JP2020184342A - Dialog management server, dialog management method, and program - Google Patents

Abstract

To provide a dialog management server capable of effectively using a manual, etc. while improving the convenience for users and the efficiency of operations within an organization, a weight determination device, a validity determination device, a dialog management method, a dialog management method, and a program.SOLUTION: A dialog management server device 1 includes: a record recording unit 11 that records a piece of text data of a descriptive text and a piece of document data in a structured format that has multiple pieces of unit information that associate text data of descriptive text with a piece of text data that represents the heading of the information and hierarchy for identifying one or more hierarchies leading to descriptive text; a reception unit 19 that receives a piece of text data of a question sentence from a user; and an answer generation unit 27 that matches the text data of the question sentence received by the reception unit 19 with the text data included in each unit information recorded in recording unit 11, extracts the unit information related to the question sentence, and generates a piece of answer output information based on the heading and descriptive text corresponding to the extracted unit information.SELECTED DRAWING: Figure 17

Description

The present invention relates to a dialogue management server, a dialogue management method, and a program.

Conventionally, in a company, when a clerical worker receives a question about a business content from a sales staff, he / she extracts necessary information from a huge business manual and answers it one by one. Here, a calling system is known for an operator to respond to an inquiry from a customer by telephone.

For example, Patent Document 1 describes a call system that manages the contents of calls made by customers and operators (Patent Document 1).

International Publication No. 2012/120624

However, in the conventional system described in Patent Document 1, if the number of inquiries from customers by telephone is large, it is necessary to have a dedicated operator to handle the inquiries, so that the efficiency of operations in the organization can be improved. Absent. In addition, depending on the content of the inquiry, the operator may not be able to answer immediately, so it cannot be said that the service is highly convenient for the customer.
In addition, in order to respond to inquiries from customers, manuals and Q & A collections for customers may be prepared. However, such manuals can be huge, so
Customers often end up making inquiries over the phone because they don't know where to look in the manual. In this way, despite the existence of manuals and the like, it cannot be said that they are being effectively utilized.

Therefore, some aspects of the present invention have been made in view of such circumstances, and in addition to being able to improve work efficiency within the organization and improving user convenience, effective utilization of manuals and the like is possible. It is an object of the present invention to provide a dialogue management server, a dialogue management method, and a program that can be used.

The dialogue management server according to one aspect of the present invention provides unit information in which the text data of the explanatory text is associated with the information for identifying one or more layers leading to the explanatory text and the text data representing the heading of the hierarchy. A recording unit that records data of multiple structured documents, a reception unit that receives text data of question texts from users, text data of question texts received by the reception department, and recorded in the recording unit. A response generation unit that matches text data included in each unit information, extracts unit information related to the question text, and generates answer output information based on the heading and explanation corresponding to the extracted unit information. To be equipped with.

The dialogue management method according to one aspect of the present invention provides unit information in which text data of an explanatory text is associated with information for identifying one or more layers leading to the explanatory text and text data representing a hierarchy heading. The step of recording the data of a document having a plurality of structured formats, the step of accepting the text data of the question text from the user, the text data of the received question text, and the text data included in each recorded unit information. Includes a step of extracting unit information related to the question text and generating answer output information based on the heading and description corresponding to the extracted unit information.

In the program according to one aspect of the present invention, the computer is subjected to unit information in which the text data of the explanatory text is associated with the information for identifying one or more layers leading to the explanatory text and the text data representing the heading of the hierarchy. A recording unit that records data in a structured format that has multiple documents, a reception unit that receives text data of question texts from users, text data of question texts received by the reception department, and each recorded in the recording unit. It functions as an answer generator that matches the text data included in the unit information, extracts the unit information related to the question text, and generates the answer output information based on the heading and description corresponding to the extracted unit information. Let me.

In the present invention, the "part" does not simply mean a physical means, but the "part" thereof.
It also includes the case where the function of the "part" is realized by software. Further, even if the function of one "part" or device is realized by two or more physical means or devices, the function of two or more "parts" or devices is realized by one physical means or device. You may.

According to the present invention, it is possible to improve the work efficiency in the organization, improve the convenience of the user, and effectively utilize the manual or the like.

It is a schematic block diagram (system block diagram) of the dialogue management system which concerns on 1st Embodiment. It is a conceptual diagram which shows an example of the process of converting a document such as business manual data created by a general word processor or spreadsheet software into a structured document such as OOXML (Office Open XML). It is a conceptual diagram which shows an example of the process of converting a structured document such as OOXML into a structured document of a specific format peculiar to this embodiment. It is a flowchart which shows an example of the dialogue management processing which concerns on 1st Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 1st Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 1st Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 2nd Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 3rd Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 4th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 5th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 5th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 6th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 6th Embodiment. It is a figure which shows an example of the screen of the display part of the operator terminal apparatus which concerns on 6th Embodiment. It is a schematic block diagram (system block diagram) of the dialogue management system which concerns on 7th Embodiment. It is a schematic block diagram (system block diagram) of the dialogue management system which concerns on 8th Embodiment. It is a schematic block diagram (system block diagram) of the dialogue management system which concerns on 9th Embodiment. It is a schematic diagram which shows an example of the subordinate concept word dictionary which concerns on 9th Embodiment. It is a flowchart which shows an example of the dialogue management processing which concerns on 9th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 9th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 9th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 9th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 9th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 9th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 9th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 9th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on 9th Embodiment. It is a figure which shows an example of the screen of the display part of the user terminal apparatus which concerns on modification 6 of the 9th Embodiment. It is a figure which shows an example of the icon of the modification 6 of the ninth embodiment. It is a figure which shows an example of the icon of the modification 6 of the ninth embodiment. It is a schematic block diagram (system block diagram) of the dialogue management system which concerns on tenth embodiment. It is a flowchart for demonstrating the process of the weight determination apparatus which concerns on 10th Embodiment. It is a flowchart for demonstrating the process of the weight determination apparatus which concerns on 10th Embodiment. It is a flowchart for demonstrating the process of the weight determination apparatus which concerns on 10th Embodiment. It is a figure which shows an example of the initial value list which concerns on 10th Embodiment. It is a figure which shows an example of the weight candidate list which concerns on 10th Embodiment. It is a schematic block diagram (system block diagram) of the dialogue management system which concerns on eleventh embodiment. It is a flowchart for demonstrating the process of the validity determination apparatus which concerns on 11th Embodiment. It is a schematic block diagram (system block diagram) of the dialogue management system which concerns on 12th Embodiment.

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. The following embodiments are examples for explaining the present invention, and the present invention is not intended to be limited only to the embodiments. Further, the present invention can be modified in various ways as long as it does not deviate from the gist thereof. Further, in each drawing, the same components are designated by the same reference numerals as much as possible, and duplicate description will be omitted.

<First Embodiment>
FIG. 1 is a schematic configuration diagram (system configuration diagram) of the dialogue management system according to the embodiment of the present invention. As shown in FIG. 1, the dialogue management system 100 is optionally configured to include a dialogue management server 1 (dialogue management server), a dialogue log database (DB) 4, a user terminal device 6, and an operator terminal device 8. ing.

The dialogue management server 1 is a server computer that manages dialogues on a predetermined network N, and exhibits a server function by operating a predetermined server program on the server computer. The “dialogue on the network N” refers to a dialogue between the user and the dialogue management server 1 via the network N, and a dialogue between the operator and the dialogue management server 1 via the network N. For example, "dialogue on network N" includes a question input from the user terminal device 6 to the dialogue management server 1 via the network N, an answer generated by the dialogue management server 1 corresponding to the question, and the said. Includes the user's response to the answer. Further, "dialogue on network N" includes a question input from the operator terminal device 8 to the dialogue management server 1 via the network N, an answer generated by the dialogue management server 1 corresponding to the question, and the said. Includes the operator's response to the answer.

The dialogue log DB 4 is a database that records the dialogue between the user and the dialogue management server 1 and the dialogue between the operator and the dialogue management server 1, and is, for example, a dialogue from the user terminal device 6 via the network N. A dialogue log (dialogue result) including a question input to the management server 1, an answer generated by the dialogue management server 1 corresponding to the question, and a user's response to the answer is recorded. The dialogue log DB 4 may be included in the dialogue management server 1.

The user terminal device 6 and the operator terminal device 8 are devices that output the contents of the dialogue.
For example, an information output device such as a laptop computer or a notebook computer. In addition, it should be noted
The user terminal device 6 and the operator terminal device 8 are, for example, a mobile phone such as a smartphone.
It may be an information output device including a tablet terminal or the like.

The network N is a communication line or a communication network related to information processing including, for example, the Internet, and its specific configuration is data between a dialogue management server 1, a dialogue log DB 4, a user terminal device 6, and an operator terminal device 8. There is no particular limitation as long as it is configured to be able to send and receive.

The dialogue management server 1 exemplifies the information processing unit 10 that executes the dialogue management process for managing the dialogue on a predetermined network N, the information required for the dialogue management process, and the information generated by the dialogue management process. It is configured to include a recording unit 11 for recording, a technical term dictionary DB 12 in which technical terms of the industry in which the document is used are stored, and a synonym dictionary DB 13 in which synonyms of words are stored.

The information processing unit 10 functionally includes, for example, an acquisition unit 15, a data conversion unit 17, and a reception unit 19.
It includes a natural language processing unit 21, a weighting setting unit 23, an answer generation unit 27, a dialogue result management unit 29, an output control unit 31, a voice recognition processing unit 33, and a voice dialogue management unit 35.

Each of the above parts of the information processing unit 10 can be realized, for example, by using a storage area such as a memory or a hard disk, or by executing a program stored in the storage area by a processor. Further, each of the DBs 12 and 13 of the dialogue management server 1 can be realized by executing the processor.

The acquisition unit 15 acquires, for example, a document to be answered. Here, the "document" means, for example, a business manual used in business, a text document such as a Q & A collection, a document created by word processor software, a document created by presentation software, or the like.

The data conversion unit 17 (conversion unit) is based on the data of a document in which the content of the document has a hierarchical structure such as editions, chapters, sections, sections, and sub-items, and the text data included in the items located at the end of the hierarchical structure. Each of the above is converted into a document in a structured format in association with information that can identify one or more layers leading to each item and headings of one or more layers leading to each item. The predetermined unit may be, for example, an item located at the end (bottom layer) of the hierarchical structure. In that case, each of the text data included in the items located at the end of the hierarchical structure is divided into information that can identify one or more layers leading to each item and headings of one or more layers leading to each item. Convert to a document in a structured format in association with each other.

Using FIGS. 2 and 3, a process of converting into a structured document of a specific format peculiar to the present embodiment will be described based on a document such as a business manual or a collection of Q & A. Here, documents such as business manual data created by a general word processor or spreadsheet software are displayed in OOXML (Office O).
An example of converting to a structured document such as pen XML) and then converting to a structured document of a specific format used by the dialogue management server 1 of the present embodiment will be described. However, when converting a document such as a business manual into a structured document in a specific format peculiar to this embodiment, it is not always OO.
It is not necessary to go through XML, and various conversion processes can be adopted.

FIG. 2 is a conceptual diagram showing an example of a process of converting a document such as business manual data created by a general word processor or spreadsheet software into a structured document such as OOXML (Office Open XML). As shown in FIG. 2, the data conversion unit 17 converts the business manual data of the word file into a file aggregate described in OOXML (Office Open XML).

FIG. 3 is a conceptual diagram showing an example of a process of converting a structured document such as OOXML (Office Open XML) into a structured document of a specific format peculiar to the present embodiment. As shown in FIG. 3, the data conversion unit 17 converts the business manual data converted into the file aggregate described in OOXML into a specific XML format. Specifically, the data conversion unit 17 extracts business manual data having a hierarchical structure so that it can be referred to for each structural unit (edition, chapter, section, section, sub-item, etc.) of the document, and ends the hierarchical structure. For each of the text data (in FIG. 3, "at the time of transaction ...") included in the structural unit (item) located in, for each extracted structural unit, information on the hierarchy of the structural unit and information on the hierarchy of the structural unit. It is converted into a document in a structured format by associating it with information related to headings in each layer, and the document is recorded in the recording unit 11 as document information BI (character information) described later.

Returning to FIG. 1, the reception unit 19 accepts the input of a question from the user. In addition, reception department 19
Accepts input from the user in response to the answer to the question.

For example, the natural language processing unit 21 reads the character information in the document and the character information included in the input question, performs morphological analysis, and cuts out each word. "Morphological analysis" is one of the natural language processing of a computer, and is based on information such as the grammar of the target language and the part of speech of a word from natural language text data without notes of grammatical information. It is an analysis process that divides into units of and discriminates part of speech etc. for each morphological element. Further, the natural language processing unit 21 may execute dependency analysis (syntax analysis), that is, investigate the modification relation of the sentence from the part of speech information.

The weighting setting unit 23 calculates the frequency of occurrence of words included in the character information in the document and the character information included in the input question, and sets the weighting for each word based on the calculated frequency of appearance of the words.

