JP2022187109A - Dialog knowledge generation apparatus and dialog knowledge generation method - Google Patents

Abstract

To automatically construct a flow model of a dialog including an asking using a business document and a past dialog record.SOLUTION: A dialog knowledge generation apparatus includes: a keyword selection unit that selects an attribute value which is a keyword of a dialog and an attribute name indicating a category of the attribute value from target data including a business document and a past dialog record using a clue phrase indicating each selection condition; a mapping unit that determines a correspondence relationship between the attribute value and the attribute name based on each of acquisition positions in the target data; and a graph generation unit that generates a knowledge graph by extracting a hierarchical relationship of a plurality of items including the attribute value or the attribute name, and combining the plurality of items based on the extracted hierarchical relationship.SELECTED DRAWING: Figure 2

Description

The present invention relates to a dialogue knowledge generation device and a dialogue knowledge generation method, and is suitable for application to a dialogue knowledge generation device and a dialogue knowledge generation method for generating dialogue knowledge by text processing using a computer.

For example, Japanese Unexamined Patent Application Publication Nos. 2003-100003 and 2003-100001 are examples of conventional techniques relating to the generation of a flowchart-type data structure using a computer. Patent Literature 1 discloses a system that generates a set of vectors from a corpus and generates a data structure (eg, cluster formation tree, phylogenetic tree). Further, Patent Document 2 discloses a dialogue management server that receives text data of a question sentence from a user and calculates a vector as a dialogue management server that manages dialogue knowledge.

JP 2019-169148 A JP 2020-184294 A

For example, in response to an inquiry such as "I received a notification that my contract is about to expire, but I don't know how to deal with it, please let me know." It is preferable to present the final answer while repeatedly asking questions to confirm unclear points. In this case, if it is possible to prepare a flowchart-type data structure (flow model) representing dialogue knowledge in advance and manage the dialogue while referring to it, even a non-expert person can understand the procedure leading up to the answer. can do. However, when using the conventional technology described above, although the tree structure of the dialogue flow model can be generated, there is the problem that it is not possible to allocate the data of the feedback questions and answers in the dialogue.

The present invention has been made in consideration of the above points, and utilizes business documents (e.g., business manuals) and past dialogue records (e.g., dialogue logs) to create a dialogue flow model including reflection (dialogue with reflection text). The object of the present invention is to propose a dialog knowledge generation method and a dialog knowledge generation apparatus capable of automatically constructing knowledge.

In order to solve this problem, the present invention provides a dialogue knowledge generation device that automatically generates a knowledge graph representing a flow model of dialogue, wherein from target data including business documents and past dialogue records, attributes that are keywords of dialogue A keyword selection unit (for example, an attribute value keyword selection unit 142 and an attribute name keyword selection unit 143, which will be described later) that selects a value and an attribute name indicating a category of the attribute value using clue phrases indicating respective selection conditions; An associating unit (for example, an attribute value-attribute name associating unit 146 to be described later) that determines the correspondence relationship between the attribute value and the attribute name selected by the keyword selecting unit based on the respective acquisition positions in the target data. ), extracting a hierarchical relationship of a plurality of items having the attribute value or the attribute name based on the correspondence determined by the associating unit, and combining the plurality of items based on the extracted hierarchical relationship. provides a dialogue knowledge generation device including a graph generation unit (for example, a knowledge graph generation unit 147 and a feedback question generation unit 148, which will be described later) that generates the knowledge graph.

Further, in order to solve such problems, the present invention provides a dialogue knowledge generation method by a dialogue knowledge generation device for automatically generating a knowledge graph representing a flow model of dialogue, wherein the dialogue knowledge generation device is composed of business documents and past a keyword selection step of selecting, from target data including dialogue records, attribute values that are keywords of dialogue and attribute names that indicate categories of attribute values, using clue phrases that indicate selection conditions for the respective attribute values; and the dialogue knowledge generation device. a matching step of determining a correspondence relationship between the attribute values selected in the keyword selecting step and the attribute names based on their acquisition positions in the target data; extracting the hierarchical relationship of a plurality of items having the attribute value or the attribute name based on the correspondence determined in the step, and combining the plurality of items based on the extracted hierarchical relationship to create the knowledge graph; and a generating graph generation step.

According to the present invention, it is possible to automatically construct a dialogue flow model including reflection by utilizing business documents and past dialogue records.

1 is a block diagram showing a hardware configuration example of a computer 10 that implements a dialogue knowledge generating apparatus 100 according to an embodiment of the present invention; FIG. 2 is a block diagram showing a functional configuration example of the dialogue knowledge generating device 100; FIG. FIG. 4 is a diagram showing a data configuration example of clue phrase information; It is a figure which shows the data structural example of attribute value keyword information. FIG. 4 is a diagram showing a data configuration example of attribute name keyword information; It is a figure which shows the data structural example of final reply keyword information. FIG. 9 is a diagram showing a data configuration example of attribute value-final answer correspondence relationship information; FIG. 10 is a diagram showing a data configuration example of attribute value-attribute name correspondence information; FIG. 4 is a diagram showing a data configuration example of knowledge graph node information; FIG. 4 is a diagram showing a data configuration example of knowledge graph edge information; FIG. 11 is a flowchart showing an example of an overall processing procedure of dialogue knowledge generation processing; FIG. FIG. 11 is a flow chart showing an example of a clue phrase estimation process; FIG. FIG. 11 is a flowchart showing an example of a processing procedure for attribute value keyword selection processing; FIG. FIG. 10 is a flowchart showing an example of a processing procedure for attribute name keyword selection processing; FIG. FIG. 11 is a flowchart showing an example of a processing procedure for final reply keyword selection processing; FIG. FIG. 10 is a flow chart showing an example of a processing procedure of attribute value-final answer association processing; FIG. FIG. 10 is a flow chart showing an example of a processing procedure for attribute value-attribute name association processing; FIG. FIG. 11 is a flowchart illustrating an example of a processing procedure of knowledge graph generation processing; FIG. FIG. 10 is a diagram for visually explaining an image of knowledge graph generation processing; FIG. 11 is a flow chart showing an example of a processing procedure of a feedback question generation process; FIG. FIG. 10 is a diagram for visually explaining an image of the feedback question generation process; FIG. 3 is a diagram showing an example (part 1) of visualization of a knowledge graph; FIG. 3 is a diagram showing an example (part 1) of visualization of a knowledge graph;

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

(1) Configuration FIG. 1 is a block diagram showing an example hardware configuration of a computer 10 that implements a dialogue knowledge generation apparatus 100 according to an embodiment of the present invention.

The computer 10 is configured by connecting a CPU 11 , a memory 12 , a storage device 13 , an input device 14 , an output device 15 and a network device 16 to each other via a bus 17 .

The CPU 11 is a processor that executes programs, such as a CPU (Central Processing Unit). The memory 12 is a main storage device such as a RAM (Random Access Memory) that stores programs and data referred to when the programs are executed. The storage device 13 is an auxiliary storage device that stores programs and data to be referenced when the programs are executed, and is, for example, a HDD (Hard Disk Drive) or an SSD (Solid State Drive). The input device 14 is a device that receives input from an operator, such as a keyboard, mouse, or touch panel. The output device 15 is a device for outputting the result of processing executed by the CPU 11 or the data held in the memory 12 or the storage device 13. For example, the output device 15 is a liquid crystal display device, an organic EL (Electro Luminescence) display, or the like. is. The network device 16 is various interfaces such as a communication interface and an input/output interface. Bus 17 is an internal communication line that connects each device within computer 10 .

Each function of the dialogue knowledge generating apparatus 100 shown in FIG. 2, which will be described later, is realized by executing a processing program corresponding to each function by the hardware resources of the computer 10 described above. Specifically, each functional unit in control unit 140 shown in FIG. 2 is realized by CPU 11 executing a predetermined program held in memory 12 or storage device 13 while referring to related data. Also, the knowledge graph visualization unit 150 is realized by the CPU 11 executing a predetermined program and further outputting the execution result to the output device 15 (or the user terminal 200 may be used). The text information DB 110, relationship information DB 120, and graph information DB 130 shown in FIG. 2 are mainly implemented by the storage device 13. FIG.

