Information processing device and information processing method

Abstract

An information processing device according to the present invention is provided with a knowledge-object generation unit and a dialog-pair generation unit. The knowledge-object generation unit extracts a plurality of terms from a semi-structured document and generates a knowledge graph representing relationships among the plurality of extracted terms. The dialog-pair generation unit generates a dialog repository including a plurality of dialog pairs on the basis of the knowledge graph. Each of the plurality of dialog pairs is a set of a question and a response to the question. A first term included in the question and a second term included in the response are two terms having a connected relationship in the knowledge graph.

Description

[Technical Field]

[0001] The present disclosure relates to an information processing device and an information processing method. [Background technology]

[0002] With the spread of smartphones and smart speakers, dialogue systems that respond appropriately to user utterances and user actions are widely used. Dialogue systems are required to respond to user utterances with appropriate speed and content.

[0003] For example, there is a known technology for generating dialogues based on knowledge, which generates responses to user questions and inquiries based on various types of knowledge stored in a knowledge base. Here, a knowledge base refers to a database of knowledge stored in a format that can be used by a computer. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2019-207648 Summary of the Invention [Problem to be solved by the invention]

[0005] However, in the above-mentioned conventional technology, a knowledge base is searched in response to a user utterance, and the knowledge obtained from the knowledge base is converted into a response sentence. However, as the size of the knowledge base increases, it becomes difficult to provide an appropriate response to the user utterance due to a delay in response caused by a decrease in the search speed of the knowledge base and the complexity of designing search rules.

[0006] Therefore, the present disclosure proposes an information processing device and an information processing method that can respond appropriately to user utterances even when the scale of the knowledge base is expanded. [Means for solving the problem]

[0007] According to the present disclosure, an information processing device includes a knowledge object generation unit that extracts a plurality of terms from a semi-structured document and generates a knowledge graph showing the relationships between the extracted plurality of terms, and a dialogue pair generation unit that generates a dialogue repository including a plurality of dialogue pairs based on the knowledge graph, each of the plurality of dialogue pairs being a set of a question sentence and a response sentence to the question sentence, and a first term included in the question sentence and a second term included in the response sentence are two terms that are connected in the knowledge graph.

[0008] Furthermore, according to the present disclosure, an information processing device includes an acquisition unit that acquires a question sentence from a user, and an inference unit that selects a dialogue pair, which is a combination of a question sentence including a first term extracted from a semi-structured document and a response sentence including a second term connected to the first term in a knowledge graph showing the relationship between multiple terms extracted from the semi-structured document, in accordance with the acquired question sentence from the user, and outputs the response sentence of the selected dialogue pair as a response sentence to the question sentence from the user. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram illustrating an overview of information processing executed by an information processing system according to an embodiment. [Figure 2] FIG. 1 is a diagram illustrating an overview of information processing executed by an information processing system according to an embodiment. [Figure 3] FIG. 2 is a diagram showing an outline of input and output data in information processing according to the embodiment. [Figure 4] FIG. 10 is a diagram illustrating an outline of processing in a case where a knowledge base specialized for a dialogue is not generated, which is different from the information processing according to the embodiment. [Figure 5]FIG. 2 is a block diagram illustrating an example of a functional configuration of a response generating device in the information processing system according to the embodiment. [Figure 6] 1 is a block diagram illustrating an example of a functional configuration of a dialogue repository generation device of an information processing system according to an embodiment. [Figure 7] 1 is a diagram for explaining a typical logical structure of a semi-structured document used for pattern matching when generating a knowledge graph from a semi-structured document in information processing according to an embodiment. FIG. [Figure 8] 1 is a diagram for explaining typical reference notations of semi-structured documents used for pattern matching when generating a knowledge graph from semi-structured documents in information processing according to an embodiment. FIG. [Figure 9] FIG. 10 is a diagram for explaining a specific example of converting semi-structured knowledge into a knowledge object in information processing according to the embodiment. [Figure 10] 10 is a diagram for explaining a specific example of creating a knowledge object for each knowledge classification in FIG. 9 in the information processing according to the embodiment. FIG. [Figure 11] 10 is a diagram for explaining a specific example of creating a knowledge object for each knowledge classification in FIG. 9 in the information processing according to the embodiment. FIG. [Figure 12] 10 is a diagram for explaining a specific example of creating a knowledge object for each knowledge classification in FIG. 9 in the information processing according to the embodiment. FIG. [Figure 13] 10 is a diagram for explaining a specific example of creating a knowledge object for each knowledge classification in FIG. 9 in the information processing according to the embodiment. FIG. [Figure 14] 10 is a diagram for explaining a specific example of creating a knowledge object for each knowledge classification in FIG. 9 in the information processing according to the embodiment. FIG. [Figure 15] 10 is a diagram for explaining a specific example of creating a knowledge object for each knowledge classification in FIG. 9 in the information processing according to the embodiment. FIG. [Figure 16] 10 is a diagram for explaining a specific example of creating a knowledge object for each knowledge classification in FIG. 9 in the information processing according to the embodiment. FIG. [Figure 17]FIG. 10 is a diagram illustrating an example of how to handle an irregular case in generating a group of dialogue pairs from a knowledge base in the information processing according to the embodiment. [Figure 18] 10 is a diagram illustrating a specific example of automatically completing content that is not explicitly stated in the terms and conditions by defining a persona in the information processing according to the embodiment. FIG. [Figure 19] FIG. 10 is a diagram illustrating a specific example of automatic completion of technical terms in the terms and conditions by replacing them with specific example sentences using numerical values calculated from persona parameters in the information processing according to the embodiment. [Figure 20] 10A and 10B are diagrams illustrating a specific example of automatic completion of content not explicitly stated in the terms and conditions using other knowledge of a semi-structured document in the information processing according to the embodiment. [Figure 21] 10A and 10B are diagrams illustrating a specific example of automatic completion of content not explicitly stated in the terms and conditions using other knowledge of a semi-structured document in the information processing according to the embodiment. [Figure 22] 10A and 10B are diagrams illustrating a specific example of automatic completion of content not explicitly stated in the terms and conditions using updated content of the terms and conditions in the information processing according to the embodiment. [Figure 23] FIG. 10 is a diagram illustrating selection of a dialogue pair using context information in information processing according to the embodiment. [Figure 24] FIG. 10 is a diagram illustrating selection of a dialogue pair using context information in information processing according to the embodiment. [Figure 25] 10 is a flowchart illustrating an example of a dialogue repository generation process as information processing according to the embodiment. [Figure 26] FIG. 10 is a diagram illustrating generation of a sentence vector according to the embodiment. [Figure 27] FIG. 10 is a diagram showing an example of dialogue pair data generated from a semi-structured document by the dialogue repository generation process according to the embodiment. [Figure 28] 10 is a flowchart illustrating an example of a response generation process as information processing according to the embodiment. [Figure 29] FIG. 2 is a block diagram illustrating an example of a hardware configuration of each device in the information processing system according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. In the following embodiments, the same components are designated by the same reference numerals, and redundant description will be omitted.

[0011] The present disclosure will be described in the following order: 1. Embodiment 1-1. Overview of information processing according to the embodiment 1-2. Functional configuration of information processing system according to the embodiment 1-2-1. Generating dialogue pairs from semi-structured documents 1-2-2. Completing dialogue pairs 1-2-3. Use of context information 1-3. Information processing procedure according to the embodiment 1-4.Hardware configuration 2. Effects of the information processing system according to the present disclosure

[0012] (1. Embodiment) [1-1. Overview of information processing according to the embodiment] An overview of information processing according to an embodiment will be described with reference to Fig. 1 to Fig. 4. Fig. 1 and Fig. 2 are diagrams showing an overview of information processing executed by an information processing system 1 according to an embodiment. Fig. 3 is a diagram showing an overview of input / output data in information processing according to an embodiment. Fig. 4 is a diagram showing an overview of processing in a case where a knowledge base KB specialized for a dialogue is not generated, which differs from the information processing according to an embodiment.