The answer generation unit 27 matches the input question with the text data included in each item of the recorded document information, extracts a plurality of items related to the question, and reaches each of the extracted items. Alternatively, an answer including a plurality of layers and their headings is generated and returned to the user as an answer to the question. Further, the answer generation unit 27 may learn matching weighting based on the dialogue result stored in the dialogue log database 4, as will be described later.

The dialogue result management unit 29 stores the dialogue result including the question input by the user, the answer corresponding to the question, and the response from the user to the answer in the dialogue log database 4. The dialogue result management unit 29 stores, for example, the voice information recognized by the voice recognition processing unit 33, which will be described later, in the dialogue log database 4 as the dialogue result.

The output control unit 31 controls to output the answer generated by the answer generation unit 27.
The output control unit 31 controls so that the generated answer is output on the display unit (not shown) of the user terminal device 6 and the operator terminal device 8.

The voice recognition processing unit 33 recognizes the voice information of the voice dialogue between the user and the operator.

The voice dialogue management unit 35 manages so that when a question and an answer to the question are repeated with a certain user a predetermined number of times, a voice dialogue with the user is possible.

The recording unit 11 records the document converted into the structured format as the document information BI. Further, the recording unit 11 records the dialogue result information TI (dialogue result) including the question input by the user, the answer corresponding to the question, and the response from the user to the answer in association with each user. Further, the recording unit 11 may record the stamp information SI which is the information input by the user via the network N.

The "stamp" is image information for dialogue input via the network N. The stamp is, for example, an illustration representing emotions, intentions, and a message to be conveyed, and may include text.

(Dialogue management process)
The dialogue management process according to the embodiment of the present invention will be described with reference to FIGS. 4 and 5. FIG. 4 is a flowchart showing an example of the dialogue management process according to the embodiment of the present invention.

(Step S1)
The acquisition unit 15 shown in FIG. 1 acquires documents such as business manuals and Q & A collections.

(Step S3)
As shown in FIG. 3, the data conversion unit 17 shown in FIG. 1 is based on the data of a document in which the content of the document has a hierarchical structure, for example, to a structural unit located at the end of the hierarchical structure, that is, a node in the lowest layer. Each of the included text data is converted into a document in a structured format in association with information that can identify one or more layers leading to each structural unit and headings of one or multiple layers leading to each structural unit. Convert.

(Step S5)
The recording unit 11 shown in FIG. 1 records a document converted into a structured format as document information BI.

(Step S7)
The reception unit 19 shown in FIG. 1 receives input of a question from the user.

FIG. 5 is a diagram showing an example of a screen of a display unit of a user terminal device according to an embodiment of the present invention. As shown in FIG. 5, as displayed on the screen G of the user terminal device 6 shown in FIG. 1, the reception unit 19 "brought the pension handbook with the identity verification document entered from the user U. Is it valid if there is no? ”Is accepted.

(Step S9)
The answer generation unit 27 shown in FIG. 1 matches the input question with the text data included in each item of the recorded document information, extracts a plurality of items related to the question, and extracts each of the extracted items. Generates answer output information including one or more layers leading to the item and its heading, and returns it to the user as the answer to the question.

As shown in FIG. 5, when the question T1 is input, the answer generation unit 27 shown in FIG. 1 generates an answer T3 based on the answer output information matching the input question T1 and presents it on the screen G. .. In the following, the answer generation process will be described more specifically.

First, the natural language processing unit 21 shown in FIG. 1 reads, for example, question T1 "I brought a pension notebook as an identity verification document. Is it valid if the address is not stated?", Performs morphological analysis, and cuts out each word. .. The natural language processing unit 21 is, for example, "identification document", "identity", "confirmation", "document", "confirmation document", "identity verification", "pension notebook", "pension", "notebook", "notebook". Bring ","
Cut out at least one or more words such as "address", "statement", and "valid". At this time, the natural language processing unit 21 may use the technical term dictionary DB 12 and the synonym dictionary DB 13 as necessary.
May be referred to. In addition, the natural language processing unit 21 executes parsing by referring to the part of speech information based on the morphological analysis result.

Next, the weighting setting unit 23 calculates the frequency of appearance of the words included in the question T1 and sets the weighting for each word based on the calculated frequency of appearance of the words. The weighting setting unit 23 is a cut-out "identity verification document", "identity", "confirmation", "document", "confirmation document", "identity verification", "pension notebook", "pension", "notebook". , "Bring", "Address", "Description", "Valid", etc. are set. The weight setting unit 23 may, for example, set a larger weight for frequently occurring words in one or more documents. In addition, the weight setting unit 23 may set a larger weight for words that appear less frequently in one or more documents. Further, the weight setting unit 23 may set a larger weight for a word that appears frequently in a shorter document. The information indicating the weighting for each set word may be recorded as the document information BI in the recording unit 11.

The answer generation unit 27 asked the words "identification document", "identity", and "confirmation" contained in the entered question T1 "I brought my pension certificate with the identity verification document. Is it valid if the address is not stated?""
, "Documents", "Confirmation documents", "Identity verification", "Pension notebook", "Pension", "Notebook", "Bring", "Address", "Description", "Valid", etc. Based on the weighting set by the setting unit 23, a sentence suitable as an answer to the input question T1 is extracted from the document information BI shown in FIG. The answer generation unit 27 may search for each constituent unit (edition, chapter, section, section, sub-item, etc.) included in the document information BI, and extract a sentence suitable as an answer.

The answer generation unit 27, for example, has each text included in the structural unit located at the end of the hierarchical structure of the document data recorded as the document information BI based on the weight set for each word included in the question T1. Of the data, text data containing more words with a larger weight is extracted as text data more suitable for question T1. The matching process is not limited to the above, and various matching processes can be adopted.

Then, the answer generation unit 27 extracts text data adapted to the question T1 extracted from the document information BI, for example, "Management Chapter 1 ...", "Foreign Exchange Chapter 6 ...", and "Management Chapter 5" shown in FIG.
Generate answer T3 with a template such as "Show top candidates. Which one do you want to see in detail?" To "Chapter ...".

According to this configuration, by generating an answer including a template, a natural dialogue can be performed in a format similar to a conversation with a human being.

As shown in FIG. 5, when a plurality of answers to the input question are generated, the output control unit 31 shown in FIG. 1 controls to preferentially output a more suitable answer as the answer to the question. The output control unit 31 displays, for example, the top three answers that are more suitable as answers to the question T1, "Management Chapter 1 ...", "Forex Chapter 6 ...", and "Management Chapter 5 ...". It may be controlled to output in order from the top on the screen G of.

According to this configuration, the user can easily grasp the answer that is more suitable as the answer to a certain question, so that the convenience of the user is further improved.

Then, as shown in the selection result T5 of FIG. 5, on the screen G, in the answer T3, "((
When "1) Management Chapter 1 ..." is selected, the output control unit 31 will perform "" as shown in the answer details T7.
(1) Management Chapter 1 ... ”is configured to be output. Specifically, the output control unit 31 is based on the selected number, and the structural unit (edition, chapter, section, section,) corresponding to the number.
The explanation of the sub-item in (sub-item, etc.) is extracted from the document information BI and output. Regarding the mode of selection in answer T3, the number for which details are to be confirmed in answer T3 may be selected by a click operation on the screen G, or the number for which details are desired to be confirmed may be entered as text in the input field EB. You may choose.

In addition, the answer generation unit 27 answers T a hot word, that is, a specific word set in advance.
In at least one of 3 and the answer details T7, the answer may be generated so as to have a display form different from that of the other words. As shown in FIG. 5, the answer generation unit 27 makes a predetermined specific word (for example, a person, a document, an identity verification document, etc.) in red in at least one of the answer T3 and the answer detail T7. Generate answers so that the other words are in the black. Also,
The answer generation unit 27 may generate an answer so that the word included in the question T1 has a display form different from that of the other words in at least one of the answer T3 and the answer detail T7.

With this configuration, the user can easily grasp where a specific word is used in the answer.

Further, when the answer includes a synonym of the word included in the question, the answer generation unit 27 generates the answer so that the user can identify the synonym in the generated answer. For example, the answer generation unit 27 emphasizes the synonym (dwelling) as "dwelling" in the generated answer when the synonym of the word (address) included in the question T1 is included in the answer T3 or the answer detail T7. Generate an answer after displaying it.

According to this configuration, the user can easily identify synonyms of the words included in the question in the answer.

Further, the dialogue result management unit 29 shown in FIG. 1 stores the dialogue result including the question, the answer to the question, and the response to the answer in the dialogue log database 4. And answer generation unit 2
7 learns the weighting of matching between the question and the answer based on the dialogue result stored in the dialogue log database 4.

FIG. 6 is a diagram showing an example of a screen of the display unit of the user terminal device 6 when the user selects the answer (1) and then reselects the answer (2). As shown in FIG. 6, when the favorite button HI is selected after the text information of the answer T9 is input, the dialogue result management unit 29 receives the dialogue result management unit 29.
From the start to the end of the dialogue, that is, the text information of T1 to T9 shown in FIG. 6 is stored in the dialogue log database 4 as the dialogue result. Here, for user U, the final answer (2) (
Answer T9) is an answer more suitable for question T1. Therefore, the dialogue result management unit 29 collectively manages the question T1 to the first answer (1) and the last answer T9, and the answer generation unit 27 manages the dialogue log when the favorite button HI is selected. The dialogue result stored in the database 4 learns that the answer (2) is a more suitable answer to the question T1 for the user U. In this way, the answer generation unit 27 learns the weighting of matching between the input question and the answer by repeatedly learning based on the dialogue result managed for each dialogue.

According to this configuration, it is possible to learn the weighting of matching between the question and the answer, so that the accuracy of the answer can be improved.

As described above, according to the first embodiment of the present invention, when a question is input, the answer generation unit 27 is input based on the set weighting corresponding to the word included in the input question. Generate answers to questions. Therefore, it is possible to improve the work efficiency in the organization, improve the convenience of the user, and effectively utilize the manual or the like.

(Modification example)
In the above explanation, the answer generation unit 27 matches the input question with the text data included in each item of the recorded document information, and extracts a plurality of items related to the question. The processing of the generation unit 27 is not limited to this. The answer generation unit 27
The input question may be matched with the text data and headings included in each item of the recorded document information, and a plurality of items related to the question may be extracted.

Further, in this case, the item related to the question extracted by matching with the heading and the item related to the question extracted by matching with the text data included in each item of the recorded document information by the weighting setting unit 23. You may change the importance of. For example, the weighting setting unit 23 sets the importance of the items extracted based on the heading to be higher than the importance of the items extracted based on the text data of the structural unit located at the end of the hierarchical structure. It may be something to do.

Supplementally, as an answer to the question, a simple explanation may contribute to the convenience of the user rather than a detailed explanation. Therefore, depending on the question, the heading may lead to a more appropriate answer than the text data contained in each item. In the answer generation unit 27 of the modified example, items related to the question can be appropriately extracted by matching not only the text data but also the headline.

<Second Embodiment>
FIG. 7 is a diagram showing an example of a screen of the display unit of the user terminal device 6 when the content of the input question is unclear. As shown in FIG. 7, when the content of the input question T11 “account” is unclear, the answer generation unit 27 shown in FIG. 1 wants to ask, for example, whether he / she wants to ask about opening an account or canceling the account. If it is unclear and unclear, a predetermined answer T13 "Please tell me the details of the question" is generated so as to prompt a more specific question for the entered question. You may. Specifically, the answer generation unit 27 comprehensively evaluates the number of characters of the input question, the search score for the question, the number of documents extracted for the question, and the like, and the content of the input question. Determine if is unclear.

As a result of being prompted by the user U to ask a more specific question in this way, the additional question T1
5 By inputting "open an account", the output control unit 31 selects a plurality of "chapter" that are more suitable for "open an account" among the structural units included in the document information BI, as shown in the answer T17. It is presented on screen G, and one or more "sections" that are more suitable are presented among the plurality of "chapter" presented. Then, the output control unit 31 may be configured to finally present a more suitable sentence extracted from this “section” on the screen G.

According to this configuration, the user can narrow down the search location of the document information BI more specifically.
Extraction accuracy can be improved.

As described above, according to the second embodiment of the present invention, when the content of the input question is unclear, the answer generation unit 27 generates a predetermined answer to the input question. Therefore, since it is possible to prompt the user for a more specific question, ambiguity of the question can be avoided, and the accuracy of the answer to the question can be improved.

<Third Embodiment>
FIG. 8 is a diagram showing an example of a screen of a display unit of a terminal device when a question and an answer to the question are repeated a predetermined number of times. As shown in FIG. 8, the voice dialogue management unit 35 shown in FIG.
Manages so that when a question and an answer to the question are repeated a predetermined number of times with a certain user U, a voice dialogue is possible with the user U. For example, as shown in the texts T21 to T24 of FIG. 8, the voice dialogue management unit 35 interacts with the operator shown in FIG. 1 via a telephone when the question and the answer are executed twice, for example, with the user U. Manage telephone lines to be accessible as possible. The predetermined number of times is not limited to two times.
It may be once or three times or more.

