Information processing system and method for vectorizing the meaning of words and searching for reasons, etc., based on vector similarity.

Abstract

This system provides an information processing system capable of outputting response data that aligns with the user's motivations or values. [Solution] The information processing system obtains content containing expressions representing reasons, intentions, backgrounds, or objectives from construction data as reason sentences, converts the obtained reason sentences into multidimensional embedding vectors using an embedding model to obtain reason vectors, generates reason data by associating the obtained reason sentences with the reason vectors, stores the generated reason data to construct a reason corpus, obtains content containing expressions representing reasons, intentions, backgrounds, or objectives from user query data as query reason sentences, converts the obtained query reason sentences into multidimensional embedding vectors using an embedding model to obtain query vectors, calculates the similarity between the query vectors and the reason vectors included in the reason corpus, and searches a predetermined number of reason data from the reason corpus based on the calculated similarity.

Description

Technical Field

[0001] The present invention generally relates to the generation of response data for user inquiry data.

Background Art

[0002] In recent years, with the development of natural language processing technology using large language models and embedding models, research and development of information processing systems that generate appropriate responses to user inquiries have been progressing. In particular, a technology that converts text data into multi-dimensional embedding vectors and retrieves relevant information based on similarity in vector space has attracted attention as enabling a search that takes into account semantic proximity compared to conventional keyword matching.

[0003] Conventional keyword-based search and rule-based search are suitable for extracting objective information such as "what", "where", and "how", but it is difficult to handle information related to the subjective intentions, reasons, backgrounds, or purposes of users such as "why are they asking that question". That is, generating responses along the motivations and values included in the inquiry has not been fully realized in the prior art.

[0004] Patent Document 1 discloses a system that uses a neural network to generate a response from context information and query input and provides an interpretable inference trace by repeating a selection step and a speculation step.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, the technology described in Patent Document 1 generates responses from contextual information using a selection neural network and an inference neural network, but it does not focus on processing the subjective intentions and motivations of the "why" contained in the user's question. Rather, its main focus is on generating inference traces based on facts and rules about the environment, making it difficult to generate semantic answers that align with the questioner's values ​​and motivations.

[0007] Based on the above, this invention proposes an information processing system, etc., that can output response data that aligns with the user's motivations or values. [Means for solving the problem]

[0008] To solve the above problems, the present invention provides an information processing system that outputs response data to user inquiry data, comprising: a first acquisition unit that acquires content including expressions representing reason, intention, background or purpose from construction data as reason sentences; a reason vectorization unit that converts the reason sentences acquired by the first acquisition unit into multidimensional embedding vectors using an embedding model to obtain reason vectors; a storage unit that generates reason data relating the reason sentences acquired by the first acquisition unit and the reason vectors converted from the reason sentences by the reason vectorization unit, stores the generated reason data and constructs a reason corpus; and from user inquiry data, The system comprises: a second acquisition unit that acquires content including expressions representing the reason, intention, background, or purpose as a query reason statement; a query vectorization unit that converts the query reason statement acquired by the second acquisition unit into a multidimensional embedding vector using the embedding model to obtain a query vector; a search unit that calculates the similarity between the query vector converted by the query vectorization unit and the reason vectors included in the reason corpus, and searches for a predetermined number of reason data from the reason corpus based on the calculated similarity; and an output unit that outputs response data to the query data based on the reason statement associated with the reason data searched by the search unit. [Effects of the Invention]

[0009] According to the present invention, it is possible to output response data that aligns with the user's motivations or values. Problems, means, and effects not mentioned above will be clarified by the following description of embodiments. [Brief explanation of the drawing]

[0010] [Figure 1] This figure shows an example of an information processing system according to the first embodiment. [Figure 2] This figure shows an example of an information processing infrastructure according to the first embodiment. [Figure 3] This figure shows an example of a data item table according to the first embodiment. [Figure 4] This figure shows an example of a data item table according to the first embodiment. [Figure 5] This figure shows an example of a flowchart illustrating the overall procedure according to the first embodiment. [Figure 6] This figure shows an example of a flowchart for the reason corpus construction process according to the first embodiment. [Figure 7] This figure shows an example of a flowchart for query processing according to the first embodiment. [Figure 8] This figure shows an example of a flowchart of the search process according to the first embodiment. [Figure 9] This figure shows a specific example of data according to the first embodiment. [Figure 10] This figure shows a specific example of data according to the first embodiment. [Modes for carrying out the invention]

[0011] (I) First Embodiment The configurations, procedures, or other elements disclosed below are provided to illustrate embodiments of the present invention and do not limit it. The descriptions based on the drawings are intended to aid understanding, and the shapes, arrangements, functions, etc., of elements may be omitted or simplified as necessary. The present invention is not limited to one or more embodiments disclosed, and can also be implemented by functionally or structurally equivalent means, or other means achieving similar technical objectives, to the extent that a person skilled in the art can understand from this specification. Unless otherwise specified, each component in this specification is construed as including "at least one." Furthermore, a singular form of a word includes the plural form, and a plural form of a word includes the singular form. In addition, terms used herein are not limited to a specific meaning, to the extent that they are clear from the context, and should be appropriately interpreted by a person skilled in the art. For example, the expression "including" is construed as meaning not limited to those listed.

