based on the words included in the latest utterance, and generates a narrowing down table 3 whose elements are information indicating the relevance between the words expressing the narrowed down attributes and the answer number (for example, "1" indicates relevance and "0" indicates no relevance). Specifically, the narrowing down processing unit 155 refers to the narrowing down table 2 and generates the narrowing down table 3 by words that are related to any of the answer candidates related to the words included in the latest utterance.

FIG. 12D is a schematic diagram showing an example of the narrowing-down table 3. As shown in FIG.
In FIG. 12D, narrowing-down table 3 stores, in correspondence with each other, words representing attributes that are the result of narrowing down the words representing attributes contained in the input narrowing-down table 2 or narrowing-down table 3 to words that are related to some of the answer candidates to which the words contained in the most recent utterance are related, along with information indicating whether each word is related to each answer number ("1" indicating related or "0" indicating not related).
For example, in FIG. 12D, for the word "remote conference" representing an attribute in narrowing-down table 3, "1" is stored, indicating that it is related to the answer of #X.

The numbers of the answer candidates stored in the narrowing-down table 3 are narrowed down to those related to the words contained from the start of the inquiry to the latest utterance, so the narrowing-down table 3 is information showing the answer candidates narrowed down by the latest utterance from the start of the inquiry. In addition, the narrowing-down table 3 also stores words representing attributes related to the answer candidates, so it contains information necessary to narrow down the answer candidates by the next utterance.
At this time, the narrowing-down processing unit 155 generates a response specific attribute table ATS3 and a response group specific attribute table ATG3 from the attributes included in the narrowing-down table 3. The response specific attribute table ATS3 and the response group specific attribute table ATG3, together with the narrowing-down table 3, constitute the narrowing-down tables 3.

The narrowing-down processing unit 155 also determines whether or not answer candidates can be uniquely identified by the words contained in the latest utterance, based on the narrowing-down table 3 generated in the narrowing-down process 3. If answer candidates can be uniquely identified, the answer candidates are used to determine answer candidates based on attribute relevance.

In addition, when the narrowing-down processing unit 155 cannot uniquely identify answer candidates based on the words included in the latest utterance based on the narrowing-down table 3 generated in the narrowing-down processing 3, the narrowing-down processing unit 155 performs a dialogue to induce the user to speak of attributes that can further narrow down answers (i.e., attributes that are effective in identifying answers). Specifically, the narrowing-down processing unit 155 adopts attributes that can narrow down answers more than at present (i.e., attributes that can uniquely identify answers or attributes that reduce answer candidates more than at present) among attributes whose column elements in the narrowing-down table 3 include "1" as attributes to be presented to the user as information for narrowing down. In addition, the narrowing-down processing unit 155 updates the attributes included in the narrowing-down table 3 to those that can narrow down answers more than at present. That is, the narrowing-down processing unit 155 excludes words that are related to all of the answer candidates related to the words included in the latest utterance (i.e., attributes that are not effective in identifying answers) in the narrowing-down table 3, and leaves words that are related to some of the answer candidates related to the words included in the latest utterance (i.e., attributes that are effective in identifying answers), and updates the narrowing-down table 3. Specifically, the narrowing-down processing unit 155 updates the narrowing-down table 3 by identifying a group of answers that are related to the words included in the latest utterance, deleting attributes of columns in which all elements of the table corresponding to the identified group of answers are "1", and leaving attributes of columns in which at least one element is "1" among other attributes. At this time, the narrowing-down table 3 is updated by excluding the words included in the latest utterance. In addition, in accordance with the update of the narrowing-down table 3, the response specific attribute table ATS3 and the response group specific attribute table ATG3 of the narrowing-down tables 3 are also updated, and the updated narrowing-down tables 3 are used as input for the narrowing-down process 3.
By performing such processing, the user can be prompted to speak attributes that are effective for narrowing down answers from the system.

The narrowing down processing unit 155 executes narrowing down processing 1, narrowing down processing 2, and narrowing down processing 3 in parallel, and sequentially generates and updates narrowing down tables 1 to 3 by narrowing down processing 1 to 3. The narrowing down processing unit 155 then outputs the results of identifying answer candidates by narrowing down processing 1 to 3 to the integrated processing unit 156. Note that narrowing down processing 1 is executed from the first utterance in the inquiry, narrowing down processing 2 is executed from the second utterance in the inquiry (utterance in response to the first dialogue performed by the system to narrow down the answers), and narrowing down processing 3 is executed from the third utterance in the inquiry (utterance in response to the second dialogue performed by the system to narrow down the answers).

Here, in the narrowing down processes 1 to 3, the generation (or updating) of the narrowing down tables 1 to 3 and the replacement (transfer) of the narrowing down tables that serve as inputs to the narrowing down processes 2 and 3 can take various forms that enable narrowing down the answers for each utterance, but as an example, in this embodiment, the following procedure is defined.
(1) Narrowing table 1 is generated sequentially when a word is extracted by an extraction filter in the latest utterance by narrowing down the words contained in the connector table by narrowing down process 1. When an answer is not uniquely identified and narrowing down table 1 is updated, the input of narrowing down process 2 is replaced with the updated narrowing down table 1.
(2) Narrowing table 1 is not updated if no words are extracted by the extraction filter in the latest utterance.

(3) Narrowing table 2 is not generated (narrowing process 2 is not executed) for an utterance immediately after initialization (immediately after narrowing tables 1 to 3 are reset).
(4) For utterances other than those immediately after initialization, if there is an attribute in the narrowing-down tables 1, which are the output of the narrowing-down process 1 for the previous utterance, that matches a word extracted in the latest utterance, the narrowing-down table 2 is generated by narrowing down the words included in the narrowing-down table 1 by the narrowing-down process 2. If the answer is not uniquely identified and the narrowing-down table 2 is updated, the input of the narrowing-down process 3 is replaced with the updated narrowing-down table 2.
(5) For utterances other than those immediately after initialization, the narrowing-down table 2 will not be generated if there are no attributes in the narrowing-down tables 1, which are the output of the narrowing-down process 1 for the previous utterance, that match the words extracted in the latest utterance.

(6) Narrowing table 3 is not generated (narrowing process 3 is not executed) for an utterance immediately after initialization (after narrowing tables 1 to 3 are reset).
(7) Narrowing table 3 is not generated (narrowing process 3 is not executed) for the second utterance after initialization (after narrowing tables 1 to 3 are reset).
(8) For an utterance other than immediately after initialization or the second time since initialization, if there is an attribute in the narrowing-down tables 2, which are the output of the narrowing-down process 2 for the previous utterance, that matches a word extracted in the latest utterance, the narrowing-down table 3 is generated by narrowing down the words included in the narrowing-down table 2 by the narrowing-down process 2. If an answer is not uniquely identified and the narrowing-down table 3 is updated, when a subsequent utterance is made to inquire, the input of the narrowing-down process 3 is replaced with the updated narrowing-down table 3.
(9) For an utterance other than immediately after initialization or the second utterance after initialization, the narrowing-down table 3 will not be generated if there is no attribute matching the word extracted in the latest utterance in the narrowing-down tables 2 which are the output of the narrowing-down process 2 for the previous utterance or in the narrowing-down tables 3 which are the output of the narrowing-down process 3 for the previous utterance.