Note that part or all of the above programs executed by the CPU 11 in the computer 10 and data referred to by the programs may be stored in advance in the memory 12 or the storage device 13, or may be stored in the network device 16 as necessary. may be configured to be stored in the storage device 13 of the computer 10 from a non-temporary storage device or a non-temporary storage medium of another device connected via a network. Further, in the following description, for the sake of simplicity, the dialogue knowledge generation device 100 is assumed to be implemented by one computer 10, but the structure of the dialogue knowledge generation device 100 according to this embodiment is not limited to this. Instead, it may be implemented by a plurality of computers 10, or may be implemented by including a database device, cloud storage, and the like.

FIG. 2 is a block diagram showing a functional configuration example of the dialogue knowledge generation device 100. As shown in FIG. Although the details will be described later, the dialogue knowledge generating device 100 accepts input of business documents, past dialogue records (hereinafter also referred to as inquiry records), and initial clue phrases from the user terminal 200, and uses attribute value keywords, attribute name keywords, and final answer keywords are generated, their corresponding relationships are specified, a knowledge graph and a reflection question are generated, and the knowledge graph (knowledge flow model) with the text of the reflection question is visualized. In this example, a business manual (also referred to as a manual) is used as an example of a business document based on text data, and a dialog log is used as an example of a past dialog record (inquiry record) based on text data.

As shown in FIG. 2, the dialogue knowledge generating apparatus 100 is configured with a text information DB 110, a relationship information DB 120, a graph information DB 130, a control section 140, and a knowledge graph visualization section 150. FIG.

The text information DB 110 stores data related to text among the data necessary for generating knowledge graphs and generating feedback questions. As shown in FIG. 2, the text information DB 110 comprises a clue phrase DB 111, an attribute value DB 112, an attribute name DB 113, and a final answer DB 114. FIG.

The clue phrase DB 111 stores clue phrase information related to "clue phrases." "Clue phrase" selects "attribute value" which is a keyword of dialogue, "attribute name" which indicates the category of attribute value, and "final answer" of dialogue from target data (business documents and past dialogue records). It is a phrase that serves as a selection condition for In the present embodiment, a clue phrase initially provided from the outside (user terminal 200) is referred to as an initial clue phrase. Further, although the details will be described later, in the present embodiment, the clue phrase estimation unit 141 generates an initial clue phrase (which may be read as a target expression) based on a target expression extracted from target data using the initial clue phrase. The clue phrase to be complemented can be estimated, and the clue phrase information stored in the clue phrase DB 111 includes these initial clue phrases and the complemented clue phrases. A data configuration example of clue phrase information will be described later with reference to FIG.

The attribute value DB 112 stores attribute value keyword information related to "attribute value" which is a type of keyword extracted from business documents (eg, business manuals) or inquiry records (eg, dialogue logs). The attribute value is information for obtaining "answer text" in the knowledge graph, which will be described later. The attribute value keyword information is extracted by the attribute value keyword selection unit 142, and its data configuration example will be described later with reference to FIG.

The attribute name DB 113 stores attribute name keyword information related to "attribute name" representing the category of the attribute value. The relationship between attribute names and attribute values corresponds to the relationship between keys and values. The attribute name is information for obtaining a "review text" in the knowledge graph, which will be described later. The attribute name keyword information is extracted by the attribute name keyword selection unit 143, and its data configuration example will be described later with reference to FIG.

The final reply DB 114 stores final reply keyword information related to "final reply", which is a type of keyword extracted from business documents or inquiry records. The final answer corresponds to the final answer in the dialogue knowledge including the reflection question, in other words, it is the information for obtaining the "final answer text" in the knowledge graph to be described later. The final answer keyword information is extracted by the final answer keyword selection unit 144, and its data configuration example will be described later with reference to FIG.

The relationship information DB 120 stores data associated with information between texts, among the data necessary for generating knowledge graphs and generating feedback questions. As shown in FIG. 2, the relationship information DB 120 includes an attribute value final answer correspondence DB 121 and an attribute value attribute name correspondence DB 122 .

The attribute value final answer correspondence DB 121 stores attribute value-final answer correspondence information indicating the correspondence between attribute value keywords and final answer keywords. The attribute value-final answer correspondence relation information is generated by the attribute value-final answer association unit 145, and its data configuration example will be described later with reference to FIG.

The attribute value-attribute name correspondence DB 122 stores attribute value-attribute name correspondence information indicating the correspondence between attribute value keywords and attribute name keywords. The attribute value-attribute name correspondence information is generated by the attribute value-attribute name association unit 146, and its data configuration example will be described later with reference to FIG.

The graph information DB 130 stores data associated with graph structure information, among the data necessary for generating a knowledge graph indicating dialogue knowledge and for generating feedback questions added as text data to the knowledge graph. As shown in FIG. 2, the graph information DB 130 is composed of a knowledge graph node DB 131 and a knowledge graph edge DB 132. FIG.

The knowledge graph node DB 131 stores knowledge graph node information relating to nodes that are constituent elements of a knowledge graph representing dialogue knowledge. The knowledge graph node information is supplied by the knowledge graph generation unit 147 and the feedback question generation unit 148, and an example of the data structure will be described later with reference to FIG.

The knowledge graph edge DB 132 stores knowledge graph edge information relating to edges that are constituent elements of a knowledge graph representing dialogue knowledge. The knowledge graph edge information is supplied by the knowledge graph generation unit 147 and the feedback question generation unit 148, and an example of the data configuration will be described later with reference to FIG.

In this embodiment, the knowledge graph is a form of a "dialogue flow model", which is a flowchart-type data structure in which text data of reflections and answers are assigned, and is a node ( A knowledge graph node) and an edge (knowledge graph edge) holding the answer text of the question-receiver are represented by a structure in which they are hierarchically connected.

The control unit 140 includes a clue phrase estimation unit 141, an attribute value keyword selection unit 142, an attribute name keyword selection unit 143, a final answer keyword selection unit 144, an attribute value-final answer association unit 145, and an attribute value-attribute name association unit. 146, a knowledge graph generation unit 147, and a feedback question generation unit 148, and provides functions possessed by each functional unit.

The clue phrase estimation unit 141 receives a business document (in this example, a business manual), an inquiry record (in this example, a dialogue log), and an input of an initial clue phrase from the user terminal 200, and uses the initial clue phrase to generate target data ( Extract new clue phrases from business documents and inquiry records). The clue phrase estimation unit 141 creates clue phrase information about the extracted clue phrase and stores it in the clue phrase DB 111 . Regarding the above-described processing (referred to as clue phrase estimation processing) by the clue phrase estimation unit 141, a detailed processing procedure example will be described later with reference to FIG.

The attribute value keyword selection unit 142 uses the clue phrase information stored in the clue phrase DB 111 to extract the attribute value keyword and its positional information in the target data, and converts the attribute value keyword information including these extraction results into attribute value keyword information. Store in the value DB 112 . With regard to the above processing (referred to as attribute value keyword selection processing) by the attribute value keyword selection unit 142, a detailed processing procedure example will be described later with reference to FIG.

The attribute name keyword selection unit 143 uses the clue phrase information stored in the clue phrase DB 111 to extract the attribute name keyword and the position information in the target data, and the attribute name keyword information including these extraction results. Store in name DB 113 . With regard to the above processing (referred to as attribute name keyword selection processing) by the attribute name keyword selection unit 143, a detailed processing procedure example will be described later with reference to FIG.

The final answer keyword selection unit 144 uses the clue phrase information stored in the clue phrase DB 111 to extract the final answer keyword and its position information in the target data, etc., and extracts the final answer keyword information including these extraction results as the final answer. Store in DB 114 . With regard to the above processing (referred to as final response keyword selection processing) by the final response keyword selection unit 144, a detailed processing procedure example will be described later with reference to FIG.

The attribute value-final answer association unit 145 uses the attribute value keyword information stored in the attribute value DB 112 and the final answer keyword information stored in the final answer DB 114 to determine the correspondence relationship between the attribute value record and the final answer record. is specified, and the attribute value-final answer correspondence information including the specified result is stored in the attribute value-final answer correspondence DB 121. FIG. With regard to the above processing (referred to as attribute value-final reply correlating processing) by the attribute value-final reply correlating unit 145, a detailed processing procedure example will be described later with reference to FIG.