[0012] The information processing system according to this embodiment performs a search process that handles expressions related to reasons, intentions, backgrounds, or purposes (hereinafter referred to as "reasons, etc.") contained in the inquiry data entered by the user. Unlike conventional search processes that are suitable for extracting objective information contained in the inquiry data, this information processing system performs a search process that focuses on expressions related to reasons, etc. contained in the inquiry data.

[0013] Conventional search processes are suitable for extracting objective information such as "what" and "where," but tend to return vague or general answers to questions about reasons, such as "why." This information processing system acquires expressions related to reasons as reason sentences, vectorizes these reason sentences, and uses them in searches, thereby improving the accuracy of information extraction that contributes to generating answer data aligned with the user's motivations or values.

[0014] This information processing system represents the meaning of words or sentences as numerical values in a multi-dimensional semantic space through a process of vectorizing the reason sentences. Based on the arrangement in the semantic space, this information processing system handles expressions containing ambiguity or polysemy and quantifies the proximity of context. Through a search process based on the proximity of context, this information processing system extracts information near the intention that is likely to be overlooked by conventional keyword matching or rule-based processing and uses it as the basis for answer data.

[0015] The inquiry reason sentence obtained in this information processing system is not limited to the case where the inquiry data explicitly contains an expression representing a reason, intention, background, or purpose. When the inquiry data is interpreted as a question that requires the presentation of reasons, etc., without extracting a part of the words contained in the inquiry data, it may be obtained as text generated with an expression representing a reason, intention, background, or purpose based on the semantic content of the inquiry data. For example, a question such as "Why should solar power generation be introduced?" does not contain the reason expression itself but is interpreted as a question that requests the presentation of the reason or purpose of introduction. Therefore, this information processing system generates an expression related to the reason for introduction or the purpose of introduction as an inquiry reason sentence based on the meaning of the inquiry data and uses this for the subsequent search process.

[0016] This information processing system may be applied to the purpose of alleviating the conflict of expressions in the dialogue space. This information processing system extracts an expression representing a reason, etc. as a reason sentence from target texts such as posted texts or comment texts, and represents the reason sentence as a reason vector. Based on the reason vector, this information processing system searches for an expression close to the intention, generates a presentation sentence including a neutral expression suppressing an aggressive expression or an assertive expression, and outputs the presentation sentence as answer data.

[0017] This information processing system may receive inquiry data including, as a specific example, "why solar power generation should be introduced". General search processing or general dialogue language processing includes superficial explanations such as the fact that it is renewable energy or that the environmental impact is low, and explanations regarding reasons and the like are likely to be mixed. This information processing system executes a search process specialized for expressions regarding reasons and the like, and generates, as response data, explanations that conform to the purposes such as reduction of greenhouse gas emissions and achievement of energy self-sufficiency, and the backgrounds such as energy security and transition to a decarbonized society.

[0018] The application target of this information processing system may include research support, policy formulation support, business decision-making support, medical support, or educational support. This information processing system uses a search process based on vectorization and similarity of reason sentences in a process of extracting the reasons or grounds of claims from a document group, a process of evaluating the necessity or consistency of policies, a process of extracting the aims or withdrawal reasons of decision-making, a process of extracting motives from consultation statements, or a process of extracting learning motives or reasons for confusion.

[0019] The information processing system 100 may obtain, as reason sentences, expressions representing the aims of decision-making, the background of decision-making, the reasons for decision-making, or the reasons for withdrawal from a document including the minutes of a meeting. The information processing system 100 may convert the reason sentences into reason vectors and execute a search process based on the similarity with the group of reason vectors included in the reason corpus. The information processing system 100 may present the similar reason sentences obtained by the search process as basis information and also present them as judgment materials for use in the explanation of decision-making. As the presentation of judgment materials, the information processing system 100 may generate a sentence explaining the aim or background supported by the reason sentence for the conclusion included in the minutes. As the generation of a sentence, the information processing system 100 may use the response generation process shown in FIG. 9 and FIG. 10 and generate a response text based on the reason sentence and the basis information.