[0013] As shown in Fig. 1, information processing according to the embodiment includes a process S1 for generating a dialogue repository DR from a semi-structured document D1, and a process S2 for generating a response R to a question (utterance U) from a user. The process S1 for generating the dialogue repository DR is executed prior to the process S2 for generating the response R to the question from the user. Here, the semi-structured document D1 (dataset) used to generate the dialogue repository DR can be a contract document such as a contractual provision, a legal document, a medical book such as a home medicine book, a textbook, or the like, as appropriate.

[0014] The process S1 for generating a dialogue repository DR is a process for automatically generating a dialogue repository DR containing multiple dialogue pairs DP from a semi-structured document D1. More specifically, the process for generating a dialogue repository DR includes a process S11 for generating a "dialogue-specific" knowledge base KB from the semi-structured document D1, and a process S12 for generating a group of dialogue pairs DP (a dialogue pair group) from the generated knowledge base KB.

[0015] The process S11 for generating a knowledge base KB is a process for automatically generating a knowledge object such as a knowledge graph, i.e., a knowledge base KB, from a semi-structured document D1. More specifically, as shown in FIG. 1, the process S11 for generating a knowledge base KB generates a knowledge graph (knowledge object) by extracting multiple terms from the semi-structured document D1 using a model M1 and converting them into knowledge objects. In other words, a knowledge graph is a knowledge structure of the semi-structured document D1 that indicates the relationships between multiple terms extracted from the semi-structured document D1. The extraction of terms from the semi-structured document D1 and the conversion into knowledge objects can be achieved, for example, by pattern matching (model M1) on the semi-structured document D1. Alternatively, the extraction of terms from the semi-structured document D1 and the conversion into knowledge objects can also be achieved using a machine learning model (model M1). As an example of a machine learning model, a DNN (Deep Neural Network) such as Seq2seq (Sequence to sequence model) using RNN-LSTM (Recurrent Neural Networks-Long short-term memory) can be used, which is trained using data extracted by pattern matching from semi-structured document D1 as training data.

[0016] The process S12 for generating a group of dialogue pairs is a process for automatically generating a dialogue repository DR containing multiple dialogue pairs DP based on the generated knowledge graph, i.e., the knowledge base KB. The dialogue pairs DP are automatically generated from the knowledge base KB constructed as a knowledge graph based on a template (model M2). Here, each of the multiple dialogue pairs DP is a pair (QA pair, QA statement) consisting of utterances Q1, Q2, and Q3 (Q statements) as question statements and response statements A1, A2, and A3 (A statements) to the question statements (Q statements). The process S1 for generating the dialogue repository DR includes a process for converting (sentence vectorization) at least the question statements (utterances Q1, Q2, and Q3) of the dialogue pairs DP into sentence vectors (utterance vectors QV1, QV2, and QV3) using a DNN (model M3). The generated dialogue repository DR is registered as a dialogue repository of the QA inference model (model M5).

[0017] The process S2 of generating a response R to a question (utterance sentence U) from a user includes a process S21 of acquiring the question (utterance text, utterance sentence U) from the user. The process S21 of acquiring the question from the user includes a process of acquiring the output of the DNN (model M4) in response to the input of the utterance sentence U from the user as an utterance sentence vector UV. The process S2 of generating the response R also includes a process S22 of selecting a dialogue pair DP to be output as the response R (response text, response sentence) to the user from the dialogue repository DR. The process S22 of selecting the dialogue pair DP is performed based on a comparison between the sentence vector of the user utterance (utterance sentence vector UV) and the sentence vector of a Q sentence (utterance sentence vectors QV1, QV2, QV3) generated in advance in the dialogue repository. More specifically, the dialogue repository DR searches for utterance sentence vectors QV1, QV2, QV3 that are closest to the utterance sentence vector UV and are equal to or greater than a threshold (less than the difference specified by the threshold), and the response sentence corresponding to the retrieved dialogue pair DP is selected. In the example shown in Figure 1, a response sentence A1 corresponding to an utterance sentence vector QV1 is selected. Furthermore, the process S2 for generating a response R includes a process S23 for outputting the selected response R. In the process S23 for outputting the response R, the response sentence A1 selected as described above is output and presented to the user. In this way, the response R to the user can be uniquely determined based on the result of comparing the sentence vectors between the user utterance (utterance sentence U) and the question sentences (utterance sentences Q1, Q2, Q3) of the dialogue pair DP.

[0018] In addition, in the process S2 for generating the dialogue repository DR, as shown in FIG. 2, QA dialogues are automatically generated that concretely represent parts not explicitly described in the semi-structured document D1 (or the knowledge base KB). The process S2 for generating the dialogue repository DR generates dialogue pairs DP1 for knowledge within the terms and conditions from a dataset DS1 containing the semi-structured document D1 such as terms and conditions, and generates dialogue pairs DP2 that further include knowledge outside the terms and conditions from a dataset DS2 containing other general knowledge. Here, the written language-to-knowledge converter C1 illustrated in FIG. 2 is an example of the model M1 illustrated in FIG. 1. The knowledge-to-dialogue pair converter C2 illustrated in FIG. 2 is an example of the model M2 (or the models M2 and M3) illustrated in FIG. 1. The spoken text UT illustrated in FIG. 2 is an example of the text representing the spoken sentence U illustrated in FIG. 1. The response text RT illustrated in FIG. 2 is an example of the text representing the response sentence A1 illustrated in FIG. 1. This allows the dialogue pairs DP to be complemented based on general knowledge. The dialogue pair DP can be supplemented by using a persona, for example, to supplement the dialogue pair DP from which specific but unnecessary information equivalent to personal information has been removed without touching on personal information. Alternatively, the dialogue pair DP can be supplemented by supplementing dialogues about general knowledge not stated in the terms and conditions.

[0019] In the knowledge base KB according to the embodiment, the pair of question and response sentences of the dialogue pair DP are terms extracted from the semi-structured document D1 and are two terms that are connected in a knowledge graph. As illustrated in FIG. 3, it is assumed that "$(term)" and "$(explanation)" are connected in the knowledge base KB. In this case, a pair of a question sentence (SRC) "What is $(term)?" and a response sentence (TGT) "It means $(explanation)" is generated as the dialogue pair DP. More specifically, FIG. 1 illustrates an example of a user utterance (Q sentence) and an example response (A sentence) in the dialogue pair DP. The example of a user utterance (Q sentence) illustrated in FIG. 1 is "What is the Sokui-rei-Seiden-no-gi?" The example response (A sentence) is "The Sokui-rei-Seiden-no-gi is a ceremony that is the center of the Sokui-rei-Seiden-no-gi, in which the enthroned Emperor declares his accession to the throne both within and outside Japan." In this example, "Sokui-rei-seiden-no-gi" and "the ceremony at the center of the Sokui-rei, in which the enthroned Emperor declares his accession to the throne both inside and outside Japan," are examples of two terms that are connected in a knowledge graph. Dialogue pair DPs are generated in a similar manner for the other terms shown in Figure 3 ("$(term)" and "$(also known as)", "$(term)" and "$(concretization)").

[0020] On the other hand, as shown in FIG. 4, unlike the information processing according to this embodiment, a case will be described in which a knowledge base KB specialized for a dialogue is not generated. When a knowledge base KB specialized for a dialogue is not generated, it is when a dataset DS1' such as a contract clause is converted by a written language-to-knowledge converter C1' to create the knowledge base KB'. In this case, a user utterance (utterance text UT') is converted into a search query SQ' by an utterance-to-query converter C3', and the knowledge base KB' is searched using the search query SQ'. Then, the knowledge obtained by the search (search result SR') is converted into a response sentence by a result-to-response converter C4', and a response sentence (response text RT') is output.

[0021] Therefore, unlike the information processing according to this embodiment, if a knowledge base KB specialized for dialogue is not generated, the search speed will decrease as the size (data size) of the knowledge base increases. The decrease in search speed causes delays in agent responses, which require immediacy. Furthermore, the cost of designing search rules increases when the size of the knowledge base is large. In other words, since designers who are familiar with the knowledge and dialogue must implement or update the search rules for the knowledge base, the human resources required to convert a huge amount of knowledge into a knowledge base increase.

[0022] As described above, the information processing according to the embodiment, which converts semi-structured documents such as terms and conditions into a knowledge graph and generates dialogue pairs (dialogue repositories) based on the knowledge graph, can realize a system that can generate dialogue directly from text documents in an end-to-end manner. In other words, the information processing according to the embodiment can improve the problem of delayed response speed and realize scalable dialogue generation that does not require manual design.