The attribute value-attribute name association unit 146 uses the attribute value keyword information stored in the attribute value DB 112 and the attribute name keyword information stored in the attribute name DB 113 to determine the correspondence relationship between the attribute value record and the attribute name record. is specified, and attribute value-attribute name correspondence information including the specified result is stored in the attribute value-attribute name correspondence DB 122 . Regarding the above processing (referred to as attribute value-attribute name association processing) by the attribute value-attribute name association unit 146, a detailed processing procedure example will be described later with reference to FIG.

The knowledge graph generation unit 147 uses the attribute value-final answer correspondence information stored in the attribute value-final answer correspondence DB 121 to generate knowledge graph nodes as information about nodes and edges that constitute a knowledge graph representing dialogue knowledge. Information and knowledge graph edge information are generated and stored in the knowledge graph node DB 131 and knowledge graph edge DB 132 . With regard to the above-described processing (referred to as knowledge graph generation processing) by the knowledge graph generation unit 147, a detailed processing procedure example will be described later with reference to FIGS. 18 and 19. FIG.

The feedback question generation unit 148 uses the attribute value-attribute name correspondence information stored in the attribute value-attribute name correspondence DB 122 to store the knowledge graph node information stored in the knowledge graph node DB 131 and the knowledge graph edge DB 132. By appropriately registering texts in the obtained knowledge graph edge information, the knowledge graph indicating the dialogue knowledge is associated with the text of the reflection question (reflection text) and the text of the answer to the question (answer text). Regarding the above processing (referred to as feedback question generation processing) by the feedback question generation unit 148, a detailed processing procedure example will be described later with reference to FIGS. 20 and 21. FIG.

The knowledge graph visualization unit 150 uses the knowledge graph node information stored in the knowledge graph node DB 131 and the knowledge graph edge information stored in the knowledge graph edge DB 132 to visualize the knowledge graph, thereby generating a reflective question with text. To output dialogue knowledge in a form that can be visually recognized by a user. A specific example of visualization of the knowledge graph by the knowledge graph visualization unit 150 is illustrated in FIGS. 22 and 23. FIG.

(2) Data Configuration Example A data configuration example of various types of information stored in each DB in the text information DB 110, the relationship information DB 120, and the graph information DB 130 will be described below.

FIG. 3 is a diagram showing a data configuration example of clue phrase information. The clue phrase information 300 shown in FIG. 3 is data stored in a list format in the clue phrase DB 111, and is composed of one or more clue phrase records (line data). Each cue phrase record holds information about one cue phrase. The clue phrase record is composed of fields of clue phrase ID 301 , target data category 302 , target expression category 303 , generation method 304 , clue phrase content 305 , degree of certainty 306 , and matching pattern 307 .

The clue phrase ID 301 holds the identifier of the clue phrase (cue phrase ID). The target data category 302 holds information indicating the category of target data targeted for keyword extraction using the clue phrase. For example, clue phrases are used when extracting keywords from either "manual" or "dialogue log."

The target expression category 303 holds information indicating the type of keyword to be extracted using the clue phrase. Specifically, "attribute name" is written for the attribute name keyword, "attribute value" is written for the attribute value keyword, and "final answer" is written for the final answer keyword.

The generation method 304 holds information indicating the generation method of the clue phrase. Specifically, if the clue phrase is an initial clue phrase input from the user terminal 200, it is described as "initial", and if it is a clue phrase obtained as a result of generating a new clue phrase, it is described as "complementary". be done.

The cue phrase content 305 holds the content of the cue phrase in text format. The degree of certainty 306 holds information (degree of certainty) regarding the likelihood of a clue phrase when the method 304 of generating the clue phrase is “complement”.

The collation pattern 307 holds information indicating the collation pattern of the target expression in keyword extraction using the clue phrase. In this example, the collation pattern 307 designates the target expression with a unique text and a regular expression (in [ ], the type of word is designated as a variable part). By holding such a collation pattern 307, it is possible to use the collation pattern and character string extraction by regular expression processing to perform keyword extraction for extracting words corresponding to variable portions as keywords. In FIG. 3, forms obtained by morphological analysis such as "common noun", "temporal noun", and "proper noun" are specified as the collation pattern 307, but the notation of the collation pattern 307 does not necessarily correspond to this. It is not limited, and may be a phrase obtained by syntactic analysis, a NamedEntity obtained by named entity extraction, or an HTML tag format obtained by an HTML parser.

FIG. 4 is a diagram showing a data configuration example of attribute value keyword information. The attribute value keyword information 400 shown in FIG. 4 is data stored in the attribute value DB 112 in list form, and is composed of one or more attribute value records (line data). The attribute value record is composed of fields of attribute value ID 401 , attribute value keyword 402 , target data position information 403 , certainty 404 , and selection source clue phrase ID 405 .

The attribute value ID 401 holds the identifier (attribute value ID) of the extracted attribute value keyword. The attribute value keyword 402 holds the text of the extracted attribute value keyword.

The target data position information 403 holds information indicating the extraction position of the attribute value keyword in the target data when the attribute value keyword is extracted from the target data based on the matching pattern 307 . For example, "13th line" shown in FIG. 4 means that the attribute value keyword is extracted from the 13th line of the manual, and "6th utterance" means that the attribute value keyword is extracted from the 6th utterance of the dialogue log. It means that the keyword has been extracted.

The degree of certainty 404 holds information (degree of certainty) regarding certainty as an attribute value keyword. The selection source clue phrase ID 405 holds the ID information (corresponding to the clue phrase ID 301 in FIG. 3) of the clue phrase record used for extracting the attribute value keyword.

FIG. 5 is a diagram showing a data configuration example of attribute name keyword information. The attribute name keyword information 500 shown in FIG. 5 is data stored in the attribute name DB 113 in list form, and is composed of one or more attribute name records (line data). The attribute name record is composed of fields of attribute name ID 501 , attribute name keyword 502 , target data position information 503 , certainty 504 , and selection source clue phrase ID 505 .

The attribute name ID 501 holds the identifier (attribute name ID) of the extracted attribute name keyword. The attribute name keyword 502 holds the text of the extracted attribute name keyword.

The target data internal position information 503 holds information indicating the extraction position of the attribute name keyword in the target data when the attribute name keyword is extracted from the target data based on the matching pattern 307 . For example, "12th line" shown in FIG. 5 means that the attribute name keyword is extracted from the 12th line of the manual, and "5th utterance" means that the attribute name keyword is extracted from the 5th utterance of the dialogue log. It means that the keyword has been extracted.

The degree of certainty 504 holds information (degree of certainty) regarding the likelihood of an attribute name keyword. The selection source clue phrase ID 505 holds the ID information (corresponding to the clue phrase ID 301 in FIG. 3) of the clue phrase record used to extract the attribute name keyword.

FIG. 6 is a diagram showing a data configuration example of final answer keyword information. The final reply keyword information 600 shown in FIG. 6 is data stored in the final reply DB 114 in list form, and is composed of one or more final reply records (line data). The final answer record is composed of fields of a final answer ID 601 , a final answer keyword 602 , positional information in target data 603 , a certainty factor 604 , and a selection source clue phrase ID 605 .

The final reply ID 601 holds the identifier (attribute name ID) of the extracted final reply keyword. The final answer keyword 602 holds the text of the extracted final answer keyword.

The target data internal position information 603 holds information indicating the extraction position of the final reply keyword in the target data when the final reply keyword is extracted from the target data based on the matching pattern 307 . For example, "14th line" shown in FIG. 6 means that the final answer keyword is extracted from the 14th line of the manual, and "7th utterance" is the final answer from the 7th utterance of the dialogue log. It means that the keyword has been extracted.

The degree of certainty 604 holds information (degree of certainty) regarding the certainty of the final answer keyword. The selection source clue phrase ID 605 holds the ID information (corresponding to the clue phrase ID 301 in FIG. 3) of the clue phrase record used to extract the final answer keyword.