FIG. 13 is a schematic diagram showing an example of generating an extraction filter.
For ease of explanation, attention will be paid only to the columns in FIG. 13 in which specific data is shown, and data in columns indicated with "..." will not be taken into consideration (the same applies to FIG. 14).
In addition, in the example shown in FIG. 13, for convenience of explanation, the attributes of the entire content representing the answer are illustrated as the target, but the attributes of a portion of the content representing the answer can also be handled in the same manner.

As shown in FIG. 13, when an extraction filter is generated, the narrowing down processing unit 155 first refers to the connector DB 174, extracts words that represent attributes that can uniquely identify answers prepared for the expected inquiry, and generates table-format data (response specific attribute table ATS0) that corresponds to the numbers assigned to each answer. The narrowing down processing unit 155 also refers to the connector DB 174, extracts words that represent attributes that can identify a group of answers (whole or part) prepared for the expected inquiry, and generates table-format data (response group specific attribute table ATG0) that corresponds to the number group assigned to the answer for each attribute. Furthermore, the narrowing down processing unit 155 acquires words included in the response specific attribute table ATS0 and synonyms of words included in the response group specific attribute table ATG0, and generates table-format data (extended attribute table) that corresponds original words included in the response specific attribute table ATS0 and the response group specific attribute table ATG0 to synonyms of the words. Then, the narrowing down processing unit 155 generates an extraction filter consisting of a string of words included in the extended attribute table.

FIG. 14A is a schematic diagram showing an example of a process for identifying answer candidates based on attribute relevance using an extraction filter.
FIG. 14B is a schematic diagram showing an example of changes in the narrowing-down tables 1 to 3 when an inquiry utterance is input multiple times.
Note that FIG. 14A shows an example in which answer candidates are identified from words extracted by the extraction filter via the response specific attribute table ATS0 and response group specific attribute table ATG0 included in the connector tables, but answer candidates are similarly identified in the narrowing down tables 1 to 3 via the response specific attribute tables ATS1 to 3 and response group specific attribute tables ATG1 to 3.
For example, assume that the following utterance example (1) is made as an utterance to start an inquiry.
Example utterance (1): "I would like to know about meeting minutes products."

In this case, it is assumed that words expressing attributes are extracted from the utterance of the utterance example (1) as follows:
Extracted words: "Minutes product" (original words)

In the case of utterance example (1), "minutes product" matches the words included in the extraction filter, and the original word "minutes product" is included in the response group specific attribute table ATG0. At this time, answers #1, #2, and #X are identified by "minutes product", and narrowing down process 1 generates narrowing down table 1 (see "Results of utterance example (1)" in FIG. 14B). Specifically, attributes associated with each answer are acquired in the identified answer groups #1, #2, and #X. Then, narrowing down table 1, response specific attribute table ATS1, and response group specific attribute table ATG1 are generated, whose elements are information indicating the association between the word representing the acquired attribute and the number of each answer.

In this case, since the answer cannot be uniquely identified, the narrowing down table 1 is referred to, and among the attributes whose column elements in the narrowing down table 1 contain "1", attributes that can narrow down the answers more than at present (i.e., attributes that can uniquely identify the answer or attributes that reduce the number of answer candidates from the current number) are adopted as attributes to be presented to the user as information for narrowing down. In addition, in the narrowing down table 1, words that are related to all of the answer candidates related to the words included in the latest utterance (i.e., attributes that are not effective in identifying the answer) are excluded, and the narrowing down table 1 is updated while leaving words that are related to some of the answer candidates related to the words included in the latest utterance (i.e., attributes that are effective in identifying the answer). As a result, a dialogue is executed that induces the user to speak of attributes that are effective in identifying the answer.
Specifically, referring to Narrowing Table 1, the answers to #1, #2, and #X can be narrowed down more than the current answers if either "XX system" or "remote conference" is spoken, so the following utterance is output to the user: "We can answer if you say 'XX system' or 'remote conference'."

When the information processing system 1 performs a dialogue to induce the user to speak an utterance of an attribute effective for identifying an answer, narrowing-down processes 1 and 2 are executed based on the words included in the utterance, and narrowing-down tables 1 and 2 are generated. As a result, the answer candidates are further narrowed down.
For example, assume that after the information processing system 1 has performed a dialogue to induce an utterance of an attribute effective for identifying an answer, the user makes an utterance such as the following additional utterance example as a second utterance.
Additional utterance example (1): "I'm interested in the XX system."

In the additional utterance example (1), "XX system" matches the words included in the extraction filter.
In the narrowing process 1, the same process as in the case of the "minutes product" described above is executed for the additional utterance example (1).
On the other hand, in the narrowing down process 2, the original word "XX system" is included in the response group specific attribute table ATG1. At this time, the responses #2 and #X are identified by "minutes product", and the narrowing down table 2 is generated by the narrowing down process 2 (see "Result of additional utterance example (1)" in FIG. 14B). Specifically, in the identified response groups #2 and #X, attributes associated with each response are acquired. Then, the narrowing down table 2, the response specific attribute table ATS2, and the response group specific attribute table ATG2 are generated, whose elements are information indicating the association between the word representing the acquired attribute and the number of each response.

In this case, since the answer cannot be uniquely specified, the narrowing down table 2 is referred to, and among the attributes whose column elements in the narrowing down table 2 contain "1", attributes that can narrow down the answers more than at present (i.e., attributes that can uniquely specify the answer or attributes that reduce the number of answer candidates more than at present) are adopted as attributes to be presented to the user as information for narrowing down. In addition, in the narrowing down table 2, words that are associated with all of the answer candidates related to the words included in the latest utterance (i.e., attributes that are not effective in specifying the answer) are excluded, and the narrowing down table 2 is updated by leaving words that are associated with some of the answer candidates related to the words included in the latest utterance (i.e., attributes that are effective in specifying the answer).
This allows a dialogue to be executed that prompts the user to speak about attributes that are effective in identifying an answer.
Specifically, by referring to Narrowing Table 2, the answers to #2 and #X can be uniquely narrowed down if "remote conference" is spoken, so the following speech is output to the user: "If you say 'remote conference', we can answer that question."