[0023] [1-2. Functional configuration of information processing system according to embodiment] An example of the functional configuration of the information processing system 1 according to the embodiment will be described using Fig. 5 and Fig. 6. Fig. 5 is a block diagram showing an example of the functional configuration of the response generation device 10 of the information processing system 1 according to the embodiment. Fig. 6 is a block diagram showing an example of the functional configuration of the dialogue repository generation device 20 of the information processing system 1 according to the embodiment. Here, the response generation device 10 and the dialogue repository generation device 20 are each an example of an information processing device according to the present disclosure.

[0024] 5 and 6, an information processing system 1 according to an embodiment includes a response generation device 10, a dialogue repository generation device 20, and an external server 30. The response generation device 10, the dialogue repository generation device 20, and the external server 30 are communicably connected via a network N.

[0025] Of the information processing according to the embodiment, the process of generating a response to a question from a user (response generation process) is executed by a response generation device 10 shown in Fig. 5 and a dialogue repository generation device 20 shown in Fig. 6. On the other hand, of the information processing according to the embodiment, the process of generating a dialogue repository from a semi-structured document (dialogue repository generation process) is executed by the dialogue repository generation device 20 shown in Fig. 6.

[0026] The response generation device 10 is an information processing terminal (information processing device) that executes the response generation process according to the present disclosure. The voice recognition and voice response process executed by the response generation device 10 may be referred to as an agent function. The response generation device 10 may also be referred to as an agent device.

[0027] The response generation device 10 may be realized by various smart devices having information processing functions, or may be realized as a dedicated device. In other words, the response generation device 10 is not limited to a general-purpose computer, but may also be a smartphone, a tablet terminal, a smart speaker, a wearable device such as a wristwatch-type or eyeglass-type terminal, a smart home appliance such as a television, an air conditioner, or a refrigerator, a smart vehicle such as an automobile, a drone, or a home robot. The response generation device 10 as various devices having information processing functions functions as the response generation device 10 according to the present disclosure by executing a program (application) for realizing a response generation process.

[0028] As shown in FIG. 5, the response generation device 10 includes a sensor 11, an input unit 12, a communication unit 13, a storage unit 14, an acquisition unit 15, a response generation unit 16, and an output unit 17.

[0029] The sensor 11 detects various types of information. For example, the sensor 11 includes a microphone that collects the voice of the user, a camera that captures the user's actions as video, and a vital sensor that detects the user's vital signs such as body temperature, breathing, and heart rate.

[0030] The input unit 12 is a device for receiving various operations from a user, and is realized by, for example, a keyboard, a mouse, a touch panel, or the like.

[0031] The communication unit 13 is connected to the network N by wire or wirelessly, and transmits and receives information to and from the dialogue repository generating device 20 and the like via the network N.

[0032] The storage unit 14 is realized by, for example, a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory, or a storage device such as a hard disk or an optical disk. In the embodiment, the storage unit 14 has, for example, a user information table 141, model information 142, and a control program 143. The user information table 141 stores information about users who use the response generation device 10. The model information 142 stores models (functions) and parameters used in the response generation process. The control program 143 stores a program (application) for realizing the response generation process.

[0033] The acquisition unit 15 is a processing unit that acquires various types of information. The acquisition unit 15 may acquire, along with the input information, context information at the time the input information is input. The acquisition unit 15 includes a detection unit 151, a registration unit 152, and a reception unit 153. The detection unit 151 detects various types of information via the sensor 11. For example, the detection unit 151 detects a voice uttered by a user via a microphone, which is an example of the sensor 11. The detection unit 151 also detects image information used as context information via a camera, which is an example of the sensor 11. The detection unit 151 also detects vital signs, such as the user's body temperature, breathing, and heart rate, via a vital sensor, which is an example of the sensor 11. The registration unit 152 accepts registration from a user via the input unit 12. For example, the registration unit 152 accepts registration of a user profile (attribute information) of a user using the response generation device 10 via a touch panel or a keyboard. The attribute information is used, for example, to select a persona. The receiving unit 153 receives various information. For example, the receiving unit 153 receives from the dialogue repository generating device 20 a response sentence selected in the dialogue repository generating process.

[0034] The response generation unit 16 executes a process (response generation process) for generating a response to a question from a user. The response generation unit 16 includes a signal processing unit 161, a voice recognition unit 162, a response unit 163, and a transmission unit 164. The signal processing unit 161 performs signal processing on input information such as the user's utterance voice, images, and vital signs from the sensor 11. The voice recognition unit 162 recognizes the signal processed by the signal processing unit 161 as voice. The response unit 163 executes a process for acquiring a question sentence (utterance text, utterance sentence) from the user in the process for generating a response to a question from the user. The response unit 163 converts the user's utterance recognized by the voice recognition unit 162 into a sentence vector. The response unit 163 also executes a process for outputting a selected response sentence. The response unit 163 generates a response to the user based on the response sentence from the dialogue repository generation device 20 acquired by the acquisition unit 15. For example, the response unit 163 converts the acquired response sentence into voice data. The transmitting unit 164 transmits the sentence vector of the user utterance voice generated by the responding unit 163 to the communication unit 13. The transmitting unit 164 transmits the voice data converted by the responding unit 163 to the output unit 17.

[0035] The output unit 17 is a mechanism for outputting various information. The output unit 17 is, for example, a speaker. The output unit 17 outputs the voice data transmitted from the response generation unit 16. The output unit 17 may be a display that displays the response sentence acquired by the acquisition unit 15.

[0036] The dialogue repository generation device 20 is an information processing device that executes the dialogue repository generation process according to the present disclosure. The dialogue repository generation device 20 may be, for example, a general-purpose computer or a server. The dialogue repository generation device 20 functions as the dialogue repository generation device 20 according to the present disclosure by executing a program (application) for realizing the dialogue repository generation process.

[0037] As shown in FIG. 6, the dialogue repository generating device 20 includes a communication unit 21, a storage unit 22, an acquisition unit 23, and a dialogue repository generating unit 24.

[0038] The communication unit 21 is connected to the network N by wire or wirelessly, and transmits and receives information to and from the response generation device 10 and the external server 30 via the network N.

[0039] The storage unit 22 is realized by, for example, a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory, or a storage device such as a hard disk or an optical disk. In the embodiment, the storage unit 14 has, for example, a user information table 221, model information 222, a dialogue repository 223, and a control program 224. The user information table 221 is information acquired from the storage unit 14 of the response generation device 10, and stores information about users who use the response generation device 10. The model information 222 stores models (functions) and parameters used in the dialogue repository generation process. The model information 222 stores patterns used in the process of extracting a knowledge structure from a semi-structured document and generating a knowledge base. The model information 222 stores templates used in the process of generating a group of dialogue pairs from the knowledge base. The dialogue repository 223 stores the generated group of dialogue pairs. The control program 224 stores a program (application) for implementing the dialogue repository process. The storage unit 22 may store semi-structured documents.

[0040] The acquisition unit 23 is a processing unit that acquires various types of information. The acquisition unit 23 includes a receiving unit 231. The receiving unit 231 acquires text information such as semi-structured documents to be converted into knowledge objects and general knowledge from the external server 30. The receiving unit 231 also acquires information about users who use the response generation device 10, user utterance texts (question sentences, Q sentences), context information, and the like from the response generation device 10.

[0041] The dialogue repository generation unit 24 executes a process of generating a dialogue repository from a semi-structured document (dialogue repository generation process). The dialogue repository generation unit 24 includes a knowledge object generation unit 241, a dialogue pair generation unit 242, an inference unit 243, and a transmission unit 244. The knowledge object generation unit 241 executes a process of generating a knowledge base. The dialogue pair generation unit 242 executes a process of generating a dialogue repository including a group of dialogue pairs from the generated knowledge base. At this time, the inference unit 243 performs sentence vectorization with respect to at least the question sentence of the dialogue pair. In the process of generating a response to a question from a user, the inference unit 243 executes a process of selecting a dialogue pair to be output as a response sentence to the user from the dialogue repository 223 and a process of outputting the selected response sentence. The transmission unit 244, for example, transmits the response sentence selected in the dialogue repository generation process to the response generation device 10 via the communication unit 21.