[0020] The information processing system 100 may obtain, as reason sentences, expressions representing the necessity of the policy, the objective of the policy, the reasons for comparing it with alternatives, the background of the issue, or the basis for implementation from documents relating to the draft law or policy. The information processing system 100 may convert the reason sentences into reason vectors and perform a search process based on the similarity with the group of reason vectors included in the reason corpus. The information processing system 100 may present similar reason sentences obtained through the search process as supporting information and present them as decision-making material to be used to explain the necessity. As a presentation of decision-making material, the information processing system 100 may extract supporting information that reinforces the explanation of necessity for the claims included in the draft law or policy, and generate an explanatory text based on the supporting information. As a generation of the explanatory text, the information processing system 100 may use the response generation process shown in Figures 9 and 10 to generate a response text based on the reason sentences and supporting information.

[0021] The information processing system 100 may obtain, as reason sentences, expressions representing evidence supporting a claim, reasons based on experimental results, reasons based on relationships with prior research, or the basis for inferences leading to a conclusion, from documents including academic literature. The information processing system 100 may convert the reason sentences into reason vectors and perform a search process based on the similarity with a group of reason vectors included in the reason corpus. The information processing system 100 may present similar reason sentences obtained through the search process as evidence information and present them as the result of evidence extraction. As a presentation of the evidence extraction results, the information processing system 100 may generate an explanatory text showing the correspondence between the claim and the evidence information. As generating the explanatory text, the information processing system 100 may use the response generation process shown in Figures 9 and 10 to generate a response text based on the reason sentences and evidence information.

[0022] The information processing system 100 may present information that assists in making judgments based on expressions representing reasons, etc., contained in a group of documents, by obtaining reason statements, generating reason vectors, and performing search processing based on similarity. In addition to its use for generating responses, the information processing system 100 may also be used for explaining the objectives of decision-making, explaining the necessity of policies, and extracting the basis for claims by presenting information that assists in making judgments.

[0023] This information processing system may be configured to take into account the characteristics of the search method or data format. This information processing system adopts a vector format that enables searching based on semantic proximity, taking into account that tree structures or label structures are suitable for hierarchical organization and classification but have difficulty handling overlapping meanings across multiple categories, symbolic logic is suitable for inference and explicit judgment but tends to be difficult to design and maintain rules, and text sentences are highly expressive but tend to be difficult to search and classify. This information processing system performs semantic vector searching based on reason sentences, taking into account that keyword searches have difficulty handling context and intent, rule-based searches tend to lack flexibility, semantic vector searches are suitable for extracting intent similarity, and label matching is weak against ambiguous intents across multiple axes.

[0024] The following will be a detailed explanation using the drawings. In this specification, elements that are identical or functionally similar to the components shown in the drawings are denoted by the same reference numerals. The use of reference numerals in this specification is not limited to the specific embodiments shown, but is also applicable to various modifications including at least one component.

[0025] Figure 1 shows an example of an information processing system 100.

[0026] In the information processing system 100, the information processing infrastructure 110, the management terminal 120, and the client terminal 130 are connected to each other via the network 101, enabling communication between them.

[0027] The information processing infrastructure 110 is a computer system responsible for various information processing operations performed within the information processing infrastructure 110. The information processing infrastructure 110 consists of, for example, server devices, virtual machines, and computing resources in a cloud environment. The information processing infrastructure 110, equipped with the components shown in Figure 2, performs reason corpus construction and query processing, and generates and outputs response data to query data received from client terminals 130.

[0028] The management terminal 120 is a terminal device that performs management operations on the information processing infrastructure 110. The management terminal 120 consists of, for example, a personal computer, tablet terminal, or mobile terminal operated by an administrator. The management terminal 120 performs tasks such as registering construction data for building a reason corpus, issuing update instructions for the reason corpus, and checking the operating status of the information processing infrastructure 110.

[0029] The client terminal 130 is a terminal device that sends query data to the information processing infrastructure 110 and receives and displays the response data output from the information processing infrastructure 110. The client terminal 130 consists of, for example, a personal computer, tablet terminal, or mobile terminal operated by the user. For example, the client terminal 130 sends a natural language question sentence entered by the user as query data to the information processing infrastructure 110 and displays the response data received from the information processing infrastructure 110 on the screen.

[0030] Network 101 is a communication network connecting the information processing infrastructure 110, the management terminal 120, and the client terminal 130. Network 101 includes, for example, the Internet, a local area network, a mobile communication network, etc. The information processing infrastructure 110 communicates with the management terminal 120 and the client terminal 130 via Network 101.