When the information processing system 1 performs a dialogue to induce the user to speak an utterance of an attribute effective for identifying an answer, the narrowing-down processes 1 to 3 are executed based on the words included in the utterance, and the narrowing-down tables 1 to 3 are generated. As a result, the answer candidates are narrowed down uniquely in the narrowing-down table 3.
For example, assume that after the information processing system 1 has performed a dialogue to induce an utterance of an attribute effective for identifying an answer, the user utters an utterance such as the following additional utterance example as a third utterance.
Additional utterance example (2) "What is a remote meeting?"

In the additional utterance example (2), "remote conference" matches the word included in the extraction filter.
For the additional utterance example (2), in the narrowing down processes 1 and 2, the same processes as those in the above-mentioned cases of "minutes product" and "XX system" are executed.
On the other hand, in the narrowing-down process 3, the original word "remote conference" is included in the response group identification attribute table ATG2. At this time, the answer of #X is identified by "remote conference", and the narrowing-down process 3 generates the narrowing-down table 3 (see "Result of additional utterance example (2)" in FIG. 14B). As a result, in the example shown here, the answer candidates are uniquely narrowed down in the narrowing-down table 3.

The integration processing unit 156 executes a process of integrating and identifying the answer candidates determined by the identified language processing and the results of narrowing down processes 1 to 3, and identifies the response content to the spoken inquiry (the final answer or the content of the dialogue that induces the utterance of an attribute that is effective in identifying the answer) as the final answer. In this embodiment, the integration processing unit 156 identifies the final answer using the following selection criteria.

That is, the integration processing unit 156
(i) If a candidate answer based on language processing is identified that has a certain degree of accuracy (for example, a score of the identified result based on language processing is above a threshold), the candidate answer based on language processing is selected as the final answer.
(ii) When no candidate answer with a certain degree of certainty has been identified through language processing, if candidate answers based on attribute relevance are narrowed down to one answer in narrowing down table 1, narrowing down table 2, or narrowing down table 3, that narrowed down answer is selected as the final answer.
(iii) If the answer candidate based on the attribute relevance is other than the result of (ii), if the answer candidate generated by language processing is included in the answer candidates in narrowing down table 1, narrowing down table 2, or narrowing down table 3, the answer candidate generated by language processing is selected as the final answer.
(iv) If the answer candidate based on the attribute relevance is not the result of (iii), refer to the narrowing down tables that had a narrowing down effect on the latest utterance (i.e., the results of narrowing down processes 1 to 3 that were able to narrow down the answers), generate a message that encourages the user to speak next including an attribute that has a narrowing down effect (an attribute that reduces the number of answer candidates from the current number), and use this as the response. However, if there are multiple narrowing down tables that had a narrowing down effect on the latest utterance, refer to the response group specific attribute table that is the output of the narrowing down process with a higher number.

[Action]
Next, the operation of the information processing system 1 will be described.
[Automatic search process]
FIG. 15 is a flowchart showing the flow of the automatic search process executed by the information processing system 1.
The automatic search process is started in response to an operation being performed on the server 20 to instruct execution of the automatic search process.
In the flowchart shown in FIG. 15, the processes in steps S1 to S5 are similar to the automatic response process in the first embodiment shown in FIG. 10A.
In this embodiment, prior to the execution of the automatic search process, the connector table, response specific attribute table ATS0, and response group specific attribute table ATG0 (connector tables) used in the narrowing down process 1 are set.

In step S 6 A, narrowing-down processing section 155 executes narrowing-down processing 1 to 3 in parallel, and outputs each of the identification results to integration processing section 156 .
In step S7A, the integration processing unit 156 determines whether or not answer candidates have been identified through language processing (whether or not answer candidates having a certain degree of accuracy have been identified).
If answer candidates can be identified through language processing, a YES determination is made in step S7A, and the process proceeds to step S13A.
On the other hand, if answer candidates have not been identified through language processing, a NO determination is made in step S7A, and the process proceeds to step S8A.

In step S8A, the integration processing unit 156 specifies answer candidates based on the results of the narrowing down processes 1 to 3, using preset selection criteria.
In step S9A, the integration processing unit 156 determines whether or not answer candidates have been uniquely identified through the narrowing-down processes 1 to 3.
If answer candidates can be uniquely identified through narrowing down processes 1 to 3, a YES determination is made in step S9A, and the process proceeds to step S13A.
On the other hand, if answer candidates have not been uniquely identified through narrowing down processes 1 to 3, a NO determination is made in step S9A, and the process proceeds to step S10A.

In step S10A, the narrowing-down processing unit 155 updates the narrowing-down tables 1 to 3. As a result, the attributes included in the narrowing-down tables 1 to 3 are updated to those that can narrow down the answers more than at present. Note that if the narrowing-down tables 1 to 3 have not been narrowed down by the words included in the latest utterance (the number of answer candidates has not decreased from the current number), the narrowing-down tables to which the narrowing-down tables belong are not updated (the process of step S10A is skipped).
In step S11A, the narrowing-down processing unit 155 determines the dialogue content that induces an utterance of an attribute.
In step S12A, the narrowing-down processing unit 155 replaces (transfers) the updated narrowing-down tables as inputs for other narrowing-down processes. Note that if the narrowing-down tables 1 to 3 have not been narrowed down by the words included in the latest utterance (the number of answer candidates has not decreased from the current number), the narrowing-down tables to which the narrowing-down table belongs are not replaced (transferred) as inputs for other narrowing-down processes (the process of step S12A is skipped).

In step S13A, the integration processing unit 156 outputs (sends to the terminal device 10) a response to the inquiry as voice or text data. At this time, if a candidate answer by language processing has been identified as a response to the inquiry, the integration processing unit 156 outputs the identified candidate answer by language processing, and if a candidate answer by language processing has not been identified, it outputs a candidate answer based on attribute relevance or dialogue content that induces the speaking of the attribute. The response data sent in step S13A is received by the response receiving unit 53 of the terminal device 10, and is output by the response output unit 54 as voice or image display. At this time, if multiple candidate answers have been narrowed down, these multiple candidate answers may be output as the final answer.

In step S14A, the integrated processing unit 156 determines whether or not a termination condition for the automatic response process is met. The termination condition for the automatic response process can be defined as, for example, a unique response to an uttered inquiry being identified and a response being output, or an operation to terminate the inquiry being performed by the user.
If the end condition for the inquiry process is not met, a negative determination is made in step S14A, and the process proceeds to step S1.
On the other hand, if the end condition for the automatic response process is met, a YES determination is made in step S14A, and the automatic search process ends.

As described above, in the information processing system 1 according to the present embodiment, multiple narrowing down processes 1 to 3 having different functions are executed in parallel as a process for selecting answer candidates based on attribute relevance, and the content of the response is identified by integrating the identification results of these parallel processes and the identification results of the identification process using language processing. Then, when an answer cannot be identified by a single utterance for inquiry, the information processing system 1 automatically narrows down attributes effective for identifying the answer, and autonomously carries out a dialogue that induces the user to make a meaningful utterance.
This can increase the likelihood that an appropriate answer will be identified as an answer candidate based on attribute relevance.