[0042] The external server 30 is a device including a storage such as a non-volatile memory that stores semi-structured documents (datasets) to be converted into knowledge objects. The external server 30 or the storage such as a non-volatile memory may be configured to be directly connected to the dialogue repository generation device 20 without going through the network N. The external server 30 may also store the dialogue repository 223 generated by the dialogue repository generation device 20.

[0043] At least two of the response generation device 10, the dialogue repository generation device 20, and the external server 30 may be realized as a single device. Also, the response generation device 10, the dialogue repository generation device 20, and the external server 30 may each be realized by multiple devices.

[0044] [1-2-1. Generating dialogue pairs from semi-structured documents] Here, the process of generating a dialogue repository from a semi-structured document (dialogue repository generation process) will be described in more detail with reference to FIGS.

[0045] FIG. 7 is a diagram illustrating a typical logical structure of a semi-structured document used for pattern matching when generating a knowledge graph from a semi-structured document. While FIG. 7 illustrates the case of a contract, other semi-structured documents have similar logical structures. As shown in FIG. 7, a semi-structured document has a logical structure, such as a term definition ST1, a link definition ST2, a conditional statement ST3, an exception ST4, and a range ST5. The term definition ST1 is a logical structure that indicates a term, such as "A is B." The link definition ST2 is a logical structure that indicates a correlation (link), such as "As A, B." The conditional statement ST3 is a logical structure that indicates a condition, such as "If A, then B," "If A, then B, then C," or "If A, B. In this case, C." The exception ST4 is a logical structure that indicates an exception, such as "If A, then B. However, if C, then D." The range ST5 is a logical structure that indicates a range, such as "A, but including B," or "A, but excluding B."

[0046] FIG. 8 is a diagram illustrating typical reference notations for semi-structured documents used in pattern matching when generating a knowledge graph from a semi-structured document. While FIG. 8 illustrates the case of a conditional statement, other logical structures also have similar reference notations. As shown in FIG. 8, the logical structure of a conditional statement ST3 has reference notations such as parentheses E1, compound statement E2, simple statement E3, and nested statement E4. The parentheses E1 is a reference notation expressed using parentheses, such as "B (only if A)." The compound statement E2 is a reference notation expressed using a compound statement, such as "A. In this case, B." The simple statement E3 is a reference notation expressed using a simple statement, such as "If A, then B." The nested statement E4 is a reference notation expressed using nesting, such as "1) If B. a) A. b) A'."

[0047] Here, the conversion of semi-structured knowledge into knowledge objects (knowledge bases) will be explained in more detail with reference to Figs. 9 to 16. Fig. 9 is a diagram for explaining a specific example of the conversion of semi-structured knowledge into knowledge objects. Fig. 9 relates to terms and conditions as semi-structured documents (semi-structured knowledge). Figs. 10 to 16 are diagrams for explaining specific examples of the conversion of knowledge objects for each knowledge classification in Fig. 9. Note that the classifications, patterns, templates, etc. shown in Figs. 9 to 16 are examples, and can be changed, added, or deleted as appropriate.

[0048] The knowledge classification (large) and knowledge classification (small) illustrated in Fig. 9 are each an example of a node in a knowledge graph. In the example illustrated in Fig. 9, the knowledge classification (large) includes the items "Term," "Summary," "Instructions," "Payment Obligation," "Right," "Prohibition," "Action," "Reference," "Scope," "Due Date," "People," "Calculation Formula," "Condition," "Application," and "Other."

[0049] The knowledge classification (minor) for "terms" includes the items "aliases," "definitions," and "interpretations." The knowledge classification (minor) for "summary" includes the items "terms summary," "article summary," and "paragraph summary." The knowledge classification (minor) for "instructions" includes the items "payment instructions," "submission instructions," "change instructions," "notification instructions," "designation instructions," and "billing instructions." The knowledge classification (minor) for "payment obligation" includes the items "paid," "no payment," and "possible payment." The knowledge classification (minor) for "rights" includes the items "claim right," "option," "right to change," "right to terminate," "right to cancel," "right to rescind," "right to prepay," "right to repay," "right to revive," "loan terms," "right to restore," "right to inquire," "approval permission," and "right to repay." The knowledge classification (minor) for "prohibition" includes the items "prohibition of revival," "prohibition of cancellation," "prohibition of modification," "prohibition of opposition," "prohibition of use," and "prohibition of disclosure." The minor knowledge category for "action" includes the items "confirmation," "issuance," "discount," "continuation," "notification," "handling," "processing," "refund," "registration," "setting," "repayment," "judgment," "addition," and "top-up." The minor knowledge category for "reference" includes the item "as follows." The minor knowledge category for "scope" includes the items "includes" and "limited." The minor knowledge category for "due date" includes the items "expiration date," "time of liability," "grace period," "registration period," "renewal date," "base date," and "effective date." The minor knowledge category for "person" includes the items "payee," "contractor," and "agent." The minor knowledge category for "formula" includes the items "metal definition," "calculation method," "minimum amount," "deduction," "provision," "appropriation," "ratio," "round-down," and "deduction." The minor knowledge category for "condition" includes the items "amount," "payment," and "notification." The knowledge classification (small) for "application" includes the items "application," "applicable mutatis mutandis," and "not applicable." The knowledge classification (small) for "other" includes the items "responsibility," "jurisdiction," "invalidity," "extinction," "no dividend," and "no refund."

[0050] Figure 10 illustrates knowledge objectification and dialogue pairing for a "term" node. Figure 10 shows an example of a pattern for extracting the knowledge classification "term" from a semi-structured document. An example of an "alternative name" pattern is "$(term) (so-called "$(alternative name)")". Examples of "definition" patterns are "$(definition) (hereinafter referred to as "$(term)")", "$(term) (referring to $(definition). The same applies below.)", and "$(term) is $(definition)." An example of a "rereading" pattern is "$(term) is read as "$(rereading)". " and "$(term 1), $(term 2), $(term 3), $(term 4), $(term 5), $(term 6) should be read as $(replacement 1), $(replacement 2), $(replacement 3), $(replacement 4), $(replacement 5), $(replacement 6) respectively." An example of a "range" pattern is "For □□, $(range) is $(term)."

[0051] As described above, a semi-structured document is a document written using a certain logical structure and reference notation. Here, as shown in Figures 9 and 10, each knowledge classification corresponds to each node in the knowledge graph. Furthermore, a pattern (model information 222) according to the logical structure and reference notation is defined for each knowledge classification (node).

[0052] For example, as shown in FIG. 10, suppose that a description "cervical syndrome" (also known as "whiplash") is extracted from a semi-structured document using the pattern "$(term) (also known as "$(also known as "also known as ")")" of "term" and "also known as "also known as "). Here, the underlined parts of the pattern (model information 222) used in pattern matching are deleted when creating a knowledge graph. In this case, "cervical syndrome" and "(also known as "whiplash")" are connected as alternative names for the term in the knowledge graph, as shown in FIG. 10. In this way, the knowledge object generation unit 241 can extract the knowledge structure of a semi-structured document by performing pattern matching using the pattern (model information 222) according to the logical structure and reference notation.

[0053] Next, the generation of a group of dialogue pairs from the generated knowledge base will be explained in more detail with reference to Fig. 10. Fig. 9 illustrates an example of a "QA type," which is a type of dialogue pair generated from the value (term) of each knowledge classification, and an example of a dialogue pair generated for each knowledge classification. Here, in each dialogue pair illustrated in Fig. 9, the underlined part indicates the part of the template (model information 222) used to generate the dialogue pair.

[0054] The dialogue pair generation unit 242 generates a dialogue pair by inserting, for example, "terms" and "alternative names" connected in the knowledge graph into a template (model information 222), as shown in Fig. 10. Specifically, the dialogue pair generation unit 242 inserts "cervical syndrome" into the template to generate a Q sentence, "What is cervical syndrome?" Similarly, the dialogue pair generation unit 242 inserts "what is so-called 'whiplash'" into the template to generate an A sentence, "what is so-called 'whiplash'."