When the question and answer are executed twice with the user U, for example, and the user U receives a telephone request T25 "Please call me", the output control unit 3 shown in FIG.
Reference numeral 1 may be configured to display a message T26 for accepting a telephone call on the screen G.

The telephone dialogue is performed by the user terminal device 6 shown in FIG. 1 and the operator terminal device 8 of the operator.
It may be executed with or via another communication device different from the user terminal device 6 and the operator terminal device 8. The voice information of the user and the operator is collected by the user terminal device 6, the operator terminal device 8, and the other communication device, and transmitted to the dialogue management server 1. In addition, it should be noted
The screen G may be displayed in common on the display unit of the user terminal device 6 and the display unit of the operator terminal device 8. The operator can answer by telephone based on the dialogue on the screen G up to the dialogue by telephone.

Here, the voice recognition processing unit 33 of the dialogue management server 1 shown in FIG. 1 recognizes the voice information of the voice dialogue between the user and the operator, and the dialogue result management unit 29 recognizes each user as the dialogue result. You may manage the registration or update of voice information.

As described above, according to the third embodiment of the present invention, when the question and the answer to the question are repeated a predetermined number of times with the user, the voice dialogue management unit 35 has a voice dialogue with the user. Manage to be possible. Therefore, by preventing the user from easily calling the operator, the number of calls to the operator is reduced, and the efficiency of work in the organization can be improved.

However, the operation of the voice dialogue management unit 35 is not limited to this, and the user's telephone may be accepted before the start of the dialogue, or both the voice dialogue by telephone and the question by character input may be accepted. ..

<Fourth Embodiment>
FIG. 9 is a diagram showing an example of a screen of the display unit of the user terminal device 6 when one or more related keywords are presented as answers to the input question. As shown in FIG. 9, the answer generation unit 27 shown in FIG. 1 serves as an answer to the input question T31 “opening an account”, for example.
Answer T including one or more related keywords "procedure", "required documents", "proxy" and "minor" recorded in advance in the recording unit 11 in association with the word "open an account".
Generate 33. Here, the related keyword of a certain word acquires the co-occurrence information of the word, and
It is generated by weighting the co-occurrence information. Specifically, "procedure", "required documents", "proxy", "minor", etc. as related keywords of "account opening" are generated by weighting the co-occurrence information of "account opening". It is a thing. Then, questions including the word "open an account", and "procedures", "required documents", "proxy" and "proxy" for the questions.
By managing the dialogue results including the answers including "minors" and the like, the accuracy of generating related keywords can be improved, so that the accuracy of the answers to the questions can be improved. The co-occurrence information of one word and another word is obtained from all the text data of the document in the structured format.

As described above, according to the fourth embodiment of the present invention, the ambiguity of the question can be avoided by presenting the keyword associated with the word included in the question in advance and letting the user select it, and the answer to the question can be answered. The accuracy can be improved.

<Fifth Embodiment>
In the fifth embodiment, when the content of the question input by the user is unclear, a hierarchical structure consisting of, for example, chapters, sections, sections, sub-items, etc. is provided so as to prompt a more specific question. The answer is output step by step based on the document data that you have. For example, in the dialogue management system according to the fifth embodiment, answer output information is generated based on the items in the knitting layer (first layer) for the input question, and is returned to the user U as the first answer to the question. .. After that, answer output information is generated for the further input question based on the items in the hierarchy (second hierarchy) such as chapters, sections, sections, or sub-items, which are lower than the edition hierarchy, and the question is asked. Is returned to user U as the second answer of.

A fifth embodiment will be described with reference to FIGS. 10 and 11. 10 and 11 are diagrams showing an example of the screen G of the display unit of the user terminal device 6 when the content of the input question is unclear. As shown in FIG. 10, the answer generation unit 27 shown in FIG. 1 determines the number of characters of the input question T41 “unify accounts”, the search score for the question, the number of documents extracted for the question, and the like. By comprehensively evaluating, it is determined whether or not the content of the input question T41 "unify accounts" is unclear. When it is unclear and unclear what kind of content the answer generation unit 27 wants to ask about the unification of accounts, the answer generation unit 27 responds to the input question.
In order to prompt a more specific question, answer output information is generated based on the "deposit edition", "common edition", etc. in the edition hierarchy for the entered question, and the user U is used as the answer T43 of the question.
Return to. In addition, the answer generation unit 27 generates, for example, when the answer T43 is output, the predetermined answer T42 "Which volume does" OOOO "correspond to? Please tell me a little more." You may reply.

After that, the answer generation unit 27 generates answer output information for the further input question T44 based on "basic matters", "handling procedures", etc. in the section hierarchy which is lower than the first part. It is returned to the user U as the answer T46 of the question. The answer generation unit 27 is, for example,
When the answer T46 is output, the predetermined answer T45 "There are multiple candidates for" OOOO ". Which of the following items is closest to?" May be generated and answered.

As shown in FIG. 11, the answer generation unit 27 responds to the further input question T47.
Answer output information is generated based on each item in each level of chapter, section, and section, and the answer T of the question is generated.
Return to the user as 49. Here, the answer generation unit 27, for example, asks the question T41 shown in FIG.
And T44, and the top three answers that are more suitable as the final answers to T47 may be generated on the screen G in order from the top. Answer generation unit 2
7 is, for example, the question T41 in FIG. 10, which was first input when the answer T49 was output.
You may generate and reply to "Unify accounts".

As described above, according to the fifth embodiment of the present invention, answers are presented step by step according to the content of the question from the user. Therefore, since the answer candidates can be efficiently narrowed down, the accuracy of the answer can be improved, and the desired answer can be quickly presented on the screen.

<Sixth Embodiment>
In the sixth embodiment, the question from the user, the answer to the question, and the evaluation for the answer are stored in association with each other as the dialogue result, and the weighting of matching between the question and the answer is learned based on the stored dialogue result. .. A sixth embodiment will be described with reference to FIGS. 12 to 14.

FIG. 12 is a screen example showing an example of a dialogue including a question and an answer of the user U in the display unit of the user terminal device 6. As shown in FIG. 12, the answer generation unit 27 asks the input question T5.
For 1, the answer output information is generated based on each item in each layer of edition, chapter, section, and section.
It is returned to the user U as the answer T53 of the question. Further, when the answer T53 is output, the predetermined answer T52 "Is there an answer in this for" OOOO "?" May be generated and replied.

FIG. 13 is a screen example showing an example of the dialogue that follows after the user U selects the first answer “(1): Calculation ...” in the answer T53 on the screen G shown in FIG. Note that FIG. 12
Regarding the mode of selection in the answer T53, the number in the answer T53 for which the details are to be confirmed may be selected by a click operation on the screen G, or the number for which the details are to be confirmed may be entered as text in a user input field (not shown). You may select by doing.

As shown in FIG. 13, when the user U selects the first answer "(1): Calculation ..." in the answer T53 on the screen G shown in FIG. 12, the answer generation unit 27 shows the answer details T55. In addition, answer output information is generated so that the specific contents of "(1): Calculation ..." can be output. Specifically, based on the selected number, the answer generation unit 27 provides a description of the sub-items in the structural unit (edition, chapter, section, section, sub-item, etc.) corresponding to the number in FIG. Document information to show B
Extract from I to generate answer output information. The answer generation unit 27 may generate answer output information so that the hot word, that is, a specific word set in advance, is displayed in the answer detail T55 in a form different from that of other words. As shown in FIG. 13, the answer generation unit 27 uses the words included in the question T51 shown in FIG. 12 (for example, the trial expense reimbursement port and the establishment).
Answer details In T55, the answer may be generated so as to have a display form different from that of other words. In addition, the answer generation unit 27 generates answer output information so that a specific preset word (for example, a trial expense reimbursement opening and opening) is in the red, and other words are in the black. May be good.

Further, the answer generation unit 27 outputs, for example, the answer T56 (evaluation request) for prompting the evaluation of the answer to the question input from the user U after the answer details T55. Answer T5
6 is, for example, "Good" (with the text "How about the evaluation of the answer?"
It is composed of a stamp ST1 associated with a specific meaning of "good" and a stamp ST2 associated with a specific meaning of "bad" (bad).

The "stamp" is image information for dialogue input via the network N. The stamp represents, for example, the user's emotions, intentions and intentions, and a message to be conveyed by an illustration, and may include text as well as image information. The recording unit 11 shown in FIG. 1 records the stamp in association with a specific meaning. For example, stamps are recorded in association with the specific meanings of "Good" (good) and "Bad" (bad). Further, the stamp may include an illustration in association with a score indicating the accuracy of the answer, for example, text information indicating 100 points, 75 points, or the like. The evaluation is not limited to selecting a stamp. For example, the "Good" button or "Ba" can be selected in answer T56.
When the "d" button is included, the user may execute the evaluation input by selecting one of the buttons. Further, the user may execute the evaluation input by inputting the text information of "Good" or "Bad" from the text input field (not shown) after the answer details T55 is output. The evaluation may include not only two patterns of "Good" (good) or "Bad" (bad), but also three or more patterns such as very good, good, normal, bad, and very bad. ..

Next, the dialogue result management unit 29 shown in FIG. 1 stores the dialogue result including the question from the user U, the answer to the question, and the input stamp (evaluation) in the dialogue log database 4. Hereinafter, the processing of the dialogue result management unit 29 will be specifically described.

As shown in FIG. 13, when the stamp ST1 including "Good" indicating a good evaluation for the answer is selected by the user U, or when the dialogue end button (not shown) is selected after the stamp ST1 is selected. The dialogue result management unit 29 stores the text information of T51 to T56 shown in FIGS. 12 and 13, and the stamp ST1 as the dialogue result in the dialogue log database 4 from the start to the end of the dialogue. In this way, the dialogue result management unit 29 evaluates that the dialogue management server 1 is a more suitable answer to the input question by managing (associating) the dialogue from the start to the end. The presented answers and the answers actually selected by the user (desired by the user) are collectively managed.

FIG. 14 is a diagram showing an example of a dialogue log management screen in the display unit of the operator terminal device 8 according to the embodiment of the present invention. For example, the information displayed on the dialogue log management screen LG corresponds to the information stored in the dialogue log database 4 by the dialogue result management unit 29. As shown in FIG. 14, the dialogue log database 4 contains a dialogue log ID, a user ID, a date and time when the dialogue log is stored, a question first entered in the dialogue, and a final answer to the question. And the number of the answer (for example, the number (1), (2) (3) of the answer T53 in FIG.
) And the evaluation of the answer are recorded in association with each other. The dialogue log database 4 may store not only the answer to the first question from the user but also one or more answers for each question input by the user, and the dialogue log management screen LG is input by the user. One or more answers for each question asked may be output.

The dialogue result management unit 29 stores the dialogue result including the stamp, and adjusts the search score for the dialogue result based on the evaluation of the answer. That is, the dialogue result management unit 29 repeatedly records the dialogue result together with the stamp reflecting the evaluation of the input answer, and records a plurality of dialogue results (a set of questions and answers) according to the evaluation of the answer, and the search score for the question. To raise. For example, in FIG. 14, the dialogue result currently having the dialogue log ID "7350", that is,
The set of question "About opening a new account" and answer "Chapter 1 of the card edition ..." is ranked second from the top, but at another time, for example, the user enters a question about "Opening a new account". , If you enter the evaluation of "Good" for the answer "Card Chapter 1 ..." similar to the above, the display position of the above set is controlled to be higher on the dialogue log management screen LG (here, It is controlled to be the highest level). On the other hand, when a certain question and answer set is evaluated as "Bad", the set is controlled to be hidden or the display position of the set is controlled to be lower on the dialogue log management screen LG. Or

The answer generation unit 27 learns the weighting of matching between the input question and the answer to the question based on the evaluation associated with the stamp stored in the dialogue log database 4 as the dialogue result.

For example, the answer generation unit 27 asks the user U to answer T51 to the question T51 shown in FIG.
Of the multiple answers included in 53, the first answer "(1): Calculation ..." is analyzed to be a more suitable answer, and the weighting of the answer "(1): Calculation ..." is other. Control to increase compared to the answer. For example, the answer generation unit 27 gives weighting to the answer "(1): calculation ..." and other answers "(2) calculation ...""(3) calculation ..." included in the answer T53. The weighting may be lowered. In this way, the answer generation unit 27 learns the weighting of matching between the input question and the answer by repeatedly learning based on the dialogue result managed for each dialogue.

As described above, according to the sixth embodiment of the present invention, the question from the user, the answer to the question, and the evaluation of the answer are stored in association with each other as the dialogue result, and based on the stored dialogue result, the question is stored.
Learn the weighting of matching between questions and answers. Therefore, since the weighting of matching between the question and the answer can be learned, the accuracy of the answer can be improved.