In addition, in the information processing system 1 according to this embodiment, in addition to identifying the entire content representing the answer as the answer, a part of the content (e.g., a paragraph, a section, or a time period of a video) is also identified as the answer.
This makes it possible to identify a part of an appropriate answer, even if the appropriate answer is included as part of the content of the answer prepared for an anticipated inquiry, as the answer, thereby further increasing the possibility of identifying an appropriate answer.
Therefore, according to the information processing system 1, when an automatic response to a speech is made, it is possible to specify more appropriate response content while shortening the processing time.

[Third embodiment]
Next, a third embodiment of the present invention will be described.
The information processing system 1 of this embodiment does not perform identification of answer candidates through language processing in the information processing system 1 of the first embodiment, but instead executes in parallel a first identification process for identifying answer candidates based on attribute relevance using an inquiry extraction filter generated from attributes of an anticipated inquiry, and a second identification process for identifying answer candidates based on attribute relevance using an answer extraction filter generated from attributes of answers prepared for the anticipated inquiry, and outputs the answer identified in the first identification process or the second identification process as the final answer.

The main parts of the system configuration of the information processing system 1, the hardware configuration of the information processing device, the functional configuration of the terminal device 10, and the like in this embodiment are similar to those in the first embodiment, so the differences will be mainly described.
The following mainly describes the functional configuration of the server 20 and the flow chart of the automatic response process, which are different from the first embodiment.

[Functional configuration]
FIG. 16 is a block diagram showing the functional configuration of the server 20 in this embodiment.
FIG. 17 is a schematic diagram showing a specific example of the functional configuration of the information processing system 1 in this embodiment.
The response content DB173 stores data in a response content table that corresponds text data (here, the name of the person in charge) representing the anticipated inquiry content with responses (here, the telephone number of the person in charge) prepared for the anticipated inquiry.

FIG. 18 is a schematic diagram showing an example of a response content table in this embodiment.
As shown in FIG. 18, the response content table of this embodiment stores, in association with each other, text data (name of person in charge) representing the anticipated inquiry content, a response prepared for the anticipated inquiry (phone number of the person in charge), a number assigned to the anticipated inquiry, and a number assigned to the response prepared for the anticipated inquiry.

The connector DB 174 stores data of a first connector table corresponding to anticipated inquiries and data of a second connector table corresponding to responses prepared to the anticipated inquiries.
FIG. 19 is a schematic diagram showing an example of the first connector table, and FIG. 20 is a schematic diagram showing an example of the second connector table.

As shown in FIG. 19, the first connector table stores, in association with each other, a number assigned to an anticipated inquiry, text data representing the anticipated inquiry (i.e., the name of the person in charge), a word representing an attribute of the anticipated inquiry (such as the name of the person in charge), and a number assigned to a response prepared to the anticipated inquiry (the person in charge's telephone number). For example, the second row of the first connector table shown in FIG. 19 stores the word "Kono" (surname) as an attribute of the anticipated inquiry for "Kono Saburo" (name) representing the anticipated inquiry.

As shown in FIG. 20, the second connector table stores, for each anticipated inquiry (i.e., person in charge), a word representing the response ability and attributes of the person in charge (department, position, product in charge, etc.) in association with a number assigned to the answer prepared for the anticipated inquiry. For example, the second row of the second connector table shown in FIG. 20 stores a word representing the attributes of the answer prepared for the anticipated inquiry (head office/development department/development section/.../XX product/...) and a number assigned to the answer prepared for the anticipated inquiry.

Returning to Figures 16 and 17, the server 20 in this embodiment does not identify answer candidates through language processing, and therefore does not include a language processing unit 154 and an integration processing unit 156. In addition, the server 20 in this embodiment is provided with a word extraction speech recognition unit 501A instead of the dictation speech recognition unit 501 of the first embodiment shown in Figure 9, in order to perform two word extraction speech recognition processes in parallel without performing dictation speech recognition processing.

Here, one of the two word extraction speech recognition processes is a speech recognition process for extracting words contained in an anticipated inquiry, and the other is a speech recognition process for extracting words contained in a prepared response to the anticipated inquiry.
Therefore, in FIG. 17 , the dictionary data of word dictionary DB 172 referred to by word extraction speech recognition unit 501A has registered therein words contained in anticipated inquiries, and the dictionary data of word dictionary DB 172 referred to by word extraction speech recognition unit 511 has registered therein words contained in responses prepared to the anticipated inquiries.

In addition, the narrowing-down processing unit 155 in Figures 16 and 17 executes a first identification process to identify answer candidates based on attribute relevance using an inquiry extraction filter generated from attributes of the anticipated inquiry, and executes a second identification process to identify answer candidates based on attribute relevance using an answer extraction filter generated from attributes of answers prepared for the anticipated inquiry.
Then, from the results of the first and second identification processes executed by the narrowing-down processing unit 155, an answer to the question is uniquely identified and is regarded as the final answer.

FIG. 21 is a schematic diagram showing an example of generating a query extraction filter.
For ease of explanation, attention will be paid only to the columns in Figure 20 in which specific data is shown, and data in columns marked with "..." will not be taken into consideration (the same applies to Figures 22 to 25).
As shown in FIG. 21, when generating an inquiry extraction filter, the narrowing down processing unit 155 first refers to the first connector table of the connector DB 174, extracts words of response specific attributes, and generates table-format data (response specific attribute table) in which the words are associated with the numbers assigned to each answer. The narrowing down processing unit 155 also refers to the first connector table of the connector DB 174, extracts words of response group specific attributes, and generates table-format data (response group specific attribute table) in which the numbers assigned to the answers are associated with each attribute. Furthermore, the narrowing down processing unit 155 expands the words included in the response specific attribute table and the words included in the response group specific attribute table with synonyms. For example, the narrowing down processing unit 155 expands the words included in the response specific attribute table with data of a group of synonyms prepared in advance for the word, or data of words that can be used as synonyms obtained from the outside via the Internet, etc. (for example, data of words representing similar concepts obtained by referring to an ontology, etc.). Furthermore, the narrowing-down processing unit 155 generates table-format data (extended attribute table) in which the original words included in the response specific attribute table and the response group specific attribute table are associated with synonyms of the original words, and then generates an inquiry extraction filter consisting of a word string included in the extended attribute table.