[0055] FIG. 11 illustrates an example of knowledge objectification and dialogue pairing for an "outline" node. FIG. 11 shows an example of a target for extracting a knowledge classification "outline" from a semi-structured document. FIG. 12 illustrates an example of knowledge objectification and dialogue pairing for an "instruction" node and a "conditional instruction" node. FIG. 13 illustrates an example of knowledge objectification and dialogue pairing for a "payment obligation" node and a "conditional payment obligation" node. FIG. 14 illustrates an example of knowledge objectification and dialogue pairing for a "right" node and a "conditional right" node. FIG. 15 illustrates an example of knowledge objectification and dialogue pairing for a "prohibition" node and a "conditional prohibition" node. FIG. 16 illustrates an example of knowledge objectification and dialogue pairing for an "action" node and a "conditional action" node. In these cases, the knowledge object generation unit 241 can generate a knowledge graph by performing pattern matching according to the pattern of the target to be extracted, in the same manner as described with reference to FIG. 10. For example, when a condition clause is extracted as shown in FIGS. 12 to 15, a "conditional XX" knowledge classification (node) is generated. For example, the "conditional right" node shown in Fig. 12 is connected to the description (term) of the condition clause and the "right" node related to the description of the main clause. Also, the dialogue pair generation unit 242 can generate dialogue pairs by inserting the terms connected in the knowledge graph into the template (model information 222) in the same way as explained using Fig. 10.

[0056] In this way, the dialogue pair generation unit 242 generates a dialogue-specific knowledge base from the semi-structured document, and then generates dialogue pairs from the knowledge base. However, when converting a semi-structured document into dialogue Q&A sentences, irregular cases that cannot be generated using templates alone may occur. Figure 17 shows an example of how to handle irregular cases when generating dialogue pairs from a knowledge base. As shown in Figure 17, for example, there may be differences in wording between the policy terms and the FAQ accompanying the policy terms. Figure 17 illustrates an example where the policy terms state "daily hospitalization benefit amount," but the FAQ uses "guaranteed amount," and another example where the policy terms state "when making a claim under this article," but the FAQ uses "when canceling." Here, the terms extracted from the policy terms can be paired with dialogues using templates. Therefore, the dialogue pair generation unit 242 generates additional questions using FAQ terms (guaranteed amount, cancel) as question variations. In this case, a question using a term from the FAQ can be generated as a dialogue pair for a question using a term extracted from the terms and conditions, for example, and the same response. Furthermore, as shown in FIG. 17, there may be cases where the terms and conditions describe cases in which payment will be made, but do not describe cases in which payment will not be made. In such cases, the dialogue pair generation unit 242 generates additional questions about cases other than those described in the terms and conditions as questions for negative cases. Responses for negative cases can be generated from the connections in the knowledge graph, just as when a template is used.

[0057] [1-2-2. Completing dialogue pairs] 18 to 22, the completion of dialogue pairs in converting semi-structured knowledge into knowledge objects (knowledge bases) will be described in more detail. As described above with reference to FIG. 2, the dialogue pair generation unit 242 can generate Q&A dialogues that automatically specify parts that are not explicitly stated in the semi-structured document (or knowledge base).

[0058] FIG. 18 is a diagram illustrating a specific example of automatically completing content not explicitly stated in the policy by defining a persona. FIG. 18 shows an example of correspondence between a user utterance (question), a response automatically generated from policy knowledge as a response to the question, and a response automatically generated by defining a persona according to a contract example. FIG. 19 is a diagram illustrating a specific example of automatically completing technical terms in the policy by replacing them with specific example sentences using numerical values calculated from persona parameters. In the example shown in FIG. 19, the persona includes the following item examples: "insured person," "gender," "payment limit per hospitalization," "daily hospitalization benefit amount," "insurance period," "premium payment period," and "individual monthly premium." It is assumed that parameters are predefined for each of these items. In the examples shown in FIGS. 18 and 19, the underlined portions are dialogue Q&A sentences automatically generated from policy knowledge as described above. The double-underlined portions are sentences embodied using the persona and are dialogue Q&A sentences to be completed. The persona parameters are stored, for example, in the user information table 221. For example, as shown in FIG. 18 , the dialogue pair generation unit 242 automatically generates a response sentence, "It will be paid to the policyholder," from the policy condition knowledge as a response sentence corresponding to a user utterance (question sentence), "Who will be paid?" On the other hand, the dialogue pair generation unit 242 automatically generates a response sentence to which the expression, "For example, if you yourself are the policyholder, it will be paid to you," is added according to the persona of the policyholder. For example, as shown in FIGS. 18 and 19 , the dialogue pair generation unit 242 generates a specific response sentence for the term "paid insurance premiums" based on the defined persona parameters and the calculation method (calculation formula) for paid insurance premiums stated in the policy condition, and automatically adds the specified QA dialogue. In other words, the dialogue pair generation unit 242 calculates the amount of paid insurance premiums according to the defined persona parameters. Then, the dialogue pair generation unit 242 automatically adds a specific response sentence such as, "For example, if the customer has paid a monthly insurance premium of 4,610 yen for five years, the total amount of insurance premiums paid will be 280,000 yen. This 280,000 yen is the premium already paid." Here, the term "premium already paid" and the calculation method are connected in the knowledge graph.

[0059] Note that the persona parameters shown in FIG. 19 are an example and are not limited to these. Two or more personas may be defined. For example, the personas to be defined can be prepared with variations of typical parameters for the items of example contracts presented in pamphlets corresponding to the terms and conditions. Note that the variations of parameters can be determined, for example, by plotting the profiles of all policyholders and extracting typical parameters by clustering the plots.

[0060] In this way, by using personas, the dialogue pair generation unit 242 can generate easy-to-understand response sentences that are closer to the format in which a human would explain them, compared to when using a template to directly convert the clause knowledge into dialogue Q&A sentences, and add them to the dialogue repository.

[0061] Figures 20 and 21 are diagrams illustrating a specific example of automatically completing content not explicitly stated in the terms and conditions using other knowledge from a semi-structured document. Figure 20 shows an example of the correspondence between a user utterance (question sentence), a response sentence automatically generated from the terms and conditions knowledge as a response sentence corresponding to the question sentence, and a response sentence automatically generated using general knowledge. In the examples shown in Figures 20 and 21, the underlined parts are dialogue Q&A sentences automatically generated from the terms and conditions knowledge as described above. The double-underlined parts are sentences that are concretized using general knowledge and are dialogue Q&A sentences to be completed. As shown in Figure 20, it is assumed that a dialogue pair can be generated for the term "bone marrow transplant benefit" based on the knowledge graph. However, it is assumed that terms corresponding to "bone marrow transplant benefit," such as "public health insurance system" and "bone marrow transplant procedure," are not defined in the terms and conditions. Dialogue pairs for terms not stated in the terms and conditions cannot be generated using a knowledge base from the terms and conditions. Therefore, as shown in the example of FIG. 21, the knowledge object generation unit 241 creates a knowledge graph for terms not included in the terms and conditions using other general knowledge from semi-structured documents such as ontologies and Wikipedia. In the example of FIGS. 20 and 21, for example, it is obtained from the general knowledge that "bone marrow transplantation" means "a treatment in which healthy bone marrow cells from a donor are transplanted intravenously into patients with intractable blood diseases such as leukemia and aplastic anemia." In this way, it is possible to automatically add Q&A dialogues (dialogue pairs) embodied using general knowledge. Here, knowledge listed on websites on the Internet via the network N can also be used as general knowledge.