[0031] In the information processing system 100, the information processing infrastructure 110 receives construction data from the management terminal 120, obtains content including expressions representing reasons from the text contained in the construction data as reason sentences, and converts the reason sentences into multidimensional embedding vectors using an embedding model to obtain reason vectors. The information processing infrastructure 110 generates reason data by associating the reason sentences and reason vectors, stores the reason data, and constructs a reason corpus.

[0032] In the information processing system 100, the information processing infrastructure 110 receives query data from the client terminal 130, obtains content including expressions representing reasons from the query data as a query reason statement, and converts the query reason statement into a multidimensional embedding vector using an embedding model to obtain a query vector. The information processing infrastructure 110 calculates the similarity between the query vector and the reason vectors included in the reason corpus, and searches for a predetermined number of reason data from the reason corpus based on the calculated similarity. Based on the reason statements associated with the retrieved reason data, the information processing infrastructure 110 generates response data for the query data and sends the response data to the client terminal 130.

[0033] For example, if a user on client terminal 130 inputs "Why should solar power generation be introduced?", client terminal 130 sends the question as query data to information processing platform 110. Information processing platform 110 obtains the query reason sentence from the query data and converts the query reason sentence into a query vector. Information processing platform 110 calculates the similarity between the query vector and the reason vectors included in the reason corpus and searches for reason data with high similarity. Based on the reason sentence associated with the searched reason data, information processing platform 110 generates answer data such as, "Solar power generation allows homes and facilities to be self-sufficient in electricity, leading to long-term reductions in electricity costs. Solar power generation contributes to reducing environmental impact by suppressing carbon dioxide emissions and is compatible with efforts toward carbon neutrality. Solar power generation is recommended because it can reduce fuel supply risks and can be used as an emergency power source during disasters." Client terminal 130 displays the answer data received from information processing platform 110.

[0034] Figure 2 shows an example of the information processing infrastructure 110.

[0035] The information processing infrastructure 110 comprises a processor 210, a storage device 220, and an interface device 230. The processor 210 is a device that performs data arithmetic and control processing. The storage device 220 is a device for storing programs, data, etc. Programs are read and executed by the processor. The interface device 230 is a device that sends and receives information with a network or other devices, or a device that performs information input and output between a user and a device, etc. The functions of the information processing infrastructure 110 (acquisition unit 221, reason vectorization unit 222, storage unit 223, query vectorization unit 224, search unit 225, output unit 226, etc.) are realized by software, hardware, or a combination thereof. The functions of the information processing infrastructure 110 can be realized, for example, by software execution on a general-purpose computer, hardware implementation using dedicated circuits, or distributed processing utilizing the cloud, virtual machines, etc. Some or all of the functions of the information processing infrastructure 110 may be shared by one or more devices.

[0036] The acquisition unit 221 switches its operating mode according to the type of input data to acquire content including expressions that represent reasons, etc. In construction mode, when construction data is input, the acquisition unit 221 acquires content including expressions that represent reasons, etc. from the construction data as a reason statement. For example, the acquisition unit 221 generates a prompt containing construction data, inputs the prompt into a large-scale language model, and acquires the text output from the large-scale language model, which includes expressions that represent reasons, etc., as a reason statement. In query mode, when query data is input, the acquisition unit 221 acquires content including expressions that represent reasons, etc., from the query data as a query reason statement. For example, the acquisition unit 221 generates a prompt containing query data, inputs the prompt into a large-scale language model, and acquires the text output from the large-scale language model, which includes expressions that represent reasons, etc., as a query reason statement. It should also be noted that the acquisition unit 221 may be referred to as the first acquisition unit and the second acquisition unit.

[0037] The reason vectorization unit 222 converts the reason sentences obtained by the acquisition unit 221 (construction mode) into multidimensional embedding vectors using an embedding model to obtain reason vectors. For example, the reason vectorization unit 222 inputs the reason sentences into the embedding model and obtains reason vectors as multidimensional numerical sequences that reflect the meaning of the reason sentences. The reason vectorization unit 222 may normalize the reason vectors, for example, to make them suitable for calculating similarity during searches.

[0038] The storage unit 223 generates reason data by associating the reason sentences acquired by the acquisition unit 221 (construction mode) with the reason vectors obtained by the reason vectorization unit 222, and stores the reason data to construct a reason corpus. The storage unit 223, for example, assigns identifiers and classification information to the reason data and stores it as a reason corpus in a format that includes reason sentences and reason vectors.