FIG. 22 is a schematic diagram showing an example of generating an answer extraction filter.
In this embodiment, an answer extraction filter is generated from attributes of answers prepared for anticipated inquiries, separate from an inquiry extraction filter generated from attributes of anticipated inquiries.
As shown in FIG. 22, when an answer extraction filter is generated, the narrowing down processing unit 155 first refers to the second connector table of the connector DB 174, extracts words of response specific attributes, and generates table-format data (response specific attribute table) in which the words are associated with the numbers assigned to each answer. The narrowing down processing unit 155 also refers to the second connector table of the connector DB 174, extracts words of response group specific attributes, and generates table-format data (response group specific attribute table) in which the numbers assigned to the answers are associated with each attribute. Furthermore, the narrowing down processing unit 155 expands the words included in the response specific attribute table and the words included in the response group specific attribute table with synonyms. For example, the narrowing down processing unit 155 expands the words included in the response specific attribute table with data of a group of synonyms prepared in advance for the word, or data of words that can be used as synonyms obtained from the outside via the Internet, etc. (for example, data of words representing similar concepts obtained by referring to an ontology, etc.). Furthermore, the narrowing-down processing unit 155 generates table-format data (extended attribute table) in which the original words included in the response specific attribute table and the response group specific attribute table are associated with synonyms of the original words, and generates an answer extraction filter consisting of a word string included in the extended attribute table.

Fig. 23 is a schematic diagram showing an example of a process for identifying answer candidates based on attribute relevance using a query extraction filter. Fig. 24 is a schematic diagram showing an example of a process for identifying answer candidates based on attribute relevance using a response extraction filter.
For example, assume that the following utterances (1) to (3) are made.
Example utterance (1): "I'd like to speak to Yamada-san from sales, please."
Example utterance (2): "I'd like to speak to Kawano-san from development, please."
Example utterance (3): "I'd like someone with technical skills who understands the XX product."

At this time, it is assumed that words expressing attributes are extracted from each example utterance as follows:
Example utterance (1): "Yamada" (from the inquiry extraction filter), "Sales" (from the answer extraction filter)
Example utterance (2): "Kawano" (from the inquiry extraction filter), "development" (from the answer extraction filter)
Example utterance (3): "XX product" (from the answer extraction filter), "technology" (from the answer extraction filter)

In the case of utterance example (1), "Yamada" matches the words contained in the inquiry extraction filter, and "sales" matches the words contained in the answer extraction filter, so answers #2 and #3 are identified from "Yamada," and answers #3 and #4 are identified from "sales." That is, in the case of utterance example (1), answer #2 (Yamada Taro) is identified from the logical product of the results of the first and second identification processes, and is taken as the final answer.

In the case of utterance example (2), "Kawano" matches the words contained in the inquiry extraction filter, and "development" matches the words contained in the answer extraction filter, so answer #1 is identified from both "Kawano" and "development." That is, in the case of utterance example (2), answer #1 (Kawano Saburo) is identified from the results of the first identification process or the second identification process and is determined to be the final answer.

In the case of utterance example (3), the attributes do not match the words contained in the inquiry extraction filter, but "XX product" and "technology" match the words contained in the answer extraction filter, and answers #1 and #3 are identified from "XX product," while answer #1 is identified from "technology." That is, in the case of utterance example (3), answer #1 (Kono Saburo) is identified from the results of the second identification process and becomes the final answer.

[Action]
Next, the operation of the information processing system 1 will be described.
FIG. 25 is a flowchart showing the flow of the automatic response process executed by the information processing system 1.
The automatic response process is started in response to an operation of the server 20 instructing execution of the automatic response process.
The automatic response processing in this embodiment differs from the automatic response processing in the first embodiment shown in FIG. 10A in that the processing of step S41-1 is changed to step S41-1A, the processing of step S41-2 is changed to step S41-2A, and the processing of steps S5 to S9 is changed to steps S105 to S107.

That is, in step S4, the speech recognition processing unit 153 of the server 20 executes speech recognition processes (S41-1A to S43-1 and S41-2A to S43-2) in parallel on the pre-processed speech data using a plurality of speech recognition engines.
Specifically, in step S41-1A, the word extraction speech recognition unit 501A processes the speech data using a speech recognition engine for performing speech recognition processing for word extraction (here, this is assumed to be speech recognition processing using a rule grammar or speech recognition processing using a DNN), extracts specific words included in the speech data, and converts them into text data. Note that the speech recognition processing in step S41-1A is a speech recognition processing for extracting words included in an anticipated inquiry, and the speech recognition processing is performed by referring to a dictionary for word extraction in which words included in an anticipated inquiry are registered in advance.

In step S41-2A, the word extraction speech recognition unit 511 processes the speech data using a speech recognition engine for performing speech recognition processing for word extraction (here, this is assumed to be speech recognition processing using a rule grammar or speech recognition processing using a DNN), extracts specific words contained in the speech data, and converts them into text data. Note that the speech recognition processing in step S41-2A is a speech recognition processing for extracting words contained in answers prepared for anticipated inquiries, and the speech recognition processing is performed by referring to a word extraction dictionary in which words contained in answers prepared for anticipated inquiries are registered in advance.
After steps S43-1 and S43-2, the process proceeds to step S105.

In step S105, the narrowing down processing unit 155 executes the first and second identification processes to identify the final answer to the spoken inquiry. That is, the narrowing down processing unit 155 determines whether the voice recognition result of the word extraction acquired by the voice recognition processing unit 153-1 (i.e., the words included in the expected inquiry) matches the words included in the inquiry extraction filter, and if they match, narrows down the answers prepared in advance based on the matching words included in the inquiry extraction filter. The narrowing down processing unit 155 also determines whether the voice recognition result of the word extraction acquired by the voice recognition processing unit 153-2 (i.e., the words included in the answer prepared for the expected inquiry) matches the words included in the answer extraction filter, and if they match, narrows down the answers prepared in advance based on the matching words included in the answer extraction filter. Then, the narrowing down processing unit 155 identifies the final answer to the spoken inquiry based on the narrowing down results of the first and second identification processes.

In step S106, the integration processing unit 156 transmits the identified final answer as voice or text data to the terminal device 10. The transmitted answer data is received by the response receiving unit 53 of the terminal device 10, and is output by the response output unit 54 as voice or image display. Note that the response output in step S106 may be a response that prompts speech to uniquely identify multiple answer candidates, or a response that outputs multiple answer candidates, in addition to a response that is a uniquely identified answer.

In step S107, the integrated processing unit 156 determines whether or not a termination condition for the automatic response process is met. The termination condition for the automatic response process can be defined as, for example, a unique response to an uttered inquiry being identified and outputted, or an operation for terminating the inquiry being performed by the user.
If the end condition for the inquiry process is not met, a negative determination is made in step S107, and the process proceeds to step S1.
On the other hand, if the end condition for the automatic response process is met, the answer in step S107 is YES, and the automatic response process ends.