FIG. 7 is a diagram showing a data configuration example of attribute value-final answer correspondence relationship information. The attribute value-final answer correspondence information 700 shown in FIG. 7 is data stored in a list format in the attribute value-final answer correspondence DB 121. From one or more attribute value-final answer correspondence records (line data) Configured. The attribute value-final answer correspondence record is composed of correspondence ID 701, attribute value ID 702, final answer ID 703, and certainty 704 fields.

The correspondence ID 701 holds the identifier of the correspondence between the attribute value keyword and the final answer keyword (attribute value-final answer correspondence). The attribute value ID 702 holds ID information (corresponding to the attribute value ID 401 in FIG. 4) of the attribute value keyword forming the corresponding relationship. The final reply ID 703 holds the ID information (corresponding to the final reply ID 601 in FIG. 6) of the final reply keyword forming the corresponding relationship. The degree of certainty 704 holds information about the certainty of matching of the corresponding relationship.

FIG. 8 is a diagram showing a data configuration example of attribute value-attribute name correspondence information. The attribute value-attribute name correspondence information 800 shown in FIG. 8 is data stored in a list format in the attribute value attribute name correspondence DB 122, and is composed of one or more attribute value-attribute name correspondence records (row data). Configured. The attribute value-attribute name correspondence record is composed of correspondence ID 801, attribute value ID 802, attribute value ID 803, and certainty 804 fields.

The correspondence ID 801 holds identifiers of correspondences (attribute value-attribute name correspondences) between attribute value keywords and attribute name keywords. The attribute value ID 802 holds ID information (corresponding to the attribute value ID 401 in FIG. 4) of the attribute value keyword forming the corresponding relationship. The attribute name ID 803 holds ID information (corresponding to the attribute name ID 501 in FIG. 5) of the attribute name keyword forming the corresponding relationship. The degree of certainty 804 holds information about the certainty of matching of the corresponding relationship.

FIG. 9 is a diagram showing a data configuration example of knowledge graph node information. The knowledge graph node information 900 shown in FIG. 9 is data stored in the knowledge graph node DB 131 in list form, and is composed of one or more knowledge graph node records (line data). The knowledge graph node record is composed of fields of node ID 901 , text 902 , and node category 903 .

The node ID 901 holds the identifiers (node IDs) of the nodes forming the knowledge graph. Text 902 holds the text of the reflection question or the text of the final answer. The node category 903 holds information identifying the category (kind) of the knowledge node. Specifically, for example, a category of "final answer" or "reflection question" is prepared.

FIG. 10 is a diagram showing a data configuration example of knowledge graph edge information. The knowledge graph edge information 1000 shown in FIG. 10 is data stored in the knowledge graph edge DB 132 in list format, and is composed of one or more knowledge graph edge records (line data). A knowledge graph edge record is composed of fields of an edge ID 1001 , an inflow source node ID 1002 , an inflow destination node ID 1003 , and a text 1004 .

The edge ID 1001 holds identifiers (edge IDs) of edges forming the knowledge graph. The inflow source node ID 1002 holds the ID information (corresponding to the node ID 901 in FIG. 9) of the inflow source knowledge graph node to which the edge is assigned in the knowledge graph. The inflow destination node ID 1003 holds the ID information (corresponding to the node ID 901 in FIG. 9) of the inflow destination knowledge graph node to which the edge is assigned in the knowledge graph. The text 1004 holds the text regarding the answer result of the reflection question.

(3) Dialogue Knowledge Generation Processing Hereinafter, the dialogue knowledge generation processing executed by the dialogue knowledge generation apparatus 100 will be described in detail while clearly indicating the functional units that are the main bodies of the processing.

FIG. 11 is a flow chart showing an example of an overall processing procedure of dialogue knowledge generation processing. In the dialogue knowledge generation process shown in FIG. 11, when the dialogue knowledge generation device 100 receives an input from the user terminal 200, the clue phrase estimation unit 141 executes the clue phrase estimation process of storing the clue phrase information 300 in the clue phrase DB 111. (step S1101).

Next, the attribute value keyword selection unit 142 executes attribute value keyword selection processing for extracting attribute value keywords from the target data using the clue phrase information 300 (step S1102). Attribute name keyword selection processing is executed to extract attribute name keywords from the target data using the information 300 (step S1103), and the final answer keyword selection unit 144 extracts the final answer keywords from the target data using the clue phrase information 300. Final answer keyword selection processing is executed (step S1104). It should be noted that the processing of steps S1102 to S1104 may be performed in a different order or may be performed in parallel.

Next, the attribute value-final answer associating unit 145 executes attribute value-final answer associating processing for associating the attribute value record with the final answer record (step S1105), and the attribute value-attribute name associating unit 146 , an attribute value-attribute name association process for associating the attribute value record with the attribute name record is executed (step S1106). It should be noted that the processing of steps S1105 and S1106 may be performed in a different order, or may be performed in parallel.

After that, the knowledge graph generation unit 147 executes knowledge graph generation processing for generating a structure of a knowledge graph indicating dialogue knowledge (step S1107). Specifically, in the knowledge graph generation process, by partially creating the knowledge graph node information 900 and the knowledge graph edge information 1000, the relationship between the nodes and edges that constitute the knowledge graph is specified, and the knowledge graph is generated. complete the structure.

Next, the reflection question generation unit 148 executes reflection question generation processing for associating the reflection question and the text of the answer (reflection text, answer text) with the knowledge graph generated in step S1107 (step S1108). Reflection text is added to the knowledge graph halfway generated by the knowledge graph generation unit 147 by executing the reflection question generation process, so that the knowledge graph as dialogue knowledge (dialogue flow model) with reflection text is completed. do.

Next, the knowledge graph visualization unit 150 executes knowledge graph visualization processing for visualizing the knowledge graph with the text of the feedback question generated through steps S1107 and S1108 (step S1109).

Thereafter, the dialogue knowledge generating apparatus 100 determines whether or not there is an additional input from the user terminal 200 (step S1110), and if there is an additional input (YES in step S1110), returns to step S1109, Depending on the content, the knowledge graph and reflection questions are revisited. On the other hand, if there is no additional input (NO in step S1110), the dialogue knowledge generation process ends.

The processing of each step in the dialogue knowledge generation processing will be described in detail below.

(3-1) Clue Phrase Estimation Processing FIG. 12 is a flowchart showing an example of the procedure of the clue phrase estimation processing. The clue phrase estimation process is the process of step S1101 in FIG. 11 and is executed by the clue phrase estimation unit 141 .

In the clue phrase estimation process, first, the clue phrase estimation unit 141 receives a business document (business manual in this example), an inquiry record (a dialogue log in this example), and an input of an initial clue phrase from the user terminal 200 (step S1201).

Next, the clue phrase estimation unit 141 generates a clue phrase record regarding the initial clue phrase received in step S1201, and adds it to the clue phrase information 300 stored in the clue phrase DB 111 (step S1202). In the clue phrase record generated in step S1202, "initial" is registered in the generation method 304, and input values are registered in other fields as appropriate. Note that a plurality of initial clue phrases may be input from the user terminal 200 in step S1201. In this case, the clue phrase estimation unit 141 newly generates new clue phrase records for the initial clue phrases received in step S1202. to generate Also, in step S1202, the clue phrase estimation unit 141 stores the business document and inquiry record received in step S1201 in an appropriate storage location (details are not shown) within the dialogue knowledge generating apparatus 100. FIG.

Next, the clue phrase estimation unit 141 determines whether or not there is an initial clue phrase from which no new clue phrase has been extracted (step S1203). If there is an initial clue phrase for which no new clue phrase has been extracted, for example, immediately after the initial clue phrase is registered in the clue phrase DB 111 in step S1202 (YES in step S1203), the process proceeds to step S1204. On the other hand, if new clue phrases have already been extracted for all initial clue phrases (NO in step S1203), the clue phrase estimation process ends.

In step S1204, the clue phrase estimation unit 141 extracts the target expression of the initial clue phrase based on the clue phrase record generated in step S1202. Then, the clue phrase estimation unit 141 extracts a new clue phrase from the extracted target expression (step S1205), generates a clue phrase record related to the extracted new clue phrase, and stores it in the clue phrase DB 111 (cue phrase information 300). (step S1206) and returns to step S1203.