[0039] The query vectorization unit 224 converts the query reason statement obtained by the acquisition unit 221 (query mode) into a multidimensional embedding vector using an embedding model to obtain a query vector. For example, the query vectorization unit 224 inputs the query reason statement into the embedding model and obtains a query vector as a multidimensional numerical sequence that reflects the meaning of the query reason statement. The query vectorization unit 224 may normalize the query vector, for example, to make it suitable for calculating similarity during a search.

[0040] The search unit 225 calculates the similarity between the query vector obtained by the query vectorization unit 224 and the reason vectors included in the reason corpus, and searches for a predetermined number of reason data from the reason corpus based on the calculated similarity. For example, the search unit 225 calculates the cosine similarity between the query vector and the reason vectors and retrieves the reason data in descending order of similarity.

[0041] The output unit 226 generates response data to the query data based on the reason sentences associated with the reason data retrieved by the search unit 225, and outputs the response data. The output unit 226 may, for example, provide the query data and reason sentences as input to a large-scale language model and generate response data that includes descriptions of the reasons contained in the reason sentences.

[0042] Figure 3 shows an example of a data item table 300 that indicates the data items that make up the reason statement data handled by the information processing infrastructure 110.

[0043] The information processing infrastructure 110 constructs a reason corpus using reason sentences obtained from construction data and generates response data for query data. For each reason sentence data handled in the reason corpus construction process, the information processing infrastructure 110 manages it by associating it with at least an identifier, classification information (category), and reason sentence (Why text).

[0044] The information processing infrastructure 110 uses identifiers as information to uniquely identify reason statement data. The information processing infrastructure 110 also uses identifiers as keys to associate reason vectors with reason statements. For example, when the information processing infrastructure 110 stores reason vectors corresponding to reason statements in the storage device 220, it stores the reason vectors with identifiers attached. In search processing based on query data, the information processing infrastructure 110 uses identifiers to refer to reason statements from reason vectors.

[0045] The information processing infrastructure 110 uses classification information (categories) as information used to classify reason statement data. The information processing infrastructure 110 also uses classification information as auxiliary information when searching or managing reason statement data. For example, the information processing infrastructure 110 automatically generates classification information based on the content of the reason statements. For example, the information processing infrastructure 110 generates classification information based on theme information, document type, and user-specified tag information included in the construction data. The information processing infrastructure 110 uses the classification information to organize reason statement data by field and can be used to narrow down the search target.

[0046] The information processing infrastructure 110 uses the reason text (Why text) as text containing expressions representing the reasons obtained by the acquisition unit 221. The information processing infrastructure 110 uses the reason text as supporting information in generating the answer text. For example, in response to the query data "Why should solar power generation be introduced?", the information processing infrastructure 110 refers to the reason texts included in the reason corpus that contain descriptions related to reducing electricity costs, reducing carbon dioxide emissions, and securing power during disasters, and generates answer data for the query data.

[0047] Figure 4 shows an example of a data item table 400 that indicates the data items that constitute the reason data managed by the information processing infrastructure 110 in the reason corpus.

[0048] The information processing infrastructure 110 generates reason text data, which includes identifiers, classification information (categories), and reason texts (Why texts), according to the data item table 300 shown in Figure 3, and treats the reason text data as intermediate data in the reason corpus construction process. In the reason corpus construction process, the information processing infrastructure 110 generates reason vectors corresponding to the reason texts contained in the reason text data, and generates reason data by adding the reason vectors to the reason text data.

[0049] The information processing infrastructure 110 uses a multidimensional numerical sequence generated from the reason statement as a reason vector. The information processing infrastructure 110 uses a single data set, which associates an identifier, classification information (category), reason statement (Why text), and reason vector, as reason data. The information processing infrastructure 110 stores the reason data in the storage device 220 as a component of the reason corpus, uses the reason vector to calculate similarity in the search process, and can refer to the reason statement based on the identifier.

[0050] The information processing infrastructure 110 manages reason data by associating it with at least an identifier, classification information (category), reason statement (Why text), and reason vector. The information processing infrastructure 110 stores the reason data in the storage device 220 according to the data item table 400 shown in Figure 4 and provides it for retrieval processing.

[0051] The information processing infrastructure 110 uses identifiers as information to uniquely identify reason data. The information processing infrastructure 110 uses identifiers as keys to associate reason statements and reason vectors. For example, the information processing infrastructure 110 generates reason vectors for reason statements using an embedding model, then assigns identifiers to them and stores the reason vectors in the reason data in the storage device 220. In the search process, the information processing infrastructure 110 can refer to the identifiers and retrieve the reason statements corresponding to the reason vectors.