As described above, in the information processing system 1 of this embodiment, a first identification process is executed in parallel to identify answer candidates based on attribute relevance using an inquiry extraction filter generated from attributes of an anticipated inquiry, and a second identification process is executed in parallel to identify answer candidates based on attribute relevance using an answer extraction filter generated from attributes of answers prepared for the anticipated inquiry, and the answer identified in the first identification process or the second identification process is output as the final answer.
Therefore, it is possible to identify an answer to a query from the content of the anticipated query and the attributes of the answers prepared for the anticipated query, without performing language processing to identify answer candidates.
Therefore, when an automatic response to a speech is made, it is possible to specify more appropriate response content while shortening the processing time.

[Modification 1]
In each of the above-described embodiments, a case has been described in which a plurality of tables are generated as search data, but the present invention is not limited to this.
In other words, the format of the search data is not particularly limited, and as long as it is possible to identify an answer by tracing the relationships between data (such as words representing attributes or answer numbers), it is possible for the data to be composed of a single table or a group of data whose relationships are defined by pointers, etc.

[Modification 2]
In the second embodiment, the narrowing-down process 2 has been described as specifying answer candidates based on attribute relevance, based on the latest utterance and the previous utterance, but is not limited to this.
That is, the narrowing-down process 2 can specify answer candidates based on the attribute relevance, based on the latest utterance and any number of past utterances preceding the latest utterance.
This makes it possible to appropriately set the number of past utterances reflected in the narrowing-down process 2 depending on statistical data or the purpose to which the information processing system 1 is applied.
In addition, as narrowing-down process 2, multiple processes (i.e., multiple-layer narrowing-down process 2) can be executed in parallel to identify answer candidates based on attribute relevance, based on the latest utterance and a predetermined number of different past utterances preceding the latest utterance.
This makes it possible to continue the narrowing down process by utilizing the narrowing down results (narrowing down table) based on the most recent predetermined number of utterances, even if the direction of the user's inquiry changes or if the user speaks words that do not conform to the guidance from the system.
Therefore, it is possible to narrow down the answers by accurately reflecting the user's intention, so that an appropriate answer to the inquiry can be identified quickly in accordance with the user's preferences.

[Modification 3]
In the first and second embodiments, the voice recognition processing unit 153 includes the first voice recognition unit 153-1 to the n-th voice recognition unit 153-n, and a case where a plurality of voice recognition processes are executed in parallel has been described as an example, but this is not limited to this. For example, the voice recognition processing unit 153 may output a result of a voice recognition process performed by one voice recognition engine (for example, a voice recognition engine that performs general-purpose dictation, etc.), and may execute in parallel a process of identifying answer candidates by language processing and a process of identifying answer candidates based on attribute relevance for this voice recognition result. For example, the narrowing-down processes 1 to 3 of the second embodiment and a process of identifying answer candidates by language processing may be executed in parallel for one voice recognition result.
In addition to identifying candidate answers based on language processing and candidate answers based on attribute relevance for one speech recognition result, it is also possible to identify the content (message) of the dialogue that induces the utterance of attributes that are effective in identifying an answer.
At this time, in response to a user's inquiry utterance, AI (Artificial Intelligence) may be used to respond with a dialogue that induces the user to utter an answer or an utterance of an attribute that is effective in identifying an answer.
In this case, it is possible to obtain more useful information than the processing result obtained by a single voice recognition process while suppressing an increase in the processing load.
The types of voice recognition processing that can be used at this time include general-purpose dictation, dictation for a specific field, and voice recognition through word extraction.

[Modification 4]
In the second embodiment, an example was given in which a voice recognition process for extracting words using an extraction filter is performed on a user's speech, and based on the extracted words, narrowing down answer candidates based on attribute relevance are narrowed down by narrowing down processes 1 to 3, but this is not limited to the example.
That is, in addition to a mode in which the user speaks including words for narrowing down answers and the words are input by voice recognition, a mode in which the words are input by operating the screen can also be used.
For example, words for an inquiry to be entered into narrowing-down process 1 can be entered by operating the screen, or words for further narrowing down the answer candidates narrowed down in narrowing-down process 1 or narrowing-down process 2 can be entered by operating the screen.
A form in which words are input by operating the screen can be a form in which words (words representing attributes) related to answer candidates are scrolled (moved and displayed) in a window on the display (output unit 816) of the terminal device 10, and the user selects a word by clicking or tapping with the mouse, etc. Also, other forms in which words are input by operating the screen can be a form in which words (words representing attributes) related to answer candidates are listed in a window on the display (output unit 816) of the terminal device 10 using a pull-down menu or the like, and the user selects a word by clicking or tapping with the mouse, or a form in which words are listed in a table format, and the user selects a word by checking a check box attached to each word.

Note that the words stored in each table included in the connector tables can be displayed as the words to be selected in the narrowing-down process 1. Also, the words stored in each table included in the narrowing-down tables 1 or 2 (i.e., the words after narrowing down by the narrowing-down process) can be displayed as the words to be selected in the narrowing-down process 2 or 3.
When inputting words by operating the screen, the user can select effective words for identifying the answer using candidate words displayed on the screen as a hint.
It is also possible to display words (words representing attributes) related to answer candidates on the screen in any of the above-mentioned formats, and for the user to select a word by speaking any of the words.
Furthermore, in addition to inputting words by operating the screen, it is also possible to input words by operating a keyboard or the like.
In this case, in response to the user inputting any text (such as a word or a sentence), the words contained in the input text can be extracted, thereby narrowing down answer candidates based on the relevance of attributes.

[Modification 5]
In the above embodiment and modified examples, the case where an inquiry is received by voice has been described, but the present invention is not limited to this.
That is, it is also possible to accept inquiries by text data, such as inquiries by chat.
Therefore, in the above-described embodiment and modified examples, the configuration for performing voice recognition processing on a voice inquiry can be appropriately replaced with a configuration for acquiring text data representing the contents of the inquiry.

The information processing system 1 configured as above includes the connector DB 174 , the voice recognition processing unit 153 , the narrowing-down processing unit 155 , and the integration processing unit 156 .
The connector DB 174 stores first attribute information regarding anticipated inquiries.
The connector DB 174 stores second attribute information regarding responses to anticipated inquiries.
The voice recognition processor 153 acquires the contents of the inquiry.
The narrowing down processing unit 155 and the integration processing unit 156 identify a response (answer or dialogue message) including an answer to an inquiry based on attribute information possessed by the voice recognition result from the voice recognition processing, the first attribute information, and the second attribute information.
This makes it possible to identify a response to an inquiry from the content of the anticipated inquiry and the attributes of the response prepared for the anticipated inquiry.
Therefore, when an automatic response to an inquiry is made, it is possible to shorten the processing time and specify more appropriate response content.

The voice recognition processing unit 153 acquires a voice recognition result obtained by subjecting an utterance for inquiry to voice recognition processing.
This makes it possible to shorten the processing time and specify more appropriate response content when an inquiry is made by voice.