<7th Embodiment>
The dialogue management system according to the seventh embodiment will be described with reference to FIG. FIG. 15 shows the seventh
It is a schematic block diagram (system block diagram) of the dialogue management system which concerns on embodiment. In the seventh embodiment, as shown in FIG. 15, the dialogue management server 1 further includes the exclusion term dictionary DB 14. The exclusion term dictionary DB 14 is a database that stores "excluded terms" whose influence is excluded when matching the input question with the text data included in each item of the document information recorded in the recording unit 11. .. Specifically, the excluded terms include particles, conjunctions, predetermined modifiers, predetermined frequent words, and the like. Here, the predetermined modifier corresponds to a term such as "various" or "various" that does not substantially limit the meaning of the word to be employed. In addition, a predetermined frequent word is a term that frequently appears among all the terms used in a structured document of a specific format.
In a Japanese document with tens of thousands of words, the top 10 terms correspond to this. For example, terms such as "confirmation" are often mentioned as frequent words.

The answer generation unit 27 generates answer output information excluding the influence of the excluded term when matching the input question with the text data included in each item of the recorded document information.

Since the dialogue management server 1 according to the seventh embodiment has the above-described configuration, the validity of the answer to the question can be enhanced. Supplementally, in general, if the text data contained in each item of recorded document information contains many terms that frequently appear in the entered question, it is likely to be a valid answer to the question. However, among all the terms in the document information, terms that appear more frequently than other terms may be used in arbitrary sentences, which may rather reduce the validity of the answer. Therefore, it may be possible to improve the validity of the answer to the question by excluding the influence of these terms at the time of matching.

<8th Embodiment>
In the eighth embodiment, the document in the structured format according to the first to seventh embodiments is the text data included in the items located at the end of the hierarchical structure, and one or more documents leading to each item. The "unit information" is a combination of information that can specify a hierarchy and headings of one or a plurality of layers leading to each item. Specifically, the data shown in FIG. 3 is defined as one unit information, and these sets are defined as the above-mentioned specific format structured document.

The dialogue management system according to the eighth embodiment will be described with reference to FIG. FIG. 16 shows the eighth
It is a schematic block diagram (system block diagram) of the dialogue management system which concerns on embodiment. The dialogue management server 1 according to the eighth embodiment further includes a reverse document frequency DB 24. The reverse document frequency DB 24 is a database that stores the reverse document frequency of words used in a structured document.
Here, the "reverse document frequency" is an index indicating the rarity of a word in the entire document, and the total number maxD of all unit information in a structured document is calculated by the number of unit information including the word DFreqs. It is the value divided.

The weighting setting unit 23 calculates the frequency of appearance of words included in the character information in the unit information, and sets the weighting for each word based on the calculated frequency of appearance of words. Further, the weight setting unit 23 corrects the weight based on the reverse document frequency.

The answer generation unit 27 extracts unit information including a predetermined number or more of words included in the input question as a candidate answer, and generates answer output information based on the candidate answer. Here, the answer generation unit 27 calculates the degree of similarity between the candidate answer and the question based on the weight set by the weight setting unit 23. Then, the answer generation unit 27 generates candidate answers having a high degree of similarity as answer output information. The algorithms of TFIDF Similarity and BM25 Similarity can be used to calculate the similarity.

With the above configuration, the dialogue management server 1 according to the eighth embodiment generates an answer having many words included in the question, so that a highly valid answer can be provided.

Further, the dialogue management server 1 according to the eighth embodiment can calculate the similarity after correcting the weight of the word based on the reverse document frequency. This makes it possible to reflect the rarity of the word in the entire document. Here, the rarity of a word in the entire document tends to indicate that the word is characteristic, and is likely to indicate the correct answer in the search of text data. Therefore, the answer generation unit 27 generates appropriate answer output information for the input question by using the similarity corrected by the reverse document frequency.

When the exclusion term dictionary DB 14 shown in FIG. 15 is combined with the dialogue management server 1 according to the eighth embodiment, the number of unit information is not counted for the exclusion terms when extracting candidate answers. ..

<9th embodiment>
In the ninth embodiment, with respect to the structured format document according to the first to seventh embodiments, the text data of the explanation and the information and the hierarchy for identifying one or more layers leading to the explanation. The "unit information" may be associated with the text data representing the heading of. For example, the data shown in FIG. 3 (the part surrounded by "extracted in a specific XML format")
Is defined as one unit information, and a set containing a plurality of unit information is defined as the above-mentioned specific format structured document. The answer generation unit 27 matches the text data of the question text received by the reception unit 19 with the text data included in each unit information, extracts the unit information related to the question text, and extracts the extracted unit information. Generate answer candidates (answer output information) based on the heading and description corresponding to. For some questions, headings may lead to more appropriate answers than text data contained in document building blocks (items) (eg, editions, chapters, sections, sections, sub-items, etc.). The dialogue management server 1 according to the embodiment makes it possible to appropriately extract items related to a question by matching not only text data of each item but also headings.

The dialogue management system according to the ninth embodiment will be specifically described with reference to FIGS. 17 to 27. FIG. 17 is a schematic configuration diagram of the dialogue management system according to the ninth embodiment. The answer generation unit 27 of the dialogue management server 1 calculates a search score using, for example, an index such as TF (Term Frequency) indicating the frequency of occurrence of words or IDF (Inverse Document Frequency) indicating the frequency of reverse documents. .. The TF indicates, for example, how many specific words appear in each structured document. That is, the TF is a value obtained by dividing the frequency of occurrence of words in a structured document by the sum of the frequency of occurrence of all words in a structured document. Specifically, the answer generation unit 27 is TF.
As a method using, a search score is calculated based on the frequency of occurrence (TF) of the extracted words in each unit information, and response output information is generated using the calculated search score.

Further, the IDF is, for example, as described in the eighth embodiment, an index indicating the rarity of a word in the entire document, and the total number of all unit information in the structured document is the number of unit information including the word. It is the value divided by. As a method using TF-IDF, the answer generation unit 27 includes the frequency of appearance of the extracted words in each unit information (TF) and the ratio of the unit information including the extracted words to the total number of unit information (IDF). The search score is calculated based on. For example, the answer generation unit 27 calculates the search score by multiplying the TF and the IDF. The answer generation unit 27 generates answer output information using the calculated search score.

As shown in FIG. 17, the dialogue management server 1 according to the ninth embodiment is the first synthetic word dictionary DB1.
2A, second synthetic word dictionary DB12B, subordinate conceptual word dictionary DB44, short talk DB
46 and are further provided. The first synthetic word dictionary DB12A and the second synthetic word dictionary DB
12B is an aspect of the technical term dictionary DB 12 shown in FIGS. 1, 15 and 16.

The first compound word dictionary DB12A is a database in which a first compound word obtained by synthesizing another word with a word of a predetermined noun is registered. In the first synthetic word dictionary DB12A, for example, "confirmation" (another word) is synthesized with "at the time of transaction" (word of a predetermined noun), or "confirmation" (word of a predetermined noun) is "transaction". "Confirmation at the time of transaction" (first composite word), which is a composite of "time" (other words), is registered. In the first synthetic word dictionary DB12A, "confirmation at the time of transaction" is not associated with "at the time of transaction" or "confirmation".

An example of a method for generating answer output information when the first synthetic word dictionary DB12A is used will be described.
For example, the answer generation unit 27 is a compound word among words extracted from a question sentence (for example, "I want to find out about confirmation at the time of transaction") (for example, "confirmation at the time of transaction", "at the time of transaction", "confirmation", etc.). When it is determined that "confirmation at the time of transaction" is registered in the first compound word dictionary DB12A, the first search score is calculated based on the frequency of appearance of each word extracted from the question sentence in each unit information, and the first search score is calculated. The second search score is calculated based on the frequency of occurrence of synthetic words (for example, "confirmation at the time of transaction") in each unit information. Then, the answer generation unit 27 generates the answer output information using both search scores.

The second compound word dictionary DB12B is a database in which a word of a predetermined noun and a second compound word obtained by synthesizing another word with the word are registered in association with each other. The second compound word dictionary DB12B is
For example, unlike the first synthetic word dictionary DB12A, "confirmation at the time of transaction" is registered in association with each of "at the time of transaction" and "confirmation".

An example of a method for generating answer output information when the second synthetic word dictionary DB12B is used will be described.
For example, the answer generation unit 27 has determined that the word (for example, "at the time of transaction") extracted from the question sentence (for example, "I want to find out about confirmation at the time of transaction") is registered in the second compound word dictionary DB12B. For each of the words extracted from the interrogative (eg, "at the time of transaction"), other words synthesized into the noun word (eg, "confirmation"), and the compound words (eg, "confirmation at the time of transaction"). , Calculate the search score based on the frequency of appearance in each unit information. Then, the answer generation unit 27 generates the answer output information using the calculated search score.

The subordinate concept word dictionary DB44 contains a predetermined word and a "subordinate concept word" which is a subordinate concept of the word.
Is a database registered in association with. FIG. 18 is a diagram showing an example of subordinate conceptual words. As shown in FIG. 18, in the subordinate concept word dictionary DB44, for example, each of the key key words "account", "confirmation", "processing", "transaction", "periodic", etc., and each key word Subordinate concept words and are registered in association with each other.

The short talk DB 46 is a database in which words and sentences for realizing a short talk between the user and the dialogue management server 1 are registered. For example, a predetermined word and a predetermined answer sentence including the predetermined word are stored. It is associated and registered. As will be described later in FIG. 20B, the short talk DB 46 is registered in association with, for example, the word "curry" and the answer sentence "I like curry" including "curry". Also, in the short talk DB46, for example, when the word "△△△" is input, a sentence template (for example, "... is"" is generated so that the answer sentence "△△△ is good."It's good. ") May be registered. Further, the short talk DB46 includes a predetermined word and
It may be registered in association with a predetermined stamp which is image information for dialogue.

Further, the short talk DB 46 may be a database in which a plurality of unit information prepared for the short talk is registered. In this case, the unit information is extracted based on the frequency of appearance of the input word and the like, and the answer sentence is generated based on the extracted unit information.

FIG. 19 is a flowchart showing an example of the dialogue management process according to the ninth embodiment. 9th
An example of the dialogue management process according to the embodiment will be described with reference to FIGS. 20 to 27 together with FIG. 20 to 27 are views showing an example of a screen of the display unit of the user terminal device 6. Figure 1
As shown in 9, first, the dialogue management server 1 accepts the input of the question text from the user (
Step S10). Next, in step S12, the dialogue management server 1 matches the text data of the received question text with the text data included in each unit information in the document information BI, and obtains the unit information related to the question text. Extract and search for the document corresponding to the extracted unit information. For example, the dialogue management server 1 matches the words extracted from the interrogative sentence with the words in each document, and when there is a matching document (in the case of No), step S2.
Proceed to 0. On the other hand, if there is no matching document (in the case of Yes), the dialogue management server 1 proceeds to step S14.

In step S14, the dialogue management server 1 determines, for example, whether or not a word included in at least a part of the interrogative sentence is registered in the short talk DB 46. For example, FIG. 20 (
As shown in A), when the question sentence T61 is "ahhh", when the dialogue management server 1 determines that "ahhh" is not registered in the short talk DB46 (No in step S14), "I'm sorry. Generates the predetermined answer T62, "No." On the other hand, as shown in FIG. 20B, when the question sentence T64 is "curry", the dialogue management server 1 determines that "curry" is registered in the short talk DB 46 (step S14).
Yes), generate the prescribed answer T65, "I like'curry'." In this way, for example, when a word included in at least a part of the question sentence is registered in the short talk DB 46, a short talk is performed between the user and the dialogue management server 1. Therefore, even if the user asks a question other than the question related to the business manual, the dialogue between the user and the dialogue management server 1 can be continued without causing a feeling of discomfort.

In step S20, the dialogue management server 1 has, for example, a search score of 8 or 9 when the maximum search score among a plurality of calculated search scores is set to a predetermined score (for example, when the search score is set in the range of 0 to 10. Etc.) and the number of matching documents is a predetermined number (for example, several files to several tens of files) or less. For example, if the dialogue management server 1 has a high maximum search score and the number of matching documents is small (in the case of Yes), the process proceeds to step S22. That is, when the maximum search score is high and the number of matching documents is small, the dialogue management server 1 determines that the accuracy of the answer to the question sentence is relatively high, and displays the answer candidates (step S22). FIG. 21 (
As shown in A), when the dialogue management server 1 determines that the accuracy of the answer candidate for the question sentence T67 "about the double stripe account" is relatively high, the dialogue management server 1 generates and outputs the answer candidate T68 corresponding to the question sentence T67.

On the other hand, the dialogue management server 1 proceeds to step S24 when the maximum search score is a relatively low score or when the number of matching documents is large (No). That is, when there are many documents whose matching accuracy (search score) is relatively low as answer candidates, the dialogue management server 1 needs to narrow down the documents as answer candidates under other conditions. ..