The new clue phrase extracted in step S1205 is a clue phrase that complements the initial clue phrase. Specifically, it is assumed that the matching pattern 307 of the initial clue phrase is "when [common noun]" and the text "when making a contract" is extracted from the manual as the target expression in step S1204. In this case, in step S1205, by using a machine learning model or a parameter set for each word in advance, an expression that complements the expression "at the time of the contract" is searched. Assuming that an expression is extracted, a clue phrase such as "When [common noun]" is extracted as a new clue phrase and stored in the clue phrase DB 111 .

(3-2) Attribute Value Keyword Selection Processing FIG. 13 is a flow chart showing an example of the processing procedure of the attribute value keyword selection processing. The attribute value keyword selection process is the process of step S1102 in FIG. 11 and is executed by the attribute value keyword selection unit 142. FIG.

In the attribute value keyword selection process, first, the attribute value keyword selection unit 142 uses the clue phrase information 300 stored in the clue phrase DB 111 to extract clue phrase records whose target expression category 303 is "attribute value" (step S1301).

Next, the attribute value keyword selection unit 142 confirms the presence of clue phrase records for which extraction of attribute value keywords by the collation pattern 307 has not been executed among the clue phrase records extracted in step S1301 (step S1302).

If there is a clue phrase record for which attribute value keyword extraction has not been performed in step S1302 (YES in step S1302), the attribute value keyword selection unit 142 selects the target data indicated by the target data category 302 of the clue phrase record. Then, by searching using the collation pattern 307 as a search key, the corresponding attribute value keyword is extracted (step S1303), the certainty factor (certainty factor 404) of the extracted attribute value keyword is calculated (step S1304), and extracted. An attribute value record related to the attribute value keyword is generated, and the attribute value record is added to the attribute value keyword information 400 of the attribute value DB 112 (step S1305). After the end of step S1305, the process returns to step S1302.

Then, if there is no clue phrase record for which attribute value keyword extraction has not been executed in step S1302 (NO in step S1302), the attribute value keyword selection process is terminated.

(3-3) Attribute Name Keyword Selection Processing FIG. 14 is a flow chart showing an example of a processing procedure for attribute name keyword selection processing. The attribute name keyword selection process is the process of step S1103 in FIG. 11 and is executed by the attribute name keyword selection unit 143. FIG.

In the attribute name keyword selection process, first, the attribute name keyword selection unit 143 uses the clue phrase information 300 stored in the clue phrase DB 111 to extract clue phrase records whose target expression category 303 is "attribute name" (step S1401).

Next, the attribute name keyword selection unit 143 confirms the presence of clue phrase records for which attribute name keywords have not been extracted by the matching pattern 307 among the clue phrase records extracted in step S1401 (step S1402).

In step S1402, if there is a clue phrase record for which attribute name keyword extraction has not been performed (YES in step S1402), the attribute name keyword selection unit 143 selects the target data indicated by the target data category 302 of the clue phrase record. Then, by searching using the collation pattern 307 as a search key, the corresponding attribute name keyword is extracted (step S1403), the certainty factor (certainty factor 504) of the extracted attribute name keyword is calculated (step S1404), and extracted. An attribute name record relating to the attribute name keyword is generated, and the attribute name record is added to the attribute name keyword information 500 of the attribute name DB 113 (step S1405). After the end of step S1405, the process returns to step S1402.

Then, in step S1402, if there is no clue phrase record for which attribute name keyword extraction has not been executed (NO in step S1402), the attribute name keyword selection process is terminated.

(3-4) Final Answer Keyword Selection Processing FIG. 15 is a flowchart showing an example of the processing procedure of the final reply keyword selection processing. The final answer keyword selection process is the process of step S1104 in FIG. 11, and is executed by the final answer keyword selection unit 144.

In the final answer keyword selection process, first, the final answer keyword selection unit 144 uses the clue phrase information 300 stored in the clue phrase DB 111 to extract clue phrase records whose target expression category 303 is "final answer" (step S1501).

Next, the final answer keyword selection unit 144 confirms the existence of a clue phrase record for which the final answer keyword has not been extracted by the matching pattern 307 among the clue phrase records extracted in step S1501 (step S1502).

In step S1502, if there is a clue phrase record for which the final answer keyword has not been extracted (YES in step S1502), the final answer keyword selection unit 144 selects the target data indicated by the target data category 302 of the clue phrase record. Then, by searching using the matching pattern 307 as a search key, the corresponding final answer keyword is extracted (step S1503), the degree of certainty (certainty degree 604) of the extracted final answer keyword is calculated (step S1504), and extracted. A final response record for the final response keyword is generated, and the final response record is added to the final response keyword information 600 stored in the final response DB 114 (step S1505). After the end of step S1505, the process returns to step S1502.

Then, in step S1502, if there is no clue phrase record for which the final answer keyword has not been extracted (NO in step S1502), the final answer keyword selection process ends.

(3-5) Attribute Value-Final Answer Correlation Processing FIG. 16 is a flow chart showing an example of a processing procedure for attribute value-final reply correlation processing. The attribute value-final answer association process is the process of step S1105 in FIG. 11 and is executed by the attribute value-final answer association unit 145. FIG.

In the attribute value-final answer association process, first, the attribute value-final answer association unit 145 refers to the attribute value DB 112 and the final answer DB 114, and extracts the attribute values constituting the attribute value keyword information 400 stored in the attribute value DB 112. A record and a final answer record constituting the final answer keyword information 600 stored in the final answer DB 114 are extracted (step S1601).

Next, the attribute value-final answer association unit 145 determines whether or not there is a final answer record not associated with the attribute value record extracted in step S1601 among the final answer records extracted in step S1601. Determine (step S1602). The determination in step S1602 can be made, for example, by referring to the attribute value-final reply correspondence information 700 stored in the attribute value final reply correspondence DB 121. FIG.

In step S1602, if there is a final answer record that is not associated with an attribute value record (YES in step S1602), the attribute value-final answer associating unit 145 creates a final answer record that is not associated with , the corresponding attribute value record is specified (step S1603), an attribute value-final answer correspondence record is generated based on the specified result, and the attribute value-final answer correspondence relationship information 700 in the attribute value final answer correspondence DB 121 is Add (step S1604) and return to step S1602.

In step S1603, when identifying the attribute value record corresponding to the final answer record, the attribute value-final answer correlating unit 145 searches for related pairs based on the proximity of the positional information of each keyword. More specifically, the attribute value-final answer associating unit 145 uses the target data position information 603 of the final answer record to be associated and the target data position information of each attribute value record extracted in step S1601. 403 to specify the attribute value record indicating the closest position information as the associated attribute value record. Note that an attribute value record whose target data is different from the final answer keyword may be excluded from the matching search target.

Then, in step S1602, if there is no final answer record that is not associated with an attribute value record (NO in step S1602), the attribute value-final answer association processing ends.

(3-6) Attribute Value-Attribute Name Correlation Processing FIG. 17 is a flow chart showing an example of a processing procedure for attribute value-attribute name correlation processing. The attribute value-attribute name association processing is the processing of step S1106 in FIG. 11, and is executed by the attribute value-attribute name association unit 146.

In the attribute value-attribute name association processing, first, the attribute value-attribute name association unit 146 refers to the attribute value DB 112 and the attribute name DB 113, and extracts the attribute values constituting the attribute value keyword information 400 stored in the attribute value DB 112. Records and attribute name records that form the attribute name keyword information 500 stored in the attribute name DB 113 are extracted (step S1701).

Next, the attribute value-attribute name associating unit 146 determines whether or not there is an attribute value record not associated with the attribute name record extracted in step S1701 among the attribute value records extracted in step S1701. Determine (step S1702). The determination in step S1702 can be made by referring to the attribute value-attribute name correspondence information 800 stored in the attribute value attribute name correspondence DB 122, for example.

If there is an attribute value record that is not associated with an attribute name record in step S1702 (YES in step S1702), the attribute value-attribute name associating unit 146 , the corresponding attribute name record is specified (step S1703), the attribute value-attribute name correspondence record is generated based on the specified result, and the attribute value-attribute name correspondence relationship information 800 of the attribute value attribute name correspondence DB 122 is stored. Add (step S1704) and return to step S1702.