The narrowing-down processing unit 155 identifies a response including an answer to the inquiry based on attribute information contained in the voice recognition result by the voice recognition processing, the first attribute information, the second attribute information, and expanded attribute information representing attributes expanded from the first attribute information and the second attribute information.
This makes it possible to easily identify an answer that matches the spoken inquiry even when unexpected expressions are used, such as when there is variation in expression or when words with similar meanings are used.

The information processing system 1 also includes a voice recognition processing unit 153 , a language processing unit 154 , a narrowing-down processing unit 155 , and an integration processing unit 156 .
The voice recognition processing unit 153 acquires a voice recognition result obtained by subjecting an utterance for inquiry to voice recognition processing.
The language processing unit 154 identifies answer candidates that match the inquiry from the meaning expressed by the voice recognition result by the voice recognition processing unit 153 through language processing.
The narrowing-down processing unit 155 extracts words contained in the inquiry from the voice recognition result by the voice recognition processing unit 153, and identifies candidate answers that match the inquiry based on the relevance between attribute information based on the words contained in the inquiry and attribute information based on the answer to the inquiry.
The integration processing unit 156 identifies a response including an answer to the inquiry based on the answer candidates identified by the language processing unit 154 and the answer candidates identified by the narrowing-down processing unit 155 .
This makes it possible to identify an answer to a query using not only an identification method based on language processing, but also an identification method based on the consistency of words expressing attributes.
Therefore, when an automatic response to a speech is made, it is possible to specify more appropriate response content while shortening the processing time.

The narrowing-down processing unit 155 identifies responses including an answer to the inquiry, based on attribute information based on the answer to the inquiry and expanded attribute information representing attributes expanded from the attribute information based on the answer to the inquiry.
This makes it possible to easily identify an answer that matches the spoken inquiry even when unexpected expressions are used, such as when there is variation in expression or when words with similar meanings are used.

When the accuracy of the answer candidate acquired by the language processing unit 154 is equal to or greater than a set threshold, the integration processing unit 156 determines that the answer candidate acquired by the language processing unit 154 is the answer, and when the accuracy of the answer candidate acquired by the language processing unit 154 is less than a set threshold, the integration processing unit 156 determines that the answer candidate identified by the narrowing down processing unit 155 is the answer.
This makes it possible, when a query is spoken, to quickly obtain a more accurate speech recognition result and identify an appropriate answer.

The information processing system 1 also includes a voice recognition processing unit 153 , a narrowing-down processing unit 155 , and an integration processing unit 156 .
The narrowing-down processing unit 155 generates search data consisting of one or more attributes associated with the answer to the expected inquiry based on first attribute information related to the expected inquiry and second attribute information related to the answer to the expected inquiry.
The voice recognition processor 153 acquires the contents of the inquiry.
The narrowing down processing unit 155 and the integration processing unit 156 identify a response including an answer to the content of the inquiry based on attribute information contained in the content of the inquiry and the search data.
This allows for rapid identification of an answer when an inquiry is made using the search data.

The voice recognition processing unit 153 acquires a voice recognition result obtained by subjecting an utterance for inquiry to voice recognition processing.
This makes it possible to quickly identify an answer when a voice inquiry is made using the search data.

The search data is prepared before the content of the inquiry is acquired, and when the content of the inquiry is acquired, the generated search data is referenced to identify an answer to the content of the inquiry.
As a result, since the search data is prepared before a query is made, it becomes possible to quickly identify an answer when a query is made.

The search data includes response identifying information that indicates an attribute that can uniquely identify a response, and response group identifying information that indicates an attribute that can identify a plurality of answer candidates.
This allows the process of identifying an appropriate answer to be performed depending on whether the words contained in the query utterance represent attributes that can uniquely identify an answer or represent attributes that can identify multiple answer candidates.

The search data includes expanded attribute information that represents attributes expanded from attributes included in the response specifying information and the response group specifying information.
This makes it possible to easily identify an answer that matches the spoken inquiry even when unexpected expressions are used, such as when there is variation in expression or when words with similar meanings are used.

When the content of an inquiry acquired by the voice recognition processing unit 153 identifies multiple answers to the inquiry, the narrowing-down processing unit 155 narrows down one or more attributes associated with the answers to the anticipated inquiry contained in the search data to an attribute that uniquely identifies the answer to the content of the inquiry, or an attribute that identifies fewer answers to the content of the inquiry.
The narrowing-down processing unit 155 executes a dialogue to induce the content of an additional inquiry to identify an answer to the content of the inquiry, as a response, based on the narrowed-down attributes.
This makes it possible to accept further inquiries and identify appropriate answers even if the user does not make an appropriate inquiry.

The narrowing down processing unit 155 executes in parallel a first narrowing down process which targets the entire original search data to identify an answer to the content of the most recent inquiry, and a second narrowing down process which targets search data whose attributes have been narrowed down from the original search data to identify an answer based on the content of a specified number of past inquiries.
This can increase the likelihood that an appropriate answer will be identified as an answer candidate based on attribute relevance.

The narrowing-down processing unit 155 executes a third narrowing-down processing in parallel with the first narrowing-down processing and the second narrowing-down processing, which identifies answers based on the content of a predetermined number of past inquiries different from the second narrowing-down processing.
This makes it possible to continue the narrowing down process by utilizing the narrowing down results (narrowing down table) from the most recent specified number of inquiries, even if the direction of the user's inquiry changes or if the user makes an inquiry that does not conform to the guidance from the system.
Therefore, it is possible to narrow down the answers by accurately reflecting the user's intention, so that an appropriate answer to the inquiry can be identified quickly in accordance with the user's preferences.

The narrowing-down processing unit 155 displays on a screen words representing attributes contained in the original search data or words representing attributes narrowed down from the original search data, and narrows down answers to the contents of the inquiry by selecting a word displayed on the screen.
This allows the user to select effective words for identifying the answer using the candidate words displayed on the screen as a hint.

The information processing system 1 also includes a voice recognition processing unit 153 and a narrowing-down processing unit 155 .
The voice recognition processing unit 153 acquires a voice recognition result obtained by subjecting an utterance for inquiry to voice recognition processing.
The narrowing-down processing unit 155 filters attribute information contained in the speech recognition result by the speech recognition processing unit 153 using a filter including, as elements, first attribute information related to an expected inquiry, second attribute information related to a response to the expected inquiry, and expanded attribute information related to the first attribute information and the second attribute information.
The narrowing-down processing unit 155 identifies responses including answers to the utterances for inquiry based on the results of the filtering process.
This makes it possible to quickly extract words expressing attributes from a query utterance and identify answer candidates based on the relevance of attributes.