[0062] FIG. 22 is a diagram illustrating a specific example of automatically completing content not explicitly stated in the policy terms and conditions using updated policy terms and conditions. In the example shown in FIG. 22, the underlined portions indicate changes made as a result of the policy terms and conditions update. The double-underlined portions are sentences embodied using the updated policy terms and conditions, and are dialogue Q&A sentences to be completed. For example, for policy terms and conditions that allow policy terms and conditions to be updated and policyholders to be transferred, the dialogue pair generation unit 242 creates a knowledge base of the differences between the policy terms and conditions before and after the update. At this time, the dialogue pair generation unit 242 generates dialogue pairs for each of the policy terms and conditions before and after the update, and automatically generates dialogue pairs indicating that a new version after the update is available for the old version before the update, and adds these to the dialogue repository. FIG. 22 illustrates an example in which, following an update to the policy terms and conditions, the definition of the term "previously paid insurance premiums (equivalent amount)" in the policy terms and conditions at the time of the contract, which was "monthly premium equivalent amount × 12 × contract," has been updated to "monthly premium equivalent amount × 11 × contract" in the latest policy terms and conditions. In such a case, the dialogue pair generation unit 242 automatically generates a response sentence that complements the updated contents of the policy, such as, "For example, the premium already paid is ____ million yen. In addition, in the latest policy updated in April 2019, the premium has been changed from 12 months to 11 months..." Note that there may be cases where only part of the policy content is updated. In such a case, the dialogue pair generation unit 242 generates dialogue pairs for the contents of both the pre-update and post-update versions. At this time, the dialogue pair generation unit 242 inserts phrases such as "In addition, in the new clause..." as exemplified in FIG. 22. In this way, the dialogue pair can be complemented using the updated contents of the policy.

[0063] [1-2-3. Use of context information] Here, an example of a process for selecting a dialogue pair to be output as a response sentence to the user from the dialogue repository 223 will be described with reference to Figures 23 and 24. Figures 23 and 24 are diagrams for explaining the selection of a dialogue pair using context information.

[0064] 23 illustrates a case where semi-structured documents D1 are appropriately used, such as insurance policy terms and conditions, legal documents such as legal documents, medical documents such as home medicine documents, regulations such as administrative rules, and textbooks such as grammar books. In such cases, the response unit 163 of the response generation device 10 converts not only the user's utterance U but also context information UC1 such as a user profile (attribute information) and context information UC2 such as vital signs and images obtained by the sensor 11 into a text vector. In other words, the text vector output from the response generation device 10 to the dialogue repository generation device 20 is a document vector combining the user's utterance U and the context information UC1 and UC2, as shown in FIG. 23. For medical information, images of related documents such as medical certificates and notes, or symptom images obtained by photographing the symptoms to be investigated, can be appropriately used.

[0065] When context information UC1 and UC2 are used in addition to the user's utterance sentence U, as described above, multiple personas can be defined or general documents can be used as knowledge to complement the dialogue pair DP. However, in this case, depending on the context information UC1 and UC2 used, a large number of redundant response sentences may be generated. However, the inference unit 243 according to the embodiment can narrow down the dialogue pairs DP to be output as responses to the user based on the similarity of the sentence vectors from the response unit 163 of the response generation device 10. Specifically, the inference unit 243 preferentially selects a dialogue pair DP related to the insurance policy terms and conditions based on a user profile indicating the insurance policyholder. Alternatively, the inference unit 243 preferentially selects a dialogue pair DPB generated using persona B defined according to the user profile from among multiple personas A, B, C, . . . , X. This configuration allows the dialogue pair DPB to be selected based on the sentence vector of the user's utterance sentence U supplemented by the context information UC1 and UC2. This reduces the user's speaking burden.

[0066] FIG. 24 illustrates a case in which medical documents, such as home medical documents, are used as the semi-structured document dataset DS1. In the example shown in FIG. 24, the image-to-disease name converter C5 of the response unit 163 converts context information UC21 indicating a symptom image I10 into a disease name based on the medical information dataset DS3, generating image text UCT1 ("molluscum contagiosum"). The vital sensor-to-symptom converter C6 of the response unit 163 converts the body temperature "36.8°C" in context information UC22 indicating the output of a vital sensor into a symptom based on the vital sign dataset DS4, generating sensor text UCT2 ("normal temperature"). The inference unit 243 selects a response sentence (response text RT) to the user based on the sentence vector of the utterance text UT combining the context information UC21 and UC22, thereby realizing a pseudo-diagnosis-like mechanism. In other words, the technology according to the embodiment can be applied even when the dataset is medical information, and a system can be realized that automatically diagnoses corresponding symptoms based on medical knowledge using user utterances, images, and vital sensor information as input.

[0067] [1-3. Information Processing Procedure According to the Embodiment] Next, the procedure of information processing according to the embodiment will be described with reference to FIGS.

[0068] FIG. 25 is a flowchart showing an example of dialogue repository generation processing as information processing according to the embodiment.

[0069] The knowledge object generation unit 241 opens an arbitrary semi-structured document (S101). The knowledge object generation unit 241 inputs each sentence of the document into a written language-to-knowledge converter to generate a knowledge base (S102). Here, the written language-to-knowledge converter is an example of a model stored in the model information 222, as described above with reference to Figures 1 to 3, 6, etc. As an example, the written language-to-knowledge converter is a DNN such as Seq2seq, whose parameters are trained using data extracted from a semi-structured document by pattern matching so as to output a knowledge object in response to an input sentence of the semi-structured document. As another example, the written language-to-knowledge converter is a function designed to output a knowledge object in response to an input sentence of the semi-structured document by pattern matching. When not all of the document has been converted into a knowledge base (S103: No), the flow of Figure 25 repeats the processes of S101 and S102.

[0070] On the other hand, when all documents have been converted into a knowledge base (S103: Yes), the dialogue pair generation unit 242 searches for the relevant knowledge from the knowledge base and inputs it into the knowledge-dialogue converter to generate a pair of Q sentence and A sentence, i.e., a dialogue pair (S104). Here, the knowledge-dialogue converter is an example of a model stored in the model information 222, as described above using Figures 1 to 3, 6, etc. As an example, the knowledge-dialogue converter can sequentially generate dialogue pairs for each term by inserting the knowledge searched from the knowledge base into a template.

[0071] Furthermore, the dialogue pair generation unit 242 extracts terms to which example sentences outside the clause knowledge can be added from the generated Q&A sentences, i.e., from the dialogue pairs (S105). When a term to which an example sentence outside the clause knowledge can be added is extracted in the process of S105, the dialogue pair generation unit 242 adds a Q&A sentence that expands knowledge related to the term (complements the dialogue pair) (S106). Thereafter, the flow in Figure 25 repeats the processes of S104 to S106 until the entire knowledge base has been converted into Q&A sentences (S107: No), and ends when the entire knowledge base has been converted into Q&A sentences (S107: Yes).

[0072] Note that, when the Q sentence is generated in the processes of S104 and S106, the dialogue pair generation unit 242 may input the generated Q sentence into a sentence-to-vector converter to generate a sentence vector for search. FIG. 26 is a diagram for explaining the generation of sentence vectors. As an example, as shown in FIG. 26, the dialogue pair generation unit 242 converts the Q sentence into a sentence vector using a sentence-to-vector converter (models M3 and M4). As the sentence-to-vector converter, for example, BERT (Bidirectional Encoder Representations from Transformers), which is a trained large-scale model (machine learning model), can be used. The sentence-to-vector converter is an example of a model stored in the model information 222. Note that the means for text vectorization is not limited to this, and other machine learning models such as Word2vec and ELMo may also be used.

[0073] Fig. 27 shows an example of data of QA sentences (dialogue pairs) generated from semi-structured documents (knowledge) by the dialogue repository generation process described using Fig. 25. In this way, the dialogue repository generation process according to the embodiment makes it possible to automatically generate QA sentences (dialogue pairs) from semi-structured documents (knowledge).

[0074] Fig. 28 is a flowchart showing an example of a response generation process as information processing according to the embodiment. The flow in Fig. 28 generates a response to a user utterance using data of a Q&A sentence (dialogue pair) generated from a semi-structured document (knowledge) by the dialogue repository generation process shown in Fig. 27.

[0075] The response unit 163 inputs image data obtained by the sensor 11 into an image-to-text converter and converts it into text data (S201). Here, the image-to-text converter is an example of a model stored in the model information 142, as described above with reference to Figures 23 and 24, etc. As an example, the image-to-text converter is a machine learning model whose parameters are determined so as to output text data in response to input of image data. The response unit 163 inputs sensing data, such as vital sensor data, obtained by the sensor 11 into a sensing log-to-text converter and converts it into text data (S202). Here, the sensing log-to-text converter is an example of a model stored in the model information 142, as described above with reference to Figures 23 and 24, etc. As an example, the sensing log-to-text converter is a machine learning model whose parameters are determined so as to output text data in response to input of sensing data. In addition, the response unit 163 converts other context information, such as a user profile, into text data (S203). Thereafter, the response unit 163 combines any user-uttered text data with the converted text data, as described above with reference to Figures 23 and 24, and generates a sentence vector using a sentence-to-vector converter. This sentence-to-vector converter is an example of a model stored in the model information 142. Note that various machine learning models such as BERT, Word2vec, and ELMo can be used as the sentence-to-vector converter, as appropriate.