In step S1703, when identifying the attribute name record corresponding to the attribute value record, the attribute value-attribute name correlating unit 146 searches for related pairs based on the proximity of the position information of each keyword. More specifically, the attribute value-attribute name associating unit 146 combines the target data position information 403 of the attribute value record to be associated and the target data position information of each attribute name record extracted in step S1701. 503 to specify the attribute name record indicating the closest position information as the associated attribute name record. Note that an attribute name record whose attribute value keyword and target data are different may be excluded from the matching search target.

Then, in step S1702, if there is no attribute value record that has not been associated with an attribute name record (NO in step S1702), the attribute value-attribute name association processing ends.

(3-7) Knowledge Graph Generation Processing FIG. 18 is a flowchart illustrating an example of a processing procedure for knowledge graph generation processing. The knowledge graph generation process is the process of step S1107 in FIG. 11 and is executed by the knowledge graph generation unit 147. FIG.

In the knowledge graph generation process, first, the knowledge graph generation unit 147 refers to the attribute value-final answer correspondence information 700 stored in the attribute value final answer correspondence DB 121, and extracts the final answer keyword and the attribute value keyword that are in correspondence. A text (final answer text) is generated using this, and a new knowledge graph node having the generated text as a value is generated. At this time, for the new knowledge graph node, the knowledge graph generation unit 147 generates a knowledge graph node record in which the above text is held in the text 902 and "final answer" is held in the node category 903; is added to the knowledge graph node information 900 of the knowledge graph node DB 131 (step S1801). The text of the final answer keyword can be obtained from the final answer keyword 602 of the final answer keyword information 600 , and the text of the attribute value keyword can be obtained from the attribute value keyword 402 of the attribute value keyword information 400 .

For example, assuming that the final answer keyword whose correspondence is indicated by the attribute value-final answer correspondence information 700 is "telephone procedure" and the attribute value keyword is "spouse", the knowledge graph generation unit 147 generates "telephone procedure (spouse)" as the final answer text.

Next, the knowledge graph generation unit 147 refers to the knowledge graph edge information 1000 stored in the knowledge graph edge DB 132, and determines whether there are two or more knowledge graph node records in which the node ID 901 is not registered in the inflow destination node ID 1002. It is determined whether or not (step S1802). The process of step S1802 corresponds to the process of determining whether or not there are two or more nodes (child nodes) with no parent node set when considering the knowledge graph as a tree diagram. Therefore, when the process proceeds to step S1802 for the first time, no parent node is set for any of the knowledge graph nodes, so there is no node with ID901 registered in the inflow destination node ID1002. All knowledge graph node records will not apply.

Next, in step S1802, if there are two or more knowledge graph node records in which the node ID 901 is not registered in the inflow destination node ID 1002 (YES in step S1802), the knowledge graph generation unit 147 generates the knowledge graph node of the corresponding record. In order to determine the parent node of , a set of knowledge graph nodes to be generated as parent nodes is selected (step S1803). Specifically, the knowledge graph generation unit 147 selects knowledge graph nodes extracted as knowledge graph node records in which the node ID 901 is not registered in the inflow destination node ID 1002, based on the similarity of the character string of the text 902. , select some node-set. The process of step S1803 corresponds to the process of selecting a child node to which a common parent node is assigned from among child nodes for which no parent node is set.

Next, the knowledge graph generation unit 147 generates a knowledge graph node record that holds “request question” in the node category 903, assigns an ID to the node ID 901, and adds Add (step S1804). The process of step S1804 corresponds to the process of generating a parent node common to the set of child nodes selected in step S1803. However, in the knowledge graph node record created in step S1804, the text 902 (reflection text) in the knowledge graph node whose node category 903 is "reflection question" remains blank. The feedback text is registered in the text 902 during the feedback question generation process (see step S2006 in FIG. 20).

Next, for each of the knowledge graph nodes selected in step S1803, the knowledge graph generation unit 147 generates a knowledge graph edge record that holds the node ID 901 of the knowledge graph node in the inflow destination node ID 1003. After adding the ID 1002 and the edge ID 1001, they are added to the knowledge graph edge information 1000 of the knowledge graph edge DB 132 (step S1805). At this time, the node ID 901 (node ID of the parent node) of the knowledge graph node record generated in step S1804 is commonly registered as the inflow source node ID 1002 . A different ID is assigned to the edge ID 1001 for each generated knowledge graph edge record. The process of step S1805 corresponds to the process of generating edges connecting each of the child nodes selected in step S1803 and the parent node generated in step S1804. After the process of step S1805, the process returns to step S1802.

If there is only one knowledge graph node record in which the node ID 901 is not registered in the inflow destination node ID 1002 in step S1802 (NO in step S1802), the knowledge graph generation process ends. Supplementally, since the most root node (root node) in the knowledge graph of the tree diagram does not need to have an inflow source edge, the node ID 901 does not need to be registered in the inflow destination node ID 1003 . Therefore, when there is only one knowledge graph node record in which the node ID 901 is not registered in the inflow destination node ID 1002, all the other knowledge graph nodes are connected to the parent node, leaving only one knowledge graph node. can be used as the root node to finish the structural generation of the knowledge graph.

FIG. 19 is a diagram for visually explaining an image of knowledge graph generation processing. Nodes 1901 to 1907 shown in FIG. 19 are specific examples of knowledge graph nodes for which the node ID 901 was determined not to be registered in the destination node ID 1002 in step S1802 of FIG. A character string shown in each node represents the text of the node (corresponding to the text 902).

Nodes 1901 to 1907 do not have parent nodes set at the time when step S1802 is determined in the knowledge graph generation process, and node 1908 and edges 1911 to 1913 shown in FIG. 19 do not exist.

Next, in step S1803, nodes 1901 to 1903, for example, are selected as child nodes to which a common parent node is assigned from among the child nodes (ie, nodes 1901 to 1907) for which no parent node is set. In the case of FIG. 19, since nodes 1901-1903 have the common character string "telephone procedure", based on the similarity of the character strings, the child nodes to which these nodes assign a common parent node. selected as a set.

Next, in step S1804, a node 1908 is generated as a parent node common to the set of child nodes (nodes 1901 to 1903) selected in step S1803. As described above with reference to FIG. 18, the text 902 of node 1908 is blank at step S1804.

In step S1805, edges 1911 to 1913 are generated as edges connecting each of the child nodes (nodes 1901 to 1903) selected in step S1803 and the parent node (node 1908) generated in step S1804. be.

By performing the above processing, the nodes 1901 to 1903 whose parent nodes have not been set in the tree diagram shown in FIG. By repeating these processes, the knowledge graph generator 147 can generate edges connecting parent nodes to all nodes except for one root node, thereby completing the structure of the knowledge graph. .

(3-8) Reflection Question Generation Processing FIG. 20 is a flow chart showing an example of the processing procedure of the reflection question generation processing. The feedback question generation process is the process of step S1108 in FIG. 11, and is executed by the feedback question generation unit 148.

In the reflection question generation process, first, the reflection question generation unit 148 refers to the knowledge graph node information 900 stored in the knowledge graph node DB 131, and extracts knowledge graph node records whose node category 903 is "reflection question" ( step S2001).

Next, the reflection question generating unit 148 determines whether or not there is a knowledge graph node record in which the content of the reflection question (reflection text) is not registered in the text 902 among the knowledge graph node records extracted in step S2001. Determine (step S2002). The process of step S2002 corresponds to the process of searching for a parent node with an empty feedback question.

Next, in step S2002, if there is a knowledge graph node record in which the reflection text is not registered in the text 902 (YES in step S2002), the reflection question generation unit 148 generates the knowledge graph edge information stored in the knowledge graph edge DB 132. With reference to 1000, the knowledge graph edge record in which the node ID 901 of the knowledge graph node record is registered as the inflow source node ID 1002 is extracted, and the inflow destination node ID group in which the inflow destination node ID 1003 of each extracted record is collected is extracted. (step S2003). The process of step S2003 is a process of searching for an edge connecting a parent node with an empty feedback question as an inflow source node, and extracting the node ID group of the inflow destination node (child node) connected by the edge. For example, this corresponds to the process of extracting the node ID group of the set of knowledge graph nodes selected in step S1803 of FIG.