In step S24, the dialogue management server 1 has the same document configuration unit "○○" in, for example, the top ten files of documents with high matching accuracy among a plurality of documents matched to the interrogative sentence. , The number of times of listening back (for example, the number of times the process returned to step S10 "input question text" after the processing of step S36 or step S38 described later was executed).
Is determined, for example, once or more. If Yes in step S24, the process proceeds to step S42. As shown in FIG. 21 (B), the dialogue management server 1 is, for example, a question sentence T7.
0 When "Tell me about" How to process when you want to open a basic account "" is entered, other probable editions are included as constituent units in order to narrow down the "edition" which is one of the constituent units of the document. Propose a document. The dialogue management server 1 generates a response candidate T71 that proposes, for example, "(1) Loan support system". In the above form, "" is one of the structural units of the document.
We are narrowing down the answer candidates based on "Hen", but not limited to this, we answer by "chapter", "section", "section", "small item", etc., which are the structural units of other documents. Candidates may be narrowed down, or a combination of answer candidates may be carried out by combining two or more constituent units of the constituent units.

On the other hand, if No in step S24, the process proceeds to step S26. Step S2
In 6, the dialogue management server 1 determines whether or not, for example, the maximum search score is equal to or higher than a predetermined score among the plurality of documents matched to the question text. The "predetermined score" in step S24 may be the same as the "predetermined score" in step S20.
The score may be higher than or lower than the "predetermined score" in step S20. In the case of Yes in step S26, the dialogue management server 1 proceeds to step S40, and for example, among a plurality of documents matched to the interrogative sentence, the document having a high search score is preferentially displayed as an answer candidate.

On the other hand, if No in step S26, the process proceeds to step S28. Step S2
In 8, the dialogue management server 1 determines whether or not the number of times of listening back is, for example, three or more times. When the number of times of listening back is three or more (in the case of Yes), the dialogue management server 1 proceeds to step S40 and displays the answer candidates. On the other hand, when the number of times of listening back is less than three (N)
In the case of o), the dialogue management server 1 proceeds to step S32. It should be noted that the set number of times "three times" regarding the number of times of listening back is merely an example, and the set number of times may be another number of times.

In step S32, the dialogue management server 1 determines, for example, whether or not each of one or a plurality of words extracted from the interrogative sentence by natural language processing is registered in the subordinate concept word dictionary DB 44 shown in FIG. .. If Yes, the dialogue management server 1 is in step S3.
If the result is No, the process proceeds to step S34. In step S34, the dialogue management server 1 determines whether or not a proposal (for example, display of answer candidates) has been implemented for the same question sentence in the past hearing. In the case of Yes, the process proceeds to step S40, and in the case of No, the process proceeds to step S36.

In step S36, the dialogue management server 1 calculates a score based on the frequency of appearance of words extracted from the question sentence in the text data included in each heading, and generates answer output information using the calculated score. For example, the dialogue management server 1 is TF or TF-IDF.
The proposed word is calculated by using a method such as. The "proposal word" means, for example, a word extracted from the text data included in each heading in the unit information including the word extracted by morphological analysis of the text data of the interrogative sentence. More specifically, the dialogue management server 1 first
For example, the document ID of the document including the extracted word by morphological analysis of the question sentence T73 shown in FIG. 22 (A) is acquired. The document ID is an identification number that uniquely identifies all the documents that make up all the explanations. Then, the dialogue management server 1 refers to the "heading word list" as shown in FIG. 22B, identifies each word used for the heading of the document based on the document ID, and sets each word to the specified word. On the other hand, TF or TF-IDF is calculated. Next, the dialogue management server 1 is TF or TF-.
Words are sorted in order of score based on IDF, words included in the question sentence are excluded, and then, for example, the top 15 words are output as proposed words. Then, as shown in FIG. 22 (A), the dialogue management server 1 generates an answer candidate T74 including the output proposed word. If the proposed word does not exist, the process proceeds to step S40.

As shown in FIG. 22 (A), when the option S1 including "there is nothing that seems to be related" included in the answer candidate T74 is selected, the process returns to step S10, and the same question text as the current question text T73 is input. It will be in the state of being. That is, in step S10 returned from step S36, the same question text as the question text T73 is input again. In this case, the number of times of listening back increases once.

In step S38, the dialogue management server 1 generates answer output information including subordinate conceptual words associated with one or a plurality of words extracted by morphological analysis of the text data of the interrogative sentence. For example, the dialogue management server 1 refers to the subordinate concept word dictionary DB44 shown in FIG. 18 with respect to the word (for example, “account”) extracted by morphological analysis of the text data of the question sentence T76 shown in FIG. Generates answer output information that includes subordinate terms associated with the word "account" (eg, "deposit account", "general account", etc.).

As shown in FIG. 23, when the option S3 "not in the candidate" included in the answer candidate T77 is selected, the process returns to step S10, and a new question sentence is input. afterwards,
When the dialogue management server 1 matches the text data of the interrogative text with the text data included in each unit information and generates the answer candidate, the word included in the answer candidate generated in the past (for example). , "Deposit account", "general account", etc.), and then generate answer candidates. In this case, the number of times of listening is increased by one. Further, when the option S5 of "I don't know" included in the answer candidate T77 is selected, the process returns to step S10, and the same question text as the current question text T76 is input. That is, in step S10 returned from step S38, the same question text as the question text T76 is input again. In this case, the number of times of listening back increases once. Further, normally, the process proceeds to step S40 through step S34.

(Characteristic)
As described above, the dialogue management server 1 according to the ninth embodiment of the present invention includes a recording unit 11, a reception unit 19, and a response generation unit 27. The recording unit 11 has a structured format having a plurality of unit information in which the text data of the explanatory text, the information for identifying one or a plurality of layers leading to the explanatory text, and the text data representing the heading of the hierarchy are associated with each other. Document data (document information BI)
To record. The reception unit 19 receives the text data of the question text from the user. The answer generation unit 27 matches the text data of the question text received by the reception unit 19 with the text data included in each unit information recorded in the recording unit 11 and extracts the unit information related to the question text. Then, the answer output information based on the heading and the explanation corresponding to the extracted unit information is generated.

Here, the dialogue management server 1 calculates the score based on the frequency of appearance of the words extracted by the answer generation unit 27 by morphological analysis of the text data of the interrogative sentence in each unit information. Then, the answer generation unit 27 generates the answer output information using the calculated score. Therefore, the explanatory text that is highly likely to be related to the interrogative text is output. For more details,
In the dialogue management server 1, the answer generation unit 27 morphologically analyzes the text data of the interrogative sentence to extract the word, the frequency of appearance in each unit information, and the unit information including the extracted word.
The score is calculated based on the ratio of the unit information to the total number. Therefore, when the word included in the question sentence is characteristic, the explanation sentence that is likely to be related to the word is output.

Therefore, by reflecting the manual or the like in the content of the unit information, it is possible to improve the work efficiency in the organization and improve the convenience of the user. Moreover, since the answer output information is generated based on the unit information, the utilization of the business manual can be promoted.

The dialogue management server 1 according to the ninth embodiment includes a first synthetic word dictionary DB12A and / or a second synthetic word dictionary DB12B. Therefore, by registering the compound word in advance, the user can configure the search score for the registered compound word to be high. In other words, it is possible by registering synthetic words so as to induce operational efficiency.

The dialogue management server 1 according to the present embodiment includes a subordinate concept word dictionary DB44. Therefore, as described above, it is possible to narrow down the words of the upper concept included in the question sentence to the information of the lower concept, and it is possible to induce the improvement of work efficiency.

The dialogue management server 1 according to the ninth embodiment has a function of the answer generation unit 27 to generate answer output information including words extracted from the question sentence in the text data included in each heading. Specifically, when the answer generation unit 27 determines that the extracted word by morphological analysis of the text data of the question sentence is not registered in the subordinate conceptual word dictionary, it is a word extracted by morphological analysis. Then, answer output information including words extracted from the question sentence in the text data included in each heading is generated. With such a function, it is possible to realize a dialogue that contributes to the convenience of the user.

The dialogue management server 1 according to the ninth embodiment matches the text data of the question text with the text data included in each unit information, extracts the unit information related to the question text, and uses the extracted unit information as the extracted unit information. Generate answer candidates (answer output information) including the corresponding headings and explanations.
Here, depending on the question, the heading may lead to a more appropriate answer than the text data contained in the structural unit of the document. Therefore, the dialogue management server 1 can appropriately extract the items related to the question by matching not only the text data of each item but also the heading.

The dialogue management server 1 according to the ninth embodiment is a TF (Term Fre) indicating the frequency of appearance of words.
In addition to quency) or IDF (Inverse Document Frequency) indicating the frequency of reverse documents, the search score may be calculated using an index such as the length of unit information. In other words, the answer generation unit 27 according to the present embodiment morphologically analyzes the text data of the interrogative sentence to extract the word, the frequency of appearance in each unit information, and the unit of the unit information including the extracted word. The score may be calculated based on the ratio to the total number of information and the length of the unit information, and the answer output information may be generated using the calculated score. That is, the score calculation method of this embodiment is BM.
It includes 25 Similarity and so on.

(Modification example 1)
In the present embodiment, the screen G displayed on the display unit of the user terminal device 6 shown in FIG. 17 can take various display forms.

For example, as shown in FIG. 24, when the user operates an input device (for example, a mouse) (not shown) of the user terminal device 6, one title of a plurality of answer candidates “(404.86)”.
) Management ... Item 1: Move the cursor C to the "Beginning of target transaction" part, and then execute a click operation (for example, right-click). In that case, when the cursor C is placed, the title "
(404.86) Management ... Item 1: Among the documents related to "Beginning of Target Transaction", for example, screen the list L of other documents in which "Edit", "Chapter" and "Section" of the title match. It may be configured to be displayed on G. The content of the list L is not limited to this, and the list L may display, for example, a list L of other documents whose titles "edition" and "chapter" match on the screen G.

Further, among the plurality of items I in the list L, the user can see, for example, the item "1 item: Target transaction".
2. When "Insurance, etc." is selected by executing a click operation (for example, left-click) using the mouse, the title "Management Chapter 24: CRS System Section 2: Overview of CRS New Customer Confirmation Item 1: Target Transactions 2. Insurance Etc. ”(for example, a pdf file of a business manual, etc.) may be configured so that it can be confirmed. When the answer candidate includes an image on the screen G, as shown in FIG. 24, the image is, for example, a thumbnail image TJ of a predetermined size so as not to be an obstacle for visually recognizing the display contents of other answer candidates. It may be displayed in a format.

Further, for example, as shown in FIGS. 25 (A) and 25 (B), the user operates a mouse connected to the user terminal device 6 to obtain one of the answer candidates, "Management ... 1".
If you move the cursor C to the "Item: Overview Beginning" part and then use the mouse to perform a click operation (for example, left-click) to select it, the title will be associated with the title "Administration ... Item 1: Overview Beginning". The document content DC corresponding to the above may be configured to be displayed on the screen G. According to this configuration, the user can easily grasp the content of the document corresponding to a specific title.

Further, for example, as shown in FIG. 26, the user operates a mouse connected to the user terminal device 6 to obtain one title of a plurality of answer candidates, "(621.04) Deposits ... Section 3: When you move the cursor C to the "Deposit 1. Reception" part (when you perform so-called mouse over), at least a part of the contents of the document corresponding to the title "(621.04) Deposit ... Section 3: Deposit 1. Reception" The included snippet image SP may be configured to be displayed on the screen G. According to this configuration, the user can easily grasp at least a part of the content of the document corresponding to a specific title.

(Modification 2)
As shown in FIG. 27, the display unit of the user terminal device 6 can display the screen G on which the contents of the dictionaries shown in FIGS. 1, 15, 16 and 17 can be changed. As shown in FIG. 27, the user first inputs, for example, the name of the dictionary to be changed in the dictionary search field S5 and searches. Then, after inputting the word to be added to the blank S1 in kanji notation or the like, inputting the reading of the word to be added to the blank S3 (for example, katakana notation), and then pressing the add button B1, the dictionary search field S
A new word is added to the dictionary searched in 5. On the other hand, when the user presses the delete button B3, for example, by operating the mouse connected to the user terminal device 6, the pre-registered word corresponding to the delete button B3 is deleted from the dictionary. Further, when the change button B5 is pressed by the user, it is possible to change the content of the pre-registered word corresponding to the change button B5.

The screen G displayed on the display unit of the user terminal device 6 is not limited to a screen in which the contents of each dictionary can be changed, and includes a screen in which the contents of the structured document can be changed. For example, on a screen where the contents of a structured document can be changed, as shown in FIG. 3, the structural units of the document (editions, chapters, sections, sections, sub-items, etc.) with respect to specific XML format data of the structured document. The content can be changed for each.