The narrowing-down processing unit 155 identifies first attribute information or second attribute information related to a word extracted from the inquiry utterance, and identifies an answer to the assumed inquiry associated with the first attribute information or the second attribute information as an answer to the inquiry utterance.
This makes it possible to easily identify answers related to words included in the query utterance.

The present invention is not limited to the above-described embodiment, and modifications and improvements within the scope of the present invention that can achieve the object of the present invention are included in the present invention.
For example, in the above embodiment, an example in which the information processing system 1 is configured as a system including a plurality of information processing devices has been described, but the present invention is not limited to this. That is, the present invention may be realized by providing the functions of the information processing system 1 to a single information processing device.

Furthermore, the functions of the server 20 may be distributed among a plurality of servers, some of the functions of the server may be implemented in the terminal device 10, or some of the functions of the terminal device 10 may be implemented in the server.
For example, in each of the above-described embodiments, the voice recognition processing is executed in the server 20, but the voice recognition processing may be executed in the terminal device 10, or the voice recognition result executed in another system, etc. may be acquired (received, etc.) and used for processing.
In addition, in each of the above-described embodiments, the search data can be updated sequentially in response to updates to the response content table or updates (addition, modification, deletion, etc.) to words expressing attributes based on the operation results.

Furthermore, the present invention can be implemented by appropriately combining the configurations described in the above-mentioned embodiments and modifications.
The above-described series of processes can be executed by hardware or software.
In other words, the functional configurations of Figures 3, 4, 9, 16, and 17 are merely examples and are not particularly limited. In other words, it is sufficient for the information processing system 1 to have a function capable of executing the above-mentioned series of processes as a whole, and the type of functional block used to realize this function is not particularly limited to the examples of Figures 3, 4, 9, 16, and 17.
Furthermore, one functional block may be configured as a single piece of hardware, a single piece of software, or a combination of both.

When the series of processes is executed by software, the program constituting the software is installed into a computer or the like from a network or a recording medium.
The computer may be a computer built into dedicated hardware, or may be a computer capable of executing various functions by installing various programs, such as a general-purpose personal computer.

The storage medium that stores the program is composed of removable media distributed separately from the device body, or storage media that is pre-installed in the device body. Removable media is composed of, for example, a magnetic disk, optical disk, magneto-optical disk, or flash memory. Optical disks are composed of, for example, CD-ROMs (Compact Disk-Read Only Memory), DVDs (Digital Versatile Disks), Blu-ray Discs (registered trademark), etc. Magneto-optical disks are composed of, for example, MDs (Mini-Disks), etc. Flash memory is composed of, for example, USB (Universal Serial Bus) memory or SD cards. Storage media that is pre-installed in the device body is composed of, for example, a ROM in which the program is stored, a hard disk, or flash memory, etc.

In this specification, the steps of describing a program to be recorded on a recording medium include not only processes that are performed chronologically according to the order, but also processes that are not necessarily performed chronologically but are executed in parallel or individually.
In addition, in this specification, the term "system" refers to an overall device that is composed of a plurality of devices, a plurality of means, etc.

1 Information processing system, 10 Terminal device, 20 Server, 30 Network, 800 Information processing device, 811 CPU, 812 ROM, 813 RAM, 814 Bus, 815 Input unit, 815a Operation input unit, 815b Voice input unit, 816 Output unit, 817 Storage unit, 818 Communication unit, 819 Drive, 820 Imaging unit, 831 Removable media, 51 Speech reception unit, 52 Speech data transmission unit, 53 Response reception unit, 54 Response output unit, 151 Speech data reception unit, 152 Preprocessing unit, 153 Speech recognition processing unit, 153-1 First speech recognition unit, 153-2 Second speech recognition unit, 154 Language processing unit, 155 Narrowing processing unit, 156 Integration processing unit, 171 Speech recognition dictionary database (speech recognition dictionary DB), 172 Word dictionary database (word dictionary DB), 173 response content database (response content DB), 174 connector database (connector DB), 501 dictation speech recognition unit, 502 first recognition result identification unit, 503 first recognition result output unit, 511 word extraction speech recognition unit, 512 second recognition result identification unit, 513 second recognition result output unit

Claims (4)

An information processing system including one or more information processing devices,
a search data generating means for generating search data composed of one or more attributes associated with the answer to the anticipated inquiry based on first attribute information related to the anticipated inquiry and second attribute information related to the answer to the anticipated inquiry;
An inquiry content acquisition means for acquiring the content of an inquiry;
a response specifying means for specifying a response including an answer to the content of the inquiry based on attribute information of the content of the inquiry and the search data;
Equipped with
the attribute information of the content of the inquiry, the first attribute information, and the second attribute information are expressed by words ;
The response identification means
a first narrowing-down process for identifying an answer to the latest inquiry from the entire original search data;
a second narrowing process for identifying an answer based on the content of the inquiry made a predetermined number of times in the past, using the search data whose attributes have been narrowed down from the original search data;
An information processing system characterized by executing the above in parallel . The information processing system according to claim 1, characterized in that the response identification means executes a third narrowing-down process in parallel with the first narrowing-down process and the second narrowing-down process, in which a response is identified based on the content of a predetermined number of past inquiries different from the second narrowing-down process. An information processing method executed in an information processing system including one or more information processing devices,
a search data generating step of generating search data composed of one or more attributes associated with the answer to the anticipated inquiry based on first attribute information related to the anticipated inquiry and second attribute information related to the answer to the anticipated inquiry;
An inquiry content acquisition step of acquiring the content of the inquiry;
a response identifying step of identifying a response including an answer to the content of the inquiry based on attribute information of the content of the inquiry and the search data;
Including,
the attribute information of the content of the inquiry, the first attribute information, and the second attribute information are expressed by words;
In the response identification step,
a first narrowing-down process for identifying an answer to the latest inquiry from the entire original search data;
a second narrowing process for identifying an answer based on the content of the inquiry made a predetermined number of times in the past, using the search data whose attributes have been narrowed down from the original search data;
4. An information processing method comprising the steps of : On the computer,
a search data generation function that generates search data composed of one or more attributes associated with the answer to the anticipated inquiry based on first attribute information related to the anticipated inquiry and second attribute information related to the answer to the anticipated inquiry;
An inquiry content acquisition function for acquiring the content of an inquiry;
a response identification function that identifies a response including an answer to the content of the inquiry based on attribute information of the content of the inquiry and the search data;
Realize this,
the attribute information of the content of the inquiry, the first attribute information, and the second attribute information are expressed by words ;
The response identification function is
a first narrowing-down process for identifying an answer to the latest inquiry from the entire original search data;
a second narrowing process for identifying an answer based on the content of the inquiry made a predetermined number of times in the past, using the search data whose attributes have been narrowed down from the original search data;
A program characterized by executing the above in parallel .

Images