Next, the feedback question generating unit 148 refers to the knowledge graph node information 900 stored in the knowledge graph node DB 131, and summarizes the text 902 of the knowledge graph node record corresponding to the inflow destination node ID group extracted in step S2003. A text group is extracted (step S2004). The process of step S2004 corresponds to the process of extracting the text group of the destination node group whose node ID is extracted in step S2003.

Next, the feedback question generating unit 148 refers to the text group extracted in step S2004 and the attribute value-attribute name correspondence information 800 stored in the attribute value attribute name correspondence DB 122 (strictly speaking, attribute Also refer to value keyword information 400 and attribute name keyword information 500), attribute value keywords used to generate each text (text 902) in the text group, and attribute name keywords corresponding to the attribute value keywords are extracted. (Step S2005). Here, each text 902 described above is generated by the knowledge graph generation unit 147 in step S1801 of the knowledge graph generation processing shown in FIG. At step S2005, the attribute value keyword used for the generation is extracted. The process of step S2005 is a process of extracting attribute value keywords included in the text group extracted in step S2004 and their corresponding attribute name keywords. The attribute value keywords extracted in step S2005 are basically corresponds to the character string difference in the set of knowledge graph nodes selected based on the character string similarity in step S1803 of FIG. The attribute name keyword extracted in step S2005 is assumed to be common to each attribute value keyword.

Next, the reflection question generation unit 148 registers the attribute name keyword extracted in step S2005 as reflection text in the text 902 of the knowledge graph node record of the parent node where the reflection question determined in step S2002 is empty. Store in the graph node DB 131 . Further, the reflection question generating unit 148 registers the difference of the attribute value keywords extracted in step S2005 as the reply text to the reflection in the text 1004 of the knowledge graph edge record used for extracting the inflow destination node ID group in step S2003. , is stored in the knowledge graph edge DB 132 (step S2006). The attribute value keyword difference extracted in step S2005 can be extracted by a known method such as the Linux (registered trademark) diff command. Such processing of step S2006 registers attribute name keywords corresponding to attribute value keywords common (or similar) in the child node text group in the parent node text (review text), This corresponds to the process of registering the attribute value keyword difference in the text (answer text) of the edge connecting the parent node and the child node. After the processing of step S2006, the process returns to step S2002.

If there is no knowledge graph node record in which the reflection text is not registered in the text 902 in step S2002 (NO in step S2002), all knowledge graph node information 900 whose node category 903 is "reflection question" Since it means that all fields of the knowledge graph record and all the knowledge graph edge records forming the knowledge graph edge information 1000 have been registered, the feedback question generation process ends. That is, by completing the feedback question generation processing, a knowledge graph as dialogue knowledge with feedback text is completed.

In this embodiment, the knowledge graph generation process shown in FIG. 18 and the feedback question generation process shown in FIG. 20 may be alternately and repeatedly executed. 20 is executed each time the processing of steps S1802 to S1805 in FIG. 18 is completed once, and once the processing of steps S2002 to S2006 in FIG. 20 is completed, the process may return to step S1802 in FIG. .

FIG. 21 is a diagram for visually explaining an image of the feedback question generation process. Nodes 1901 to 1903 and 1908 and edges 1911 to 1913 shown in FIG. 21 correspond to the same reference numerals shown in FIG. That is, nodes 1901 to 1903 are knowledge graph nodes as child nodes to which a common parent node (node 1908) is assigned, node 1908 is a knowledge graph node to be a parent node of nodes 1901 to 1903, and edge 1911-1913 are knowledge graph edges connecting the node 1908 and the nodes 1901-1903.

When the determination in step S2002 is made in the reflection question generation processing, node 1908 is recognized as a knowledge graph node record in which no reflection text is registered in text 902 in FIG. Note that at this time, the text “relationship” is not registered in the node 1908 . Edges 1911 to 1913 have already been generated in the previous knowledge graph generation process (see FIG. 19), but the texts "spouse", "child", and "parent" for each edge are not registered at this time. No.

Next, in step S2003, the knowledge graph edge information 1000 is referenced to search for edges 1911 to 1913 that connect the parent node (node 1908) with an empty reflection question as an inflow source node. A node ID group of nodes 1901 to 1903) is extracted.

Next, in step S2004, with reference to the knowledge graph node information 900, a text group is extracted in which the texts 902 of the knowledge graph nodes (nodes 1901 to 1903) corresponding to the inflow destination node ID group extracted in step S2003 are summarized. In the case of FIG. 21, specifically, three texts are extracted: "telephone procedure (spouse)", "telephone procedure (child)", and "telephone procedure (parents)".

Next, in step S2005, attribute value keywords and corresponding attribute name keywords included in the text group extracted in step S2004 are extracted. In FIG. 21, combinations of extracted attribute name keywords and attribute value keywords are shown as attribute name keywords 2101 and attribute value keywords 2102 . Each attribute name keyword is obtained from the attribute name keyword 502 of the attribute name keyword information 500 , and each attribute value keyword is obtained from the attribute value keyword 402 of the attribute value keyword information 400 . Here, referring to FIG. 21, it can be seen that the attribute name keyword 2101 is common for "relationship" and the attribute value keyword 2102 differs for each node.

Next, in step S2006, a process of registering an attribute name keyword 2101 corresponding to an attribute value keyword common (or similar) in the text group of the child nodes (nodes 1901 to 1903) in the text 902 of the parent node (node 1908); A process of registering the difference of the attribute value keyword 2102 in the text group of the child node in the text 1004 of the edges (edges 1911 to 1913) connecting the parent node and the child node is performed. In the case of FIG. 21, since the attribute value keywords 2102 are different for each node, the attribute value keywords 2102 are used as they are for the differences in the attribute value keywords 2102 .

By performing the above processing, "relationship" can be registered as the reflection text for the node 1908 set as the parent node of the nodes 1901 to 1903 in the tree diagram shown in FIG. For edges 1911 to 1913 connecting node 1908 and nodes 1901 to 1903, "spouse", "child", and "parent" can be registered as respective answer texts. Thus, all fields in all records of knowledge graph node information 900 and knowledge graph edge information 1000 have been registered, and a knowledge graph with feedback text can be completed.

(3-9) Knowledge Graph Visualization Processing FIG. 22 is a diagram showing an example (part 1) of knowledge graph visualization. A display screen 2200 shown in FIG. 22 is an example of a knowledge graph visualization screen displayed and output to the output device 15 (or the user terminal 200) by the knowledge graph visualization unit 150 by the process of step S1109 in FIG. , is an example of visualizing a knowledge graph in a network format.

The display content of the display screen 2200 is formed based on the knowledge graph node information 900 and the knowledge graph edge information 1000 . For example, in the case of FIG. 22, a node 2201 is a knowledge graph node corresponding to a root node, and a node 2202 is one of knowledge graph nodes corresponding to child nodes of the node 2201 . Information on these nodes 2201 and 2202 can be obtained from the knowledge graph node information 900 . Specifically, the "time to subscribe to the service" indicated by the node 2201 and the "payment form" indicated by the node 2002 are the question texts (review texts) at each node, and are obtained from the text 902 of the knowledge graph node information 900. can be obtained. Information on the edge 2211 connecting the node 2201 and the node 2202 and information on the edge 2212 connecting the node 2202 and subsequent nodes can be obtained from the knowledge graph edge information 1000 . Specifically, "before 2019" indicated at the edge 2212 and "monthly" indicated at the edge 2212 are answer texts for each edge, which can be obtained from the text 1004 of the knowledge graph edge information 1000. .

FIG. 23 is a diagram illustrating an example (part 1) of visualization of a knowledge graph. A display screen 2300 shown in FIG. 23 is an example of a knowledge graph visualization screen displayed and output to the output device 15 (or may be the user terminal 200) by the knowledge graph visualization unit 150 by the process of step S1109 in FIG. , is an example of visualizing a knowledge graph in a table format.