[0052] The information processing infrastructure 110 uses classification information (categories) as information used to classify reason data. The information processing infrastructure 110 can use the classification information to assist in narrowing down and managing search targets. For example, the information processing infrastructure 110 automatically generates classification information based on the content of the reason statement and stores the reason data in the storage device 220 with the classification information attached.

[0053] The information processing infrastructure 110 can reference reason statements corresponding to search results by including and storing the reason statements (Why text) in the reason data. If the information processing infrastructure 110 adopts a configuration in which the reason statements (Why text) are not included and stored in the reason data, it can reference reason statements stored in a different storage area than the storage device 220 using an identifier.

[0054] The information processing platform 110 uses the reason vector as a multidimensional numerical sequence obtained by inputting the reason sentence (Why text) into an embedding model. The information processing platform 110 uses the reason vector to calculate the similarity with the query vector obtained from the query data. When the search process is executed, the information processing platform 110 can calculate the cosine similarity by calculating the L2 norm of the reason vector, calculating the L2 norm of the query vector, and dividing the calculated dot product by the product of the calculated L2 norms. Based on the calculated cosine similarity, the information processing platform 110 can obtain a predetermined number of reason data as search results from the reason corpus.

[0055] Figure 5 is a diagram showing an example of a flowchart illustrating the overall procedure of the processing performed by the information processing infrastructure 110.

[0056] In step S501, the information processing infrastructure 110 selects a processing mode. If the information processing infrastructure 110 selects reason corpus construction as the processing mode, it transitions to the reason corpus construction process in step S502. If the information processing infrastructure 110 selects query as the processing mode, it transitions to the query processing in step S503.

[0057] In step S502, the information processing infrastructure 110 executes the reason corpus construction process. For example, if the information processing infrastructure 110 receives a description of a proposal or specification as construction data, the reason corpus construction process acquires text containing expressions that represent the reasons included in the description as reason sentences, generates reason vectors based on the reason sentences, and stores them as a reason corpus. Details of the reason corpus construction process will be described later with reference to Figure 6.

[0058] In step S503, the information processing infrastructure 110 executes query processing. The information processing infrastructure 110 obtains the query reason statement from the query data, obtains the query vector corresponding to the query reason statement, searches for reason data based on the similarity with the reason vector included in the reason corpus, and generates answer data based on the search results. Details of the query processing will be described later with reference to Figure 7.

[0059] The information processing infrastructure 110 terminates processing after completing the processing in step S502 or step S503.

[0060] Figure 6 shows an example of a flowchart for the reason corpus construction process in the information processing system 100.

[0061] In step S601, the information processing infrastructure 110 receives construction data to be used in constructing the reason corpus. The information processing infrastructure 110 may receive construction data from the management terminal 120. The information processing infrastructure 110 receives, for example, explanatory texts, proposals, specifications, meeting minutes, or similar document data as construction data.

[0062] In step S602, the information processing infrastructure 110 obtains a reason statement based on the construction data. The information processing infrastructure 110 obtains text containing expressions that represent the reasons as the reason statement. For example, if the construction data includes "electricity costs can be reduced by introducing solar power generation," "carbon dioxide emissions can be reduced," and "power can be secured in the event of a disaster," the information processing infrastructure 110 can obtain text containing the above expressions as the reason statement.

[0063] More specifically, the information processing infrastructure 110 generates a prompt containing construction data. The information processing infrastructure 110 includes instructions in the prompt to insert text contained in the construction data and extract expressions that represent reasons, etc. The information processing infrastructure 110 can generate a prompt containing, for example, the following instructions: The information processing infrastructure 110 instructs to "extract expressions that represent reasons, etc. from the text" and to "output the extraction results in JSON format with the key "why". For example, the information processing infrastructure 110 generates an instruction statement that says, "Extract expressions that represent reasons, etc. from the following text and return them in JSON format. The format should be { "why" : "extracted reason text"}. Text: <text from construction data>".