(Modification 3)
Further, the dialogue management server 1 sets the timing at which the answer candidate (answer output information) for the question text is displayed on the display unit of the user terminal device 6 with the text data of the question text and the text data included in each unit information. It may be configured to be changed according to the matching accuracy (search score) of. For example, the response generation unit 27 of the dialogue management server 1 extracts one or a plurality of unit information for which the search score within a predetermined range is calculated, and the output control unit 31 extracts the unit information according to the number of the extracted unit information. , Adjust the output time of the answer output candidate. When a relatively large number of explanatory texts with similar search scores are matched, the output control unit 31 displays the timing of displaying the answer candidates corresponding to the matched explanatory texts as compared with the case where the number of matching documents is small. Delay the response timing to the question). The response timing is expressed by the following mathematical formula (0), for example.

(Number 0)
Response timing = 1 [second] + 0.2 [second] x (number of unit information with similar search scores)
The response timing is, for example, 2.5 [seconds] at the longest after the question text is input. Also,
When the response timing is 2.0 to 2.5 [seconds], the output control unit 31 displays a filler expression (for example, "um", "um", etc.) before displaying the answer candidate on the screen G. You may.

According to this configuration, the dialogue management server 1 sets the answer candidates for the question text to the user terminal device 6.
The timing of displaying on the display unit of is changed according to the accuracy of matching (search score) between the text data of the question text and the text data included in each unit information. Therefore, the dialogue between the user and the dialogue management server 1 is executed more naturally. In other words, dialogue management server 1
The output of is close to the human reaction. Further, the output control unit 31 displays the filler expression before displaying the answer candidate on the screen G so that the delay in the display of the answer candidate is not due to the abnormality of the dialogue management system in the present embodiment. It is possible to notify.

(Modification example 4)
Further, the dialogue management server 1 may be configured to further include a stop word dictionary in addition to the dictionaries shown in FIGS. 1, 15, 16 and 17. A stop word is a word that is excluded from processing when executing natural language processing, for example, particles, conjunctions, and "various".
And include words such as "to some extent". The dialogue management server 1 can generate answer candidates with high matching accuracy for interrogative sentences by excluding specific words from natural language processing by referring to the stop word dictionary.

(Modification 5)
Further, the answer generation unit 27 of the dialogue management server 1 displays at least one of the icon image corresponding to the questioner or the icon image corresponding to the respondent (dialogue management server 1) on the screen G as an answer candidate for the question text. It may be configured to change according to the matching accuracy of the document for. For example, the answer generation unit 27 determines the difference between the first average score of one or more unit information having a relatively high search score and the second average score lower than the first average score of one or more unit information. Correspondingly, the display form of the icon image corresponding to the respondent associated with the answer candidate is changed. Specifically, when the answer generation unit 27 generates answer candidates based on the top 10 unit information in descending order of search score, the average of the scores of the 1st to 5th unit information is set as the first average score. The average of the scores of the unit information of the 6th to 10th places is calculated as the second average score, and the display mode of the icon image is changed according to the difference between them.

Further, when the icon image corresponding to the respondent is a face image, the answer generation unit 27 may change the display as the quality of the question is improved by continuing the dialogue including the question and the answer. For example, in the example shown in FIG. 28, the character icon K is an expressionless face image (
(See FIG. 29) changes to a smiling face image (see FIG. 30). Further, the degree of improvement in the quality of the interrogative sentence may be expressed by using a color scale or the like instead of or in combination with the change of the face image.

(Modification 6)
Further, the answer generation unit 27 of the dialogue management server 1 is a compound word included in the text data of the interrogative sentence, and is, for example, a compound word registered in the first compound word dictionary DB12A or the second compound word dictionary DB12B. Depending on the number, the display form of the icon image corresponding to the respondent associated with the answer candidate may be changed. The fact that a question sentence including a synthetic word registered in the synthetic word dictionaries DB12A and 12B in advance is input confirms that the quality of the question is improved. Therefore, similarly to the above, for example, when the icon image corresponding to the respondent is a face image, the answer generation unit 27 changes the display as the quality of the question is improved by continuing the dialogue including the question and the answer. You may let it. Similar to the above, as shown in FIG. 28, the icon K of the character may be changed from the expressionless face image (see FIG. 29) to the smiling face image (see FIG. 30). Further, the degree of improvement in the quality of the interrogative sentence may be expressed by using a color scale or the like instead of or in combination with the change of the face image.

(Modification 7)
Further, the dialogue management server 1 according to the ninth embodiment calculates the search score of the unit information by, for example, a method using TF-IDF, but when calculating the search score, the total value of the scores calculated for each item is used. Alternatively, the total value obtained by multiplying the score calculated for each item by the weight may be used. Further, the score having the highest score among the scores for each item may be adopted as the search score. Further, the score having the maximum score among the scores obtained by multiplying the score calculated for each item by the weight may be adopted as the search score. For example, dialogue management server 9
May generate answer output information using the larger score of the score calculated from the explanation and the score calculated from one heading.

(Modification 8)
Further, when the explanatory text is composed of a plurality of texts, the dialogue management server 1 according to the ninth embodiment may correct the search score by the following method. The correction by these methods is executed by the information processing unit 10.

(A) Score correction using the distance between words First, the input question sentence is morphologically analyzed to extract words. Then, the extracted words are matched with the words of each sentence constituting the explanation sentence. If all of the extracted words match, the distance between the matched words is measured. Then, using the inter-word distance, "inter-word correction coefficient h"
The search score of the unit information is corrected by calculating "a" and multiplying by this inter-word correction coefficient ha.

For example, when the extracted words are three words A, B, and C, each sentence is divided into words and word A is used.
And find the position where the word B appears. Then, the minimum value of the absolute value between the appearance position of the word A and the appearance position of the word B is obtained. The same processing is performed for word B and word C, and word C and word A. Next, the average value of these minimum values is calculated as the "distance between words". Then, the "word-to-word correction coefficient ha" is calculated using the word-to-word distance. The inter-word correction coefficient ha is calculated by, for example, the following mathematical formula (1).

(Number 1)
Interword correction coefficient ha = (11-distance between words) x 0.01 x a + 1
Then, the search score of the unit information is corrected based on the distance between words. Here, it is assumed that the inter-word distance does not increase by 11 or more. For example, if the distance between words is 12, then 11
The inter-word correction coefficient is calculated by replacing with. Further, the constant a is a numerical value that can be adjusted as appropriate.

For example, when the inter-word distance is 3 and the constant a = 1 for the three extracted words A, B, and C, the inter-word correction coefficient ha is 1.08. Therefore, the inter-word correction coefficient ha = 1.08 is multiplied to correct the search score of the unit information.

(B) Score correction using the number of word occurrences in each sentence First, the input question sentence is morphologically analyzed to extract words. Then, the extracted words are matched with the words of each sentence constituting the explanation sentence. Here, the maximum number of words that match the words included in the explanation in one unit information is set as the "word appearance count", and the number of sentences having the maximum value is set as the "maximum number of sentences". Then, using the number of word appearances and the maximum number of sentences, "
The appearance number correction coefficient hb ”is calculated, and the search score of the unit information is corrected by multiplying the appearance number correction coefficient hb. The occurrence frequency correction coefficient hb is calculated by the following mathematical formula (2). Here, the constants b1 and b2 are numerical values that can be appropriately adjusted.

(Number 2)
Number of occurrences correction coefficient hb = (number of occurrences of words x 0.01 x constant b1) + ((maximum number of sentences-1)
) × 0.01 × constant b2) +1

(C) Score correction using the degree of similarity with each sentence (C1) Pre-processing First, pre-processing will be described. A sentence ID is set for each sentence constituting each explanatory sentence.
Next, each sentence is morphologically analyzed to extract words. Then, a fixed-length vector of each sentence ID and each word is acquired by using a predetermined neural network.

Here, the neural network uses, for example, consecutive words w (t-2) and w (t-1) included in a sentence (uth sentence) with sentence ID = u, followed by a word w (t). It's a guess.

Specifically, the information of d (u), w (t-2), and w (t-1) is input to the input layer. d
(U) is a one-hot vector representation of a sentence with sentence ID = u, and has a (U + 1) dimension. w (t) is represented by the one-hot vector of the t-th word, and (N).
+1) It has a dimension. The total number of sentences is U, the total number of words is N, and the number of vectorized dimensions is S.

Next, the information sent from the input layer to the intermediate layer is converted into S-dimensional vector information using the following mathematical formula (3).
(Number 3)
h = average (d (u) D + w (t-2) W + w (t-1) W)
Here, "average" means an average. Further, the symbol D represents a sentence weight matrix, and has elements in (U + 1) rows and S columns. Further, the symbol W represents a word weight matrix and has elements of (N + 1) rows and S columns.

Subsequently, the information sent from the intermediate layer to the output layer is (N + 1) using the following mathematical formula (4).
It is converted into dimensional vector information.
(Number 4)
h = softmax (hW')
Here, "softmax" means a softmax function. The symbol W'is a transposed matrix of the word weight matrix W.

Then, the information output from the output layer is compared with the inferred word w (t), and the elements of the sentence weight matrix D and the word weight matrix W are updated so as to reduce the difference between the two. The same processing is performed for all sentences and words, and the update of the sentence weight matrix D and the word weight matrix W is repeated. As a result, the u-th line of the sentence weight matrix D corresponds to the fixed-length vector of the sentence with the sentence ID = u. Further, the t-th row of the word weight matrix W corresponds to the fixed-length vector of the t-th word.

(C2) Inference processing Next, the inference processing will be described. When a question sentence is input, the input question sentence is morphologically analyzed to extract words. Next, the fixed-length vector of the interrogative sentence is acquired using the same neural network used in the preprocessing. Here, consecutive words w included in the question sentence
(T-2), w (t-1) performs a task of inferring the following word w (t). Information of d (U + 1), w (t-2), and w (t-1) is input to the input layer. d (U + 1) is a one-hot vector representation of the input sentence (= interrogative sentence), and has a (U + 1) dimension. w (t) is represented by the one-hot vector of the t-th word, and (N).
+1) It has a dimension. If the question text contains an unknown word, the information is stored in w (N + 1). Further, in the intermediate layer, the information sent from the input layer is converted into the information of the S-dimensional vector by using the following mathematical formula (5).

(Number 5)
h = average (d (U + 1) D + w (t-2) W + w (t-1) W)
Subsequently, the information sent from the intermediate layer is input to the output layer, and the following formula (6) is used (
N + 1) Converted to dimensional vector information.

(Number 6)
h = softmax (hW')
Compare the information output from the output layer with the word w (t) and reduce the difference between the two.
Update the elements of the sentence weight matrix D. The same processing is performed for all sentences and words existing in the question sentence. Then, the last row of the sentence weight matrix D is a fixed-length vector corresponding to the input question sentence.

Next, the cosine similarity between the fixed-length vector of the input question sentence and the fixed-length vector corresponding to each sentence ID is calculated. Then, the maximum value of the cosine similarity in the explanation in one unit information is calculated as the maximum similarity. Then, from the maximum similarity, "similarity correction coefficient hc"
Is calculated, and the search score of the unit information is corrected by multiplying this similarity correction coefficient. The similarity correction coefficient hc is calculated by the following mathematical formula (7). The constant c is a numerical value that can be adjusted as appropriate.

(Number 7)
Similarity correction coefficient hc = (maximum similarity x constant c) + 1
For example, when the maximum similarity ms = 0.8 and the constant c = 0.1, the similarity correction coefficient hc = 1.
It becomes 08.

In the above description, the above-mentioned neural network was used to acquire the fixed-length vector of the interrogative sentence, but any sentence model can be adopted as long as the fixed vector of the sentence can be generated.

According to the present inventors, as an example, the number of explanatory sentences of unit information is about 500, the total number of sentences U constituting the explanatory sentences is about 5000, and the number of words N used for them is 100,000 words and 5000 kinds. When the vectorized dimension number S is about 100 to 500,
It was confirmed that the fixed-length vector of the interrogative sentence can have significant information.

<10th Embodiment>
In the tenth embodiment, the dialogue management server 1 according to the ninth embodiment has a plurality of headings, and the search score is calculated after setting the priority order with each heading and the explanation. That is, in the dialogue management server 1 according to the tenth embodiment, the heading included in the unit information (
There are a plurality of layers), the recording unit 11 has a weight for each layer, and the answer generation unit 27 calculates the score using the weight. Other configurations of the tenth embodiment are the same as those of the ninth embodiment. In the dialogue management server 1 according to the tenth embodiment, the search score is calculated by a method such as TF-IDF, but the total value of the scores calculated for each item according to each layer may be used as the search score, or for each item. The total value obtained by multiplying the score calculated in 1 by the weight may be used as the search score. Further, the score having the highest score among the scores for each item may be adopted as the search score. further,
The score calculated for each item multiplied by the weight and having the highest score may be adopted as the search score.

Further, the dialogue management server 1 according to the tenth embodiment may be one in which the weight determination device 150 automatically determines the weight for each item according to each layer. The weight determination device 150 may be incorporated in the same device as the dialogue management server 1, or may be connected via the network N as shown in FIG. 31. Specifically, the weight determining device 150 is shown in FIGS. 32 and 3.
The processing as shown in 3 and 34 is executed. The weight determination device 150 is composed of a computer, and various processes are executed by the information processing unit 155 thereof.