As with the display screen 2200 shown in FIG. 22, the display contents of the display screen 2300 are formed based on the knowledge graph node information 900 and the knowledge graph edge information 1000 . It should be noted that the block diagram shown in the balloon in FIG. 23 shows the configuration relationship of each node represented in the table of the display screen 2300 for reference.

According to the display screen 2300 of FIG. 23, when a question is asked about the "relationship" held by the node 2301 with the node ID "5", the process branches to edges 2311 and 2312 according to the answer. If the answer is "spouse", it is connected to the node 2302 with the node ID "3" via the edge 2311, thereby guiding the telephone procedure for the spouse as the final answer. On the other hand, if the answer is "child", it is connected to the node 2303 with the node ID "4" via the edge 2312, and the telephone procedure for the child is guided as the final answer.

As shown in FIGS. 22 and 23, the dialogue knowledge generation apparatus 100 visualizes a knowledge graph representing dialogue knowledge (dialogue flow model) with reflection text and provides the user with the dialogue knowledge. In addition to being able to easily understand the overall picture of the dialogue, it is possible to grasp the necessary reflection questions according to the progress of the dialogue.

As described above, the dialogue knowledge generation device 100 according to this embodiment combines a plurality of data sources such as business documents (business manuals) and inquiry records (dialogue logs) to reduce manual workload and construction costs. A dialogue flow model (knowledge graph) with reflection text can be built automatically while reducing. 22 and 23, the interactive knowledge generating apparatus 100 can visualize the generated knowledge graph and provide it to the user. Dialogue management can be performed while referring to it, and even a non-skilled worker can proceed with dialogue work while asking back questions about unclear points and easily present an appropriate final answer to the user of the dialogue. can.

In addition, the present invention is not limited to the above-described embodiments, and includes various modifications. For example, the matching pattern 307 of the clue phrase information 300 may not be used in extracting attribute value keywords, attribute name keywords, and final answer keywords. In that case, for example, a prediction model generated in advance by machine learning may be used to extract keywords and calculate the degree of certainty. Further, for example, estimation of the correspondence relationship between the attribute value keyword and the attribute name keyword does not have to use the position information 403, 503 in the target data. In that case, for example, external open data or ontology may be used to specify that the superordinate concept of "child" is "relationship", and the process may be performed while narrowing down candidates for correspondence.

Further, the above-described embodiments are described in detail for easy understanding of the present invention, and are not necessarily limited to those having all the described configurations. Moreover, it is possible to add, delete, or replace a part of the configuration of the embodiment with another configuration.

Further, each of the above configurations, functions, processing units, processing means, and the like may be realized by hardware, for example, by designing a part or all of them using an integrated circuit. Moreover, each of the above configurations, functions, etc. may be realized by software by a processor interpreting and executing a program for realizing each function. Information such as programs, tables, and files that implement each function can be stored in recording devices such as memories, hard disks, SSDs (Solid State Drives), or recording media such as IC cards, SD cards, and DVDs.

Further, in the drawings, control lines and information lines are shown as necessary for explanation, and not all control lines and information lines are necessarily shown on the product. In fact, it may be considered that almost all configurations are interconnected.

10 computer 11 CPU
12 Memory 13 Storage Device 14 Input Device 15 Output Device 16 Network Device 17 Bus 100 Dialogue Knowledge Generator 110 Text Information DB
111 Clue phrase DB
112 attribute value database
113 Attribute name DB
114 Final Answer DB
120 Relationship information DB
121 Attribute Value Final Answer Correspondence DB
122 attribute value attribute name correspondence DB
130 graph information DB
131 knowledge graph node DB
132 knowledge graph edge DB
140 control unit 141 clue phrase estimation unit 142 attribute value keyword selection unit 143 attribute name keyword selection unit 144 final answer keyword selection unit 145 attribute value-final answer association unit 146 attribute value-attribute name association unit 147 knowledge graph generation unit 148 Reflective question generation unit 150 Knowledge graph visualization unit 200 User terminal 300 Clue phrase information 400 Attribute value keyword information 500 Attribute name keyword information 600 Final answer keyword information 700 Attribute value-final answer correspondence information 800 Attribute value-attribute name correspondence information 900 Knowledge graph node information 1000 Knowledge graph edge information 2200, 2300 Display screen

Claims (12)

A dialogue knowledge generation device for automatically generating a knowledge graph representing a dialogue flow model,
a keyword selection unit that selects, from target data including business documents and past dialogue records, attribute values that are keywords for dialogue and attribute names that indicate categories of attribute values, using clue phrases that indicate selection conditions for each;
an associating unit that determines a correspondence relationship between the attribute value selected by the keyword selecting unit and the attribute name based on their respective acquisition positions in the target data;
By extracting hierarchical relationships of a plurality of items having the attribute values or attribute names based on the correspondence determined by the association unit, and combining the plurality of items based on the extracted hierarchical relationships, a graph generator that generates a knowledge graph;
A dialogue knowledge generation device comprising: a clue phrase estimation unit that extracts a target expression from the target data using an initial clue phrase input from the outside and estimates a new clue phrase that complements the extracted target expression;
The keyword selection unit selects the attribute value and the attribute name from the target data using the initial clue phrase and the new clue phrase estimated by the clue phrase estimation unit. Item 1. The interactive knowledge generation device according to item 1. 2. The interactive knowledge generation apparatus according to claim 1, further comprising a knowledge graph visualization unit that outputs the knowledge graph generated by the graph generation unit in a visible form. 3. The knowledge graph is represented by a structure in which nodes holding texts of questions or final answers in dialogue and edges holding texts of answers to the questions are hierarchically connected. 2. The interactive knowledge generation device according to 1. The graph generation unit
a knowledge graph generation unit that generates the structure of the knowledge graph;
5. The method according to claim 4, further comprising: a feedback question generation unit that adds the content of the question in the dialogue and the content of the answer to the question to the structure of the knowledge graph generated by the knowledge graph generation unit. interactive knowledge generator. In generating the knowledge graph, the knowledge graph generation unit selects a set of child nodes based on the similarity of text held from a plurality of the nodes for which a parent node to be connected is not created, and selects a set of child nodes. 6. The interactive knowledge generation apparatus according to claim 5, wherein a common parent node is created for sets. 7. The dialogue knowledge generating apparatus according to claim 6, wherein said reflection question generation unit assigns said attribute name corresponding to a similar portion of text in said set of child nodes to said parent node as reflection text. The reflection question generation unit assigns the attribute value corresponding to the text difference in the set of child nodes to the edge connecting the child node having the difference and the parent node, as an answer text. 7. The interactive knowledge generation device according to claim 6. The keyword selection unit further selects a final answer to the dialogue from the target data using the clue phrase,
6. The method according to claim 5, wherein the associating unit further determines a corresponding relationship between the attribute value selected by the keyword selecting unit and the final answer based on the respective acquisition positions in the target data. An interactive knowledge generator as described. The knowledge graph generation unit generates a final answer node based on the correspondence relationship between the attribute value and the final answer determined by the association unit, and combines the attribute value and the final answer having the correspondence relationship. 10. The dialogue knowledge generation apparatus according to claim 9, wherein is assigned to the node of the final answer as the final answer text. 2. The dialogue knowledge generating apparatus according to claim 1, wherein said target data is text data. A dialogue knowledge generation method by a dialogue knowledge generation device for automatically generating a knowledge graph representing a dialogue flow model, comprising:
The dialogue knowledge generation device extracts attribute values, which are keywords for dialogue, and attribute names representing categories of attribute values, from target data including business documents and past dialogue records, using clue phrases representing respective selection conditions. a keyword selection step to be selected;
an association step in which the dialogue knowledge generation device determines a correspondence relationship between the attribute value selected in the keyword selection step and the attribute name based on their acquisition positions in the target data;
The dialogue knowledge generating device extracts a hierarchical relationship of a plurality of items having the attribute value or the attribute name based on the correspondence determined in the matching step, and extracts the plurality of items based on the extracted hierarchical relationship. a graph generation step of generating the knowledge graph by combining items;
A dialogue knowledge generation method characterized by comprising:

Images