[0064] The information processing infrastructure 110 may execute preprocessing logic to exclude elements other than expressions representing reasons from the construction data for the purpose of generating text to be input for extracting expressions representing reasons, etc. The information processing infrastructure 110 may, as preprocessing logic, divide the descriptions contained in the construction data into sentence units and classify each divided sentence into one of the following: a sentence corresponding to What which indicates the content of the subject being explained, a sentence corresponding to How which indicates the procedure, a sentence corresponding to When which indicates the time, a sentence corresponding to Where which indicates the location, or a sentence corresponding to an expression representing a reason, etc. The information processing infrastructure 110 may, as a classification, determine that a sentence containing a time expression or date expression corresponds to When. The information processing infrastructure 110 may, as a classification, determine that a sentence containing an instruction expression or enumeration expression which indicates the procedure corresponds to How. The information processing infrastructure 110 may, as a classification, determine that a sentence containing a place name expression or location expression which indicates the location corresponds to Where. The information processing infrastructure 110 may, as a classification, determine that a sentence containing an expression which describes the attributes or state of the object corresponds to What. The information processing platform 110 may, as preprocessing logic, exclude sentences corresponding to What, How, When, and Where from the extraction target, and leave sentences corresponding to expressions representing reasons, etc., as the extraction target. The information processing platform 110 may, as preprocessing logic, concatenate the sentences that remain after exclusion and generate extraction text to be input to the extraction process for expressions representing reasons, etc.

[0065] The information processing infrastructure 110 inputs a prompt to an external large-scale language model and receives text output from the external large-scale language model. When inputting to the external large-scale language model, the information processing infrastructure 110 sends a request including the prompt via a communication interface. The external large-scale language model may be a language processing model selected by the operator (e.g., a large-scale language model, a question answering model, an information extraction model). The information processing infrastructure 110 may also include an identifier specifying the large-scale language model in the request. If the response text received from the external large-scale language model contains JSON format, the information processing infrastructure 110 extracts the string corresponding to the key "why" and obtains the extracted string as a reason statement. If the response text received from the external large-scale language model cannot be parsed as JSON format, the information processing infrastructure 110 can adopt a configuration that treats a predetermined error text as a reason statement.

[0066] The information processing infrastructure 110 may perform post-processing to remove elements other than expressions representing reasons if the response text received from an external large-scale language model contains elements other than expressions representing reasons. The information processing infrastructure 110 may also adopt a configuration in which, as post-processing, it extracts the string corresponding to the key "why" if the response text contains JSON format, and does not treat strings corresponding to keys other than "why" as reason statements.

[0067] The information processing infrastructure 110 assigns an identifier to the acquired reason statement and supplies it as reason statement data to the subsequent storage process. The information processing infrastructure 110 uses the identifier as a key to associate the reason statement with the reason vector.

[0068] In step S603, the information processing infrastructure 110 stores the reason statements obtained in step S602 as reason statement data according to the data item table 300 shown in Figure 3. The information processing infrastructure 110 stores the reason statements in the storage device 220. The information processing infrastructure 110 assigns an identifier to the reason statements and manages the reason statements by identifier. The information processing infrastructure 110 assigns classification information (category) to the reason statements and organizes the reason statements by classification information.

[0069] In step S604, the information processing infrastructure 110 obtains a reason vector corresponding to the reason statement. The information processing infrastructure 110 inputs the reason statement as input text into the embedding model and obtains the embedding vector output from the embedding model as the reason vector. The information processing infrastructure 110 obtains the reason vector as a multidimensional numerical sequence.

[0070] For example, prior to obtaining the reason vector, the information processing infrastructure 110 reads secret information used for communication authentication from environment variables. The information processing infrastructure 110 sends a request including the reason statement to the embedding generator and obtains the numerical sequence included in the response received from the embedding generator as the reason vector.

[0071] The information processing infrastructure 110 includes a model identifier that specifies the embedding model in the above request. The information processing infrastructure 110 specifies a text embedding model that outputs high-dimensional embedding vectors based on the meaning of the input text as the model identifier. The information processing infrastructure 110 may also use a large-scale text embedding model that outputs embedding vectors that reflect the semantic relationships between words as the text embedding model.

[0072] The information processing infrastructure 110 may specify an identifier representing a text embedding model selected by the operator as a model identifier to be included in the request sent to the embedding generation device. The information processing infrastructure 110 may specify embedding-3-large as an example of a model identifier. The information processing infrastructure 110 may adopt a large-scale text embedding model that outputs a high-dimensional embedding vector based on the meaning of the input text as the text embedding model selected by the operator. If the text embedding model selected by the operator is fixed during the operational period, the information processing infrastructure 110 may continue to specify the same identifier as the model identifier.

[0073] In step S605, the information processing infrastructure 110 stores the reason vector in the storage device 220. For example, the information processing infrastructure 110 stores the reason vector obtained in step S604 as reason data according to the data item table 400 shown in Figure 4. More specifically, the information processing infrastructure 110 generates reason data by associating the identifier, classification information (category), reason statement (Why text) with the reason vector, and stores the reason data in the storage device 220.