(A) Creation of Weight Candidate List First, the weight determination device 150 selects an arbitrary nth layer from a plurality of layers as the “first layer” (step ST11). Here, it is assumed that there are nine layers. Further, as an example, it is assumed that the first layer (n = 1) is selected as the first layer. Next, the weight of the first layer (here, p (1)) is selected from the initial value list (step ST12).
In the initial value list, candidate values of weights that can be taken for each layer are stored. Specifically, as shown in FIG. 35, numerical values from 100 to 10000 are stored side by side at intervals of 100 in the initial value list. Here, as the weight of the first layer, the first weight p (1) = 1
It is assumed that 00 is selected.

Subsequently, the information processing unit 155 defines the weights p (i ≠ n) of the layers other than the first layer, that is, the weights p (2) to p (9), in the first weight operation defined by the following mathematical formula (8). Calculated from the formula (
Step ST13). In addition, i means the number of the hierarchy. Further, here, the "initial maximum value of the weight" is set to 10000.

Then, the weight p (1) of the first layer selected in step ST11 and step ST1
Search for an explanatory text that matches the sample answer A corresponding to the sample question sentence Q using the first weight set including the weights p (2) to p (9) of the layers other than the first layer calculated in 3. The ranking of the scores is calculated (step ST14). That is, the weight determination device 150 sets the weight sets p (1) to p (9) in the dialogue management server 1 and inputs the sample interrogative sentences Q (steps ST14a to 14c). Then, the answers are extracted in the order of the search score (step ST14).
d, ST14e). It is determined whether or not the output answers match the sample answer A, and if they match, the order is extracted (step ST14f-Yes, 14g). Here, the sample question sentence Q and the sample answer A have a one-to-one correspondence, and a set of a plurality of sample question sentences and the sample answer is sequentially input to the dialogue management server 1. That is, a plurality of ranks are individually extracted for a set of a plurality of sample question sentences and sample answers (step S).
T14h, ST14i).

Subsequently, the information processing unit 155 uses the ranking calculated in step ST14 and the following mathematical formula (9).
And the points for the first weight set are calculated based on the point calculation formula represented by (10) (step ST15).

Note that S is the total number of sample interrogative sentences. Further, g (s) is an arbitrary numerical value (for example, -10) of 0 or less when the rank value calculated in step ST14 is 11 or more. For example, if there are seven sample question sentences and the search order of the answers that match the sample answers corresponding to each sample question sentence is 3,1,1,2,1,13,4, respectively, the point is the second decimal place. Is rounded to 62.9.

Next, the process returns to step ST12, and other weights (other than p (1) = 100) with respect to the first layer.
Is selected, and step ST13, step ST14, and step ST15 are repeatedly executed until all the weights that the first layer can take are selected (step ST16).

Subsequently, the maximum value of the points obtained in step ST15 is extracted, and the threshold value of the points for the first layer selected in step ST11 is calculated based on the maximum value of these points (
Step ST17). Specifically, the value obtained by multiplying the maximum value of points by a predetermined constant (0.9 or the like) is set as the threshold value.

Next, the weight of the first layer when the point calculated in step ST15 is equal to or greater than the threshold value calculated in step ST17 is specified, and the initial value list is updated to the one excluding the other weights (step ST18). ).

Then, the process returns to step ST11, and until all other layers are selected as the first layer.
Steps ST12 to ST18 are repeatedly executed (step ST19). As a result, the initial value list is updated.

Next, an arbitrary m-th layer from a plurality of layers is selected as the "second layer" (step S).
T21). Here, it is assumed that the fifth layer (m = 5) is selected as the second layer. Subsequently, the maximum value of the weight p (5) of the second layer is calculated from the total value of the minimum values of the weights p (1) to p (4) and p (6) to p (9) of the layers other than the second layer. Calculate (step ST22).

Then, in step ST21, another layer (here, a layer other than the fifth layer) is selected as the second layer, and step ST22 is repeatedly executed (step ST23). Then, the same process is executed until the selection of all layers is completed, and a "weight candidate list" as shown in FIG. 36 is generated from the initial value list (step ST24). In other words, step ST22
Then, a weight candidate list is generated by excluding a weight larger than the maximum value of the weight p (m) of the second layer from the initial value list.

(B) Weight Setting First, the information processing unit 155 of the weight determination device 150 selects an arbitrary o-th layer from a plurality of layers as the “third layer” (step ST31). Here, it is assumed that the ninth layer is selected. Next, the information processing unit 155 has weights p (1) to other than the selected hierarchy.
p (8) is randomly extracted from the weight candidate list (step ST32). Next, the information processing unit 155 uses the weights p (1) to p (8) extracted in step ST32 to calculate the weight p (9) of the hierarchy selected in step ST31 from the following weight calculation formula (11). )
Is calculated (step ST33).

The second term on the right side represents the sum of the layers other than the third layer. Further, here, the "initial maximum value of the weight" is set to 10000 (see FIG. 35).

Here, if the calculated weight p (9) of the third layer does not exist in the weight candidate list, the calculation result is discarded and the process returns to step ST32 (step ST34).

Next, the information processing unit 155 uses a weight set including the weights p (9) calculated in step ST33 and the weights p (1) to p (8) extracted in step ST32, and uses the sample question sentence Q. The ranking of the score of the explanatory text corresponding to the sample answer A corresponding to is calculated (step ST35). That is, the weight determination device 150 sets the weight on the dialogue management server 1.
Set (1) to p (9) and enter the sample question text Q. Then, the answers are output in the order of the search score. It is determined whether or not there is an answer that matches the sample answer A, and if it matches, the order is output. There is a one-to-one correspondence between the sample question sentence Q and the sample answer A. A set of a plurality of sample question sentences and sample answers is sequentially input to the dialogue management server 1. That is, a plurality of ranks are individually calculated for a set of a plurality of sample question sentences and sample answers. The process of step ST35 is the above-mentioned step ST14.
(See FIG. 33).

Next, the information processing unit 155 calculates points for the weight set based on the calculated rank and the above-mentioned mathematical expressions (9) and (10) (step ST36). Then, step S
Execution from T32 to step ST36 a predetermined number of times (step ST37). Here, 1
The process is repeated 0,000 times. Then, the weight of each layer is determined based on the points after the execution 100,000 times (ST38). Specifically, among the points calculated 100,000 times, the weight set when the value of the points has the maximum value is determined as the weight of each layer.

<11th Embodiment>
In the eleventh embodiment, the dialogue management server 1 according to the ninth embodiment determines the validity of the synthetic word registered in the first synthetic word dictionary DB12A or the second synthetic word dictionary DB12B. Hereinafter, for convenience, a method for determining the validity of the compound word registered in the first compound word dictionary DB 12A will be described, but the validity of the compound word registered in the second compound word dictionary 12B will also be determined by the same processing. it can.

Specifically, in the dialogue management server 1 according to the ninth embodiment, the validity of the first synthetic word dictionary DB 12A is determined by the validity determination device 200. The validity determination device 200 may be incorporated in the same device as the dialogue management server 1, or may be connected via the network N as shown in FIG. 37.

The validation device 200 executes the process as shown in FIG. 38. Validation device 200
Is composed of a computer, and various processes are executed by the information processing unit 205.

First, the information processing unit 205 of the validity determination device 200 generates a "duplicate dictionary" that duplicates the contents of the first synthetic word dictionary DB 12A (step ST51). Next, the information processing unit 205
Initialize the first synthetic word dictionary DB12A of the dialogue management server 1 (step ST52). When the processing shown below is interrupted, the first compound word dictionary DB1 can be overwritten with the replication dictionary.
Return 2A to its original state.

Using the first synthetic word dictionary DB12A initialized after step ST52, the ranking of the scores of the explanatory texts matching the sample answer A corresponding to the sample question text Q is calculated (step ST53). Specifically, the sample question sentence Q is input. Then, the answers are output in the order of the search score. It is determined whether or not there is an answer that matches the sample answer A, and if it matches, the order is output. There is a one-to-one correspondence between the sample question sentence Q and the sample answer A. In addition, a set of a plurality of sample question sentences and sample answers is the dialogue management server 1
Is input sequentially to. That is, a plurality of ranks are individually calculated for a set of a plurality of sample question sentences and sample answers. The processing of such step ST53 is the same as that of step ST14 described above (see FIG. 33).

Then, the information processing unit 205 uses the rank calculated in step ST53 and the lower mathematical formula (12).
) And (13), the points are calculated based on the predetermined point calculation formula (step ST54).

Next, the information processing unit 205 extracts an arbitrary synthetic word from the synthetic word registered in the duplicate dictionary and registers it in the first synthetic word dictionary 12A after the initialization (step ST55).

Subsequently, the information processing unit 205 describes the first synthetic word dictionary DB12A after registration in step ST55.
Is used to calculate the ranking of the scores of the explanatory texts that match the sample answers corresponding to the sample question texts (step ST56).

Subsequently, the information processing unit 205 calculates the points for the composite words extracted in step ST55 based on the rank calculated in step ST56 and the predetermined point calculation formula (step ST57).

Then, the information processing unit 205 determines the validity of the composite word from the comparison between the points calculated in step ST54 and the points calculated in step ST57 (step ST58). Here, if the points are increased, it is determined that the validity of the composite word registered in step ST55 is high. On the other hand, if the point is descending, step ST55
It is judged that the validity of the compound word registered in is low. Words that are determined to be of low validity are recorded as candidates for deletion.

After that, the compound words registered in the duplicate dictionary are registered word by word, and the same process is repeated until all the compound words are registered (step ST59). However, in step 58, the information processing unit 205 replaces the comparison between the points calculated in step ST54 and the points calculated in step ST57 with the points calculated in step ST57 and the previous time (a new composite word is registered). The validity of the composite word is determined from the comparison with the points calculated in step ST57 (before the information processing is performed).

As described above, the validity determination device 200 according to the present embodiment registers the synthetic words registered in the replication dictionary in the first synthetic dictionary after the initialization step, and compares the points before and after the registration. To do. Thereby, it can be determined whether or not the synthetic word registered in the first synthetic word dictionary DB12A is appropriate. The validity of the composite word registered in the second composite word dictionary DB12B can be determined by executing the same processing.

<12th Embodiment>
FIG. 39 is a schematic diagram showing the configuration of the dialogue management system according to the twelfth embodiment of the present invention.
The dialogue management server 1 according to the twelfth embodiment further includes a word distributed expression database (DB) 201.

The word distributed expression DB 201 is a database that stores a distributed expression of any word. Here, the "distributed expression of words" is a fixed-length vector in which the characteristics of each word are stored, and there is a one-to-one correspondence between numerical expressions in vector form for one word.

In addition, the answer generation unit 27 according to the present embodiment operates as follows. First, the answer generation unit 27 selects a word from the words extracted by morphological analysis of the text data of the question sentence according to a predetermined rule. Here, when a plurality of words are extracted by morphological analysis, the longest word is selected. Subsequently, the answer generation unit 27 extracts words within a certain distance in the distributed expression as "related words" from the word distributed expression DB 201 with respect to the selected word.

Next, the answer generation unit 27 calculates the score based on the frequency of appearance of the words and related words extracted by analyzing the text data of the interrogative sentence in each unit information. And answer generation unit 2
7 generates answer output information using the calculated score.

For example, when the question sentence includes an expression such as "clearing the ETC card", the answer generation unit 27 morphologically analyzes the expression into two words "ETC card" and "clearing". Here, it is assumed that the "ETC card" is registered in the first synthetic word dictionary DB12A or the second synthetic word dictionary DB12B. Next, the answer generation unit 27 has 2 of "ETC card" and "clearing".
Among the words, select "ETC card" which is a word with a long number of characters. Subsequently, the answer generation unit 27 extracts words within a certain distance from the "ETC card" as related words in the word distribution expression. Here, for example, words such as "car" and "credit card" are extracted as related words. Then, the answer generation unit 27 calculates the search score including these related words, and generates the answer output information from the unit information in descending order of the search score.

If the related words are not extracted, the answer output information is generated based on the morphologically analyzed words.

As described above, in the dialogue management server 1 according to the present embodiment, the document information B including related words is included.
Since I is searched, the probability of presenting necessary information to the user can be increased.

It should be noted that each of the above embodiments is for facilitating the understanding of the present invention, and does not limit the interpretation of the present invention. The present invention can be modified / improved without departing from the spirit thereof, and the present invention also includes an equivalent thereof. In addition, the present invention can form various disclosures by appropriately combining the plurality of components disclosed in each of the above embodiments.
For example, some components may be removed from all the components shown in the embodiments. Further, the components may be appropriately combined in different embodiments.

Further, each step of the flowchart does not necessarily have to be executed according to the above-mentioned steps. For example, steps S1, S3, and S5 shown in FIG. 4 may be executed after step S7.