[0076] If the Q sentence generated from the knowledge base has not been converted in the processes of S104 and S106 of the dialogue repository generation process described above with reference to Fig. 25, the dialogue pair generation unit 242 may input the Q sentence generated from the knowledge base to a sentence-to-vector converter to generate a sentence vector for search (S205). At this time, the A sentence may be further converted into a sentence vector.

[0077] The dialogue pair generation unit 242 searches for and selects a Q sentence that has the closest vector distance to the sentence vector corresponding to the user utterance from the response unit 163 from among the Q sentences corresponding to the dialogue repository (S206). Thereafter, the dialogue pair generation unit 242 transmits an A sentence corresponding to the selected Q sentence to the response generation device 10 as a response. The response generation device 10 presents the A sentence from the dialogue pair generation unit 242 to the user. Thereafter, the flow in FIG. 28 ends.

[0078] In this way, according to the response generation process according to the embodiment, it is possible to select an appropriate response from a Q&A sentence (dialogue pair) automatically generated from a semi-structured document (knowledge).

[0079] Each machine learning model according to the embodiment is a parameterized composite function obtained by combining multiple functions and is defined by a combination of multiple adjustable functions and parameters. The machine learning model may be any parameterized composite function defined by a combination of multiple adjustable functions and parameters. The machine learning model may be a convolutional neural network (CNN) or a fully connected network. The parameters of the trained machine learning model are stored in, for example, the storage unit 14 and the storage unit 22.

[0080] [1-4. Hardware Configuration] 29 is a block diagram showing an example of the hardware configuration of each device in the information processing system 1 according to the embodiment. Information devices such as each device (the response generation device 10, the dialogue repository generation device 20, and the external server 30) in the information processing system 1 according to the embodiment described above are realized by, for example, a computer 1000 having the configuration shown in FIG.

[0081] The computer 1000 includes a CPU 1100, a RAM 1200, a ROM (Read Only Memory) 1300, a HDD (Hard Disk Drive) 1400, a communication interface 1500, and an input / output interface 1600. The components of the computer 1000 are connected via a bus 1050 so as to be able to communicate with each other.

[0082] The CPU 1100 operates and controls each unit based on programs stored in the ROM 1300 or the HDD 1400. For example, the CPU 1100 loads programs stored in the ROM 1300 or the HDD 1400 into the RAM 1200 and executes processing corresponding to the various programs.

[0083] The ROM 1300 stores boot programs such as a Basic Input Output System (BIOS) executed by the CPU 1100 when the computer 1000 is started, and programs that depend on the hardware of the computer 1000 .

[0084] The HDD 1400 is a computer-readable recording medium that non-temporarily records programs executed by the CPU 1100 and data used by such programs. Specifically, the HDD 1400 is a recording medium that records the control programs 143, 224 according to the present disclosure, which are examples of program data 1450.

[0085] The communication interface 1500 is an interface for connecting the computer 1000 to the network N or an external network 1550 (e.g., the Internet). For example, the CPU 1100 receives data from other devices and transmits data generated by the CPU 1100 to other devices via the communication interface 1500.

[0086] The input / output interface 1600 is an interface for connecting the input / output device 1650 and the computer 1000. For example, the CPU 1100 receives data from an input device such as a keyboard or a mouse via the input / output interface 1600. The CPU 1100 also transmits data to an output device such as a display, a speaker, or a printer via the input / output interface 1600. The input / output interface 1600 may also function as a media interface for reading programs and the like recorded on a predetermined recording medium. Examples of media include optical recording media such as a DVD (Digital Versatile Disc) or a PD (Phase Change Rewritable Disk), magneto-optical recording media such as an MO (Magneto-Optical disk), tape media, magnetic recording media, and semiconductor memories.

[0087] For example, the CPU 1100 of the computer 1000 executes a control program 143 loaded onto the RAM 1200 to realize each function of the response generation device 10. For example, the CPU 1100 of the computer 1000 executes a control program 224 loaded onto the RAM 1200 to realize each function of the dialogue repository generation device 20. The HDD 1400 stores the control programs 143 and 224 according to the present disclosure and data in the storage units 14 and 22. The CPU 1100 reads and executes program data 1450 from the HDD 1400, but as another example, the CPU 1100 may obtain these programs from another device via an external network 1550.

[0088] (2. Effects of the Information Processing System According to the Present Disclosure) The dialogue repository generation device 20 (information processing device) includes a knowledge object generation unit 241 and a dialogue pair generation unit 242. The knowledge object generation unit 241 extracts a plurality of terms from a semi-structured document and generates a knowledge graph showing the relationships between the extracted plurality of terms. The dialogue pair generation unit 242 generates a dialogue repository including a plurality of dialogue pairs based on the knowledge graph. Each of the plurality of dialogue pairs is a set of a question sentence and a response sentence to the question sentence. A first term included in the question sentence and a second term included in the response sentence are two terms that are connected in the knowledge graph.

[0089] This allows the dialogue repository generation device 20 (information processing device) to automatically generate Q&A sentences (dialogue pairs) from semi-structured documents (knowledge). Therefore, even if the scale of the knowledge base is expanded, the complexity of search rule design can be suppressed, and appropriate responses to user utterances can be made.

[0090] In the dialogue repository generation device 20 (information processing device), the knowledge object generation unit 241 generates a knowledge graph that further includes terms extracted from other knowledge information of the semi-structured document, and the dialogue pair generation unit 242 complements the dialogue pairs using terms extracted from other knowledge information of the semi-structured document.

[0091] This allows the dialogue repository generation device 20 (information processing device) to complement dialogue pairs using general knowledge that is not included in semi-structured documents such as terms and conditions, and therefore to generate appropriate responses to questions from users.

[0092] In the dialogue repository generating device 20 (information processing device), the dialogue pair generating unit 242 complements the dialogue pairs using at least one persona.

[0093] This makes it possible to generate response sentences that are closer to the way a human would explain things and are easier for users to understand, compared to when clause knowledge is directly converted into dialogue Q&A sentences using a template.

[0094] The dialogue repository generation device 20 (information processing device) includes an acquisition unit 23 and an inference unit 243. The acquisition unit 23 acquires a question sentence from a user. The inference unit 243 selects a dialogue pair, which is a combination of a question sentence including a first term extracted from a semi-structured document and a response sentence including a second term connected to the first term in a knowledge graph indicating the relationship between multiple terms extracted from the semi-structured document, in accordance with the acquired question sentence from the user, and outputs the response sentence of the selected dialogue pair as a response sentence to the question sentence from the user.

[0095] This allows a response to a user's question to be selected from the dialogue repository without searching the knowledge base when a response is generated in response to a user utterance. Therefore, even if the knowledge base is expanded, delays in response due to a decrease in the knowledge base search speed can be suppressed, allowing for an appropriate response to user utterances.

[0096] In the dialogue repository generating device 20 (information processing device), the inference unit 243 selects a dialogue pair based on the vector distance of the document vector between the question sentence of the dialogue pair and the question sentence from the user.

[0097] This makes it possible to suppress response delays that accompany a decrease in the speed of searching the knowledge base, even when the scale of the knowledge base is expanded, and therefore makes it possible to respond appropriately to user utterances.

[0098] In the dialogue repository generating device 20 (information processing device), the inference unit 243 selects a dialogue pair based on a document vector generated by combining a question sentence from a user and context information corresponding to the question sentence from the user.

[0099] This allows the question sentence uttered by the user to be supplemented with context information, thereby reducing the burden on the user's utterance and enabling the generation of a more appropriate response sentence to the question sentence from the user.

[0100] In the dialogue repository generating device 20 (information processing device), the acquiring unit 23 further acquires context information corresponding to the question sentence from the user, and the inferring unit 243 preferentially selects a dialogue pair corresponding to the acquired context information.