1 ... Dialogue management server, 4 ... Dialogue log DB, 6 ... User terminal device, 8 ... Operator terminal device, 10 ... Information processing unit, 11 ... Recording unit, 12 ... Technical term dictionary DB, 12A ... First compound word dictionary DB , 12B ... 2nd compound word dictionary DB, 13 ... synonymous dictionary DB, 14 ... excluded term dictionary DB,
15 ... Acquisition unit, 17 ... Data conversion unit, 19 ... Reception unit, 21 ... Natural language processing unit, 23 ... Weighting setting unit, 24 ..., Reverse document frequency DB, 27 ... Answer generation unit, 29 ... Dialogue result management unit, 31
... Output control unit, 33 ... Voice recognition processing unit, 35 ... Voice dialogue management unit, 44 ... Subordinate concept word dictionary D
B, 46 ... Short talk DB, 100 ... Dialogue management system, 150 ... Weight determination device, 1
55 ... Information processing unit, 200 ... Validity judgment device, 201 ... Word distributed expression DB, 205 ... Information processing unit

Claims (24)

Data in a structured format document having a plurality of unit information in which the text data of the explanatory text, the information for identifying one or more layers leading to the explanatory text, and the text data representing the heading of the hierarchical text are associated with each other. A recording unit that records
The reception department that accepts text data of question sentences from users,
The text data of the question text received by the reception unit was matched with the text data included in each unit information recorded in the recording unit, and the unit information related to the question text was extracted and extracted. Answer generation unit that generates answer output information based on headings and explanations corresponding to unit information,
Dialog management server with. The answer generation unit
The score of each unit information is calculated based on the frequency of appearance in each unit information of the words extracted by morphological analysis of the text data of the interrogative sentence.
Generate answer output information using the calculated score,
The dialogue management server according to claim 1. The answer generation unit
Each unit information is based on the frequency of appearance of the words extracted by morphological analysis of the text data of the interrogative sentence in each unit information and the ratio of the unit information including the extracted words to the total number of unit information. Calculate the score of
Generate answer output information using the calculated score,
The dialogue management server according to claim 1 or 2. Further provided with a first synthetic word dictionary storage unit for storing a first synthetic word dictionary in which a synthetic word obtained by synthesizing another word with a word of a predetermined noun is registered.
The answer generation unit
When it is determined that the extracted word is registered in the first synthetic word dictionary by morphological analysis of the text data of the interrogative sentence, the first word is based on the frequency of appearance of the extracted word in each unit information. In addition to calculating the score, the second score is calculated based on the frequency of appearance of the compound word in each unit information.
The answer output information is generated using the first score and the second score.
The dialogue management server according to any one of claims 1 to 3. A second compound word dictionary storage unit for storing a second compound word dictionary registered by associating a word with a predetermined noun with a compound word obtained by synthesizing another word with the word is further provided.
The answer generation unit
When it is determined that the extracted word by morphological analysis of the text data of the question sentence is registered in the second synthetic word dictionary, the extracted word, another word synthesized into the word of the noun, and the other word synthesized into the noun word, and A score is calculated based on the frequency of appearance in each unit information for each of the compound words, and answer output information is generated using the score.
The dialogue management server according to any one of claims 1 to 3. Further, a subordinate concept word dictionary storage unit for storing a subordinate concept word dictionary registered by associating a predetermined word with a subordinate concept word which is a subordinate concept of the word is further provided.
The answer generation unit
When it is determined that the extracted word is registered in the subordinate concept word dictionary by morphological analysis of the text data of the interrogative sentence, answer output information including the subconcept word associated with the word is generated.
The dialogue management server according to any one of claims 1 to 5. The answer generation unit generates answer output information based on the text data included in each heading of unit information including words extracted by morphological analysis of the text data of the interrogative sentence.
The dialogue management server according to claims 1 to 6. The answer generation unit calculates the score of the unit information related to the question sentence based on the morphological analysis of the text data of the question sentence, and sets the score within a predetermined range to the number of the calculated unit information. An output control unit for adjusting the output time of the answer output information is further provided accordingly.
The dialogue management server according to any one of claims 1 to 7. When the unit information related to the interrogative sentence cannot be extracted, the answer generation unit generates answer output information corresponding to a predetermined answer sentence based on the extracted word by morphological analysis of the interrogative sentence.
The dialogue management server according to any one of claims 1 to 8. The answer generation unit
Based on the morphological analysis of the text data of the interrogative sentence, the score of the unit information related to the interrogative sentence is calculated.
Corresponds to the answer output information according to the difference between the first average score of one or more unit information having a higher score and the second average score of one or more unit information lower than the first average score. Change the display format of the user image showing the user
The dialogue management server according to any one of claims 1 to 9. It also has a compound word dictionary storage unit that stores a compound word dictionary in which a compound word obtained by synthesizing another word with a word of a predetermined noun is registered.
The answer generation unit responds to the number of synthetic words included in the text data of the interrogative sentence.
The display form of the user image indicating the user associated with the answer output information is changed.
The dialogue management server according to any one of claims 1 to 10. The answer generation unit
The frequency of appearance of words extracted by morphological analysis of the text data of the question sentence in each unit information, the ratio of the unit information including the extracted words to the total number of unit information, and the length of the unit information. Calculate the score of each unit information based on
Generate answer output information using the calculated score,
The dialogue management server according to any one of claims 1 to 11. The description contains one or more sentences.
The answer generation unit
Based on the morphological analysis of the text data of the interrogative sentence, the score of the unit information related to the interrogative sentence is calculated.
When the words extracted by the morphological analysis and the words constituting the sentence included in the explanatory text are compared and the first word and the second word different from each other are identified, the appearance position of the first word and the above The score of the unit information is corrected based on the distance between words with the appearance position of the second word.
The answer output information is generated using the score.
The dialogue management server according to any one of claims 1 to 12. The answer generation unit
Based on the morphological analysis of the text data of the interrogative sentence, the score of the unit information related to the interrogative sentence is calculated.
The words extracted by the morphological analysis are compared with the words constituting the sentence included in the explanatory sentence, and the score of the unit information is corrected based on the number of matching word types in one sentence.
The answer output information is generated using the score.
The dialogue management server according to any one of claims 1 to 13. The answer generation unit
The score is corrected based on the number of sentences in which the number of matching word types is equal to or greater than a predetermined value.
The dialogue management server according to claim 14. The answer generation unit
Based on the morphological analysis of the text data of the interrogative sentence, the score of the unit information related to the interrogative sentence is calculated.
A fixed-length vector is calculated from the interrogative using a sentence model that converts the sentences included in the explanatory text into individual fixed-length vectors.
The score is corrected based on the degree of similarity between the fixed-length vector of the interrogative sentence and the fixed-length vector of the sentence included in the explanatory sentence.
The answer output information is generated using the score.
The dialogue management server according to any one of claims 1 to 15. There are multiple layers,
The storage unit has a weight for each layer.
The answer generation unit
Based on the morphological analysis of the text data of the interrogative sentence and the weight, the score of the unit information related to the interrogative sentence is calculated.
The answer output information is generated using the score.
The dialogue management server according to any one of claims 1 to 16. The weight determining device for determining the weight in the dialogue management server according to claim 17.
A selection step for selecting an arbitrary layer from the plurality of layers, and
An extraction step that randomly extracts weights other than the hierarchy selected by the selection step from a predetermined weight candidate list, and
A weight calculation step for calculating the weight of the hierarchy selected by the selection step from a predetermined weight calculation formula using the weight extracted by the extraction step, and
Score rank calculation to calculate the rank of the score of the explanatory text that matches the sample answer corresponding to the sample question text using the weight set including the weight calculated by the weight calculation step and the weight extracted by the extraction step. Steps and
A point calculation step for calculating points for the weight set based on the rank calculated by the score rank calculation step and a predetermined point calculation formula, and
A weight that executes the extraction step, the weight calculation step, the score ranking calculation step, and the determination step of determining the weight of each layer based on the points after executing the point calculation step a predetermined number of times. Decision device. A first layer selection step that selects an arbitrary layer as the first layer from a plurality of the above layers, and
The first weight selection step of selecting the weight of the first layer from the initial value list, and
A first weight calculation step of calculating the weights of layers other than the first layer from a predetermined first weight calculation formula, and
Using the first weight set including the weight of the first layer selected by the first weight selection step and the weight of the layer other than the first layer calculated by the first weight calculation step, the sample question text The first score ranking calculation step, which calculates the ranking of the scores of the explanations that match the corresponding sample answers,
A first point calculation step for calculating points for the first weight set based on the rank calculated by the first score rank calculation step and a predetermined point calculation formula, and
When the points calculated by the first point calculation step are less than a predetermined value, the deletion step of deleting the weight of the first layer from the initial value list and the deletion step.
In the first weight selection step, another weight is selected, and the first weight calculation step, the first score ranking calculation step, the first point calculation step, and the deletion step are repeatedly executed to perform the initial value list. The first update step to update and
In the first layer selection step, another layer is selected, and the first weight selection step, the first weight calculation step, the first score ranking calculation step, the first point calculation step, the deletion step, and the above. The second update step of repeatedly executing the first update step to update the initial value list, and
A second layer selection step that selects an arbitrary layer as the second layer from the plurality of layers,
A second weight calculation step for calculating the weight of the second layer from the total value of the minimum values of the weights of the layers other than the second layer, and
In the second layer selection step, another layer is selected, the second weight calculation step is repeatedly executed, and a weight candidate list creation step for creating a weight candidate list from the initial value list, and a weight candidate list creation step.
18. The weighting device according to claim 18. A validity determination device for determining the validity of a compound word registered in the first compound word dictionary according to claim 4.
A duplicate dictionary generation step of generating a duplicate dictionary that duplicates the contents of the first synthetic word dictionary, and
The initialization step for initializing the first compound word dictionary and
A registration step of extracting an arbitrary compound word from the compound words registered in the duplicate dictionary and registering it in the first compound word dictionary after the initialization step,
Using the first synthetic word dictionary after the registration step, a score ranking calculation step for calculating the score ranking of the explanatory text matching the sample answer corresponding to the sample question text, and a score ranking calculation step.
A point calculation step for calculating points for a composite word extracted by the registration step based on the ranking calculated by the score ranking calculation step and a predetermined point calculation formula, and a point calculation step.
By repeating the registration step, the score ranking calculation step, and the point calculation step, and comparing the points before and after the compound word is registered in the first compound word dictionary after the initialization step, the registered compound word Validation step to determine validity and
A validation device that executes. A validity determination device for determining the validity of a compound word registered in the second compound word dictionary according to claim 5.
A duplicate dictionary generation step of generating a duplicate dictionary that duplicates the contents of the second synthetic word dictionary, and
The initialization step for initializing the second compound word dictionary and
A registration step of extracting an arbitrary compound word from the compound words registered in the duplicate dictionary and registering it in the second compound word dictionary after the initialization step,
Using the second compound word dictionary after the registration step, a score ranking calculation step for calculating the score ranking of the explanatory text matching the sample answer corresponding to the sample question text, and a score ranking calculation step.
A point calculation step for calculating points for a composite word extracted by the registration step based on the ranking calculated by the score ranking calculation step and a predetermined point calculation formula, and a point calculation step.
By repeating the registration step, the score ranking calculation step, and the point calculation step, and comparing the points before and after the compound word is registered in the second compound word dictionary after the initialization step, the registered compound word Validation step to determine validity and
A validation device that executes. It also has a word distributed expression storage unit that stores distributed expressions of arbitrary words.
The answer generation unit
A word is selected from the words extracted by morphological analysis of the text data of the interrogative sentence according to a predetermined rule.
With respect to the selected word, words within a certain distance in the distributed expression are extracted as related words.
A score is calculated based on the frequency of appearance of the words extracted by analyzing the text data of the interrogative sentence and the related words in each unit information.
The answer output information is generated using the score.
The dialogue management server according to any one of claims 1 to 17. Data in a structured format document having a plurality of unit information in which the text data of the explanatory text, the information for identifying one or more layers leading to the explanatory text, and the text data representing the heading of the hierarchical text are associated with each other. And the steps to record
Steps to accept text data of question sentences from users,
The text data of the received question text is matched with the text data included in each recorded unit information, the unit information related to the question text is extracted, and the heading and explanation corresponding to the extracted unit information are extracted. Steps to generate answer output information based on sentences,
When,
Dialogue management methods, including. Computer,
Data in a structured format document having a plurality of unit information in which the text data of the explanatory text, the information for identifying one or more layers leading to the explanatory text, and the text data representing the heading of the hierarchical text are associated with each other. Recording unit, which records
Reception department that accepts text data of question texts from users,
The text data of the question text received by the reception unit was matched with the text data included in each unit information recorded in the recording unit, and the unit information related to the question text was extracted and extracted. Answer generation unit that generates answer output information based on the heading and description corresponding to the unit information,
A program that functions as.