[0101] This makes it possible to uniquely select dialogue pairs that are close to the context of the user profile, even when a large number of dialogue pairs are generated using persona and context information, thereby reducing the burden on the user's utterances and generating more appropriate responses to questions from the user.

[0102] In the dialogue repository generating device 20 (information processing device), the context information includes at least one of an image, a vital sign, and a user profile.

[0103] This reduces the burden on the user when speaking, and also makes it possible to generate a more appropriate response to a question from the user.

[0104] The information processing method includes extracting a plurality of terms from a semi-structured document, generating a knowledge graph showing the relationships between the extracted plurality of terms, and generating a dialogue repository including a plurality of dialogue pairs based on the knowledge graph, each of the plurality of dialogue pairs being a pair of a question statement and a response statement to the question statement, and a first term included in the question statement and a second term included in the response statement being two terms in a connected relationship in the knowledge graph.

[0105] This allows the dialogue repository generation device 20 (information processing device) to automatically generate Q&A sentences (dialogue pairs) from semi-structured documents (knowledge). Therefore, even if the scale of the knowledge base is expanded, the complexity of search rule design can be suppressed, and appropriate responses to user utterances can be made.

[0106] The information processing method includes acquiring a question from a user, selecting a dialogue pair, which is a combination of the question including a first term extracted from a semi-structured document and a response sentence including a second term connected to the first term in a knowledge graph showing the relationship between multiple terms extracted from the semi-structured document, in accordance with the acquired question from the user, and outputting the response sentence of the selected dialogue pair as a response sentence to the question from the user.

[0107] This allows a response to a user's question to be selected from the dialogue repository without searching the knowledge base when a response is generated in response to a user utterance. Therefore, even if the knowledge base is expanded, delays in response due to a decrease in the knowledge base search speed can be suppressed, allowing for an appropriate response to user utterances.

[0108] The effects described in this specification are merely examples and are not limiting, and other effects may also be present.

[0109] The present technology can also be configured as follows. (1) a knowledge object generation unit that extracts a plurality of terms from a semi-structured document and generates a knowledge graph that indicates the relationships between the extracted plurality of terms; a dialogue pair generation unit that generates a dialogue repository including a plurality of dialogue pairs based on the knowledge graph; Equipped with each of the plurality of dialogue pairs is a pair of a question sentence and a response sentence to the question sentence; a first term included in the question sentence and a second term included in the response sentence are two terms that are connected in the knowledge graph; Information processing device. (2) the knowledge object generation unit generates the knowledge graph, which further includes terms extracted from other knowledge information of the semi-structured document; the dialogue pair generation unit complements the dialogue pairs using terms extracted from other knowledge information of the semi-structured document; The information processing device according to (1) above. (3) The information processing device according to (1) or (2), wherein the dialogue pair generation unit complements the dialogue pair using at least one persona. (4) an acquisition unit that acquires a question sentence from a user; an inference unit that selects a dialogue pair, which is a combination of a question sentence including a first term extracted from a semi-structured document and a response sentence including a second term connected to the first term in a knowledge graph indicating a relationship between a plurality of terms extracted from the semi-structured document, in accordance with the acquired question sentence from the user, and outputs the response sentence of the selected dialogue pair as a response sentence to the question sentence from the user; An information processing device comprising: (5) The information processing device according to (4), wherein the inference unit selects the dialogue pair based on a vector distance of a document vector between a question sentence of the dialogue pair and a question sentence from the user. (6) The information processing device according to (4) or (5), wherein the inference unit selects the dialogue pair based on a document vector generated by combining a question sentence from the user and context information corresponding to the question sentence from the user. (7) the acquisition unit further acquires context information corresponding to a question sentence from the user; the inference unit preferentially selects a dialogue pair corresponding to the acquired context information; The information processing device according to any one of (4) to (6). (8) The information processing device according to (6) or (7), wherein the context information includes at least one of an image, a vital sign, and a user profile. (9) extracting a plurality of terms from a semi-structured document and generating a knowledge graph showing relationships between the extracted plurality of terms; generating a dialogue repository based on the knowledge graph, the dialogue repository including a plurality of dialogue pairs; Including, each of the plurality of dialogue pairs is a pair of a question sentence and a response sentence to the question sentence; a first term included in the question sentence and a second term included in the response sentence are two terms that are connected in the knowledge graph; Information processing methods. (10) Obtaining a question from a user; selecting a dialogue pair, which is a combination of a question sentence including a first term extracted from a semi-structured document and a response sentence including a second term connected to the first term in a knowledge graph showing the relationship between a plurality of terms extracted from the semi-structured document, in accordance with the acquired question sentence from the user, and outputting the response sentence of the selected dialogue pair as a response sentence to the question sentence from the user; An information processing method including: [Explanation of symbols]

[0110] 1. Information Processing Systems 10 Response Generation Device 11 Sensors 12 Input section 13 Communications Department 14 Storage section 141 User Information Table 142 Model Information 143 Control Program 15 Acquisition Department 151 Detection unit 152 Registration Department 153 Receiving unit 16 Response Generation Unit 161 Signal Processing Unit 162 Voice Recognition Unit 163 Response Section 164 Transmitter 17 Output section 20 Dialogue repository generator 21 Communications Department 22 Memory section 221 User Information Table 222 Model Information 223 Dialogue Repository 224 Control Program 23 Acquisition Department 231 Receiving unit 24 Interaction repository generation unit 241 Knowledge Object Generation Unit 242 Dialogue Pair Generation Unit 243 Reasoning part 244 Transmitter 30 External Servers

Claims

1. a knowledge object generation unit that extracts a plurality of terms from a semi-structured document and generates a knowledge graph that indicates the relationships between the extracted plurality of terms; a dialogue pair generation unit that generates a dialogue repository including a plurality of dialogue pairs based on the knowledge graph; Equipped with each of the plurality of dialogue pairs is a pair of a question sentence and a response sentence to the question sentence; a first term included in the question sentence and a second term included in the response sentence are two terms that are connected in the knowledge graph; Information processing device.

2. the knowledge object generation unit generates the knowledge graph, which further includes terms extracted from other knowledge information of the semi-structured document; the dialogue pair generation unit complements the dialogue pairs using terms extracted from other knowledge information of the semi-structured document; The information processing device according to claim 1 .

3. The information processing device according to claim 1 , wherein the dialogue pair generation unit complements the dialogue pairs using at least one persona.

4. an acquisition unit that acquires a question sentence from a user; an inference unit that selects a dialogue pair, which is a combination of a question sentence including a first term extracted from a semi-structured document and a response sentence including a second term connected to the first term in a knowledge graph indicating a relationship between a plurality of terms extracted from the semi-structured document, in accordance with the acquired question sentence from the user, and outputs the response sentence of the selected dialogue pair as a response sentence to the question sentence from the user; An information processing device comprising:

5. The information processing apparatus according to claim 4 , wherein the inference unit selects the dialogue pair based on a vector distance of a document vector between a question sentence of the dialogue pair and a question sentence from the user.

6. The information processing apparatus according to claim 4 , wherein the inference unit selects the dialogue pair based on a document vector generated by combining a question sentence from the user and context information corresponding to the question sentence from the user.

7. the acquisition unit further acquires context information corresponding to a question sentence from the user; the inference unit preferentially selects a dialogue pair corresponding to the acquired context information; The information processing device according to claim 4 .

8. The information processing device according to claim 7 , wherein the context information includes at least one of an image, a vital sign, and a user profile.

9. An information processing device, extracting a plurality of terms from a semi-structured document and generating a knowledge graph showing relationships between the extracted plurality of terms; generating a dialogue repository based on the knowledge graph, the dialogue repository including a plurality of dialogue pairs; Including, each of the plurality of dialogue pairs is a pair of a question sentence and a response sentence to the question sentence; a first term included in the question sentence and a second term included in the response sentence are two terms that are connected in the knowledge graph; Information processing methods.

10. An information processing device, Obtaining a question from a user; selecting a dialogue pair, which is a combination of a question sentence including a first term extracted from a semi-structured document and a response sentence including a second term connected to the first term in a knowledge graph indicating the relationship between a plurality of terms extracted from the semi-structured document, in accordance with the acquired question sentence from the user, and outputting the response sentence of the selected dialogue pair as a response sentence to the question sentence from the user; An information processing method including:

Images