[0074] The storage device 220 may store the reason data set in a file-formatted storage area. The storage device 220 may adopt the JSON format as the file format and store the reason data set, which associates identifiers, classification information, reason statements, and reason vectors, as JSON formatted data. The storage device 220 may accumulate the reason vector set in a file-formatted storage area and hold it as a reference for similarity-based search processing. The storage device 220 may also statically accumulate the reason vector set in a format that can be referenced during the operational period.

[0075] In this embodiment, when constructing a reason corpus, the results of acquiring reason statements may be treated as intermediate data used to generate reason vectors. If the information processing infrastructure 110 adopts a configuration that does not store intermediate data, the process of acquiring reason statements and the process of acquiring reason vectors are executed sequentially and stored in the storage device 220 as reason data according to the data item table 400 shown in Figure 4, and Figure 3, which shows the data items of the results of acquiring reason statements, may be omitted.

[0076] The information processing infrastructure 110 may update the reason corpus each time it receives additional construction data from the management terminal 120. The information processing infrastructure 110 may adopt a configuration in which, as an update, it performs the processing from steps S602 to S605 on the additionally received construction data and adds the generated reason data to the group of reason data stored in the storage device 220. The information processing infrastructure 110 may also add the reason vectors generated from reason statements derived from the additionally received construction data to the group of reason vectors stored in the storage device 220.

[0077] Figure 7 shows an example of a flowchart for query processing in the information processing system 100.

[0078] In step S701, the information processing infrastructure 110 receives query data from the client terminal 130. The query data includes the question entered by the user. For example, if the user on the client terminal 130 enters "Why should solar power generation be installed?", the information processing infrastructure 110 receives query data including the above question.

[0079] In step S702, the information processing infrastructure 110 obtains a query reason statement based on the query data. The information processing infrastructure 110 obtains the query reason statement using the acquisition unit 221. The information processing infrastructure 110 obtains text that includes expressions representing the reasons, etc., as the query reason statement. For example, if the query data includes "Why should solar power generation be introduced?", the information processing infrastructure 110 obtains text that includes a description of "reasons for introduction," "purpose of introduction," or "background of introduction" as the query reason statement.

[0080] The information processing infrastructure 110 may generate a prompt containing query data when obtaining the query reason statement and input the prompt into the large-scale language model. The information processing infrastructure 110 may also obtain the text containing expressions that represent the reason, etc., from the text output from the large-scale language model as the query reason statement.

[0081] More specifically, the information processing infrastructure 110 inserts the text contained in the query data into the prompt and includes instructions in the prompt to extract expressions that represent reasons, etc. The information processing infrastructure 110 may also include instructions in the prompt to output the extraction results in a predetermined structured format. The information processing infrastructure 110 may specify a JSON format with the key "why" as the structured format.

[0082] For example, the information processing infrastructure 110 generates an instruction statement that instructs the system to "extract expressions representing reasons, etc., from the following text and output the extraction results in JSON format with the key 'why'," and generates a prompt by concatenating the text included in the query data to the instruction statement. The information processing infrastructure 110 obtains the string corresponding to the key "why" from the JSON text output from the large-scale language model and adopts the obtained string as the query reason statement.

[0083] Furthermore, the query reason statement is not limited to cases where the query data explicitly includes expressions representing reasons, intentions, backgrounds, or objectives. It may also be obtained as separate text containing expressions representing reasons, intentions, backgrounds, or objectives, generated by a large-scale language model based on the semantic content of the query data. For example, if the query data includes "Why should solar power generation be introduced?", the large-scale language model may generate text expressing the gist of the question, such as "Reasons for introduction," "Purpose of introduction," or "Background of introduction," and this generated text may be adopted as the query reason statement.

[0084] In step S703, the information processing infrastructure 110 executes a search process. Based on the query reason sentence obtained in step S702, the information processing infrastructure 110 obtains a query vector and calculates the similarity between the query vector and the reason vectors included in the reason corpus. Based on the similarity, the information processing infrastructure 110 selects a predetermined number of reason data from the reason corpus. Details of the search process will be described later using Figure 8.

[0085] In step S704, the information processing infrastructure 110 acquires reason data. Based on the identifier selected by the search process in step S703, the information processing infrastructure 110 acquires reason data from the storage device 220. The information processing infrastructure 110 refers to the reason statements (Why text) contained in the reason data and acquires the reason statements as reason statement data. The information processing infrastructure 110 may also refer to the classification information (category) contained in the reason data and acquire the reason statement data while associating the classification information with it.

[0086] In step S705, the information processing infrastructure 110 generates response data based on the inquiry data and reason data obtained up to step S704, and retrieves the generated response data. When the information processing infrastructure 110 generates a presentation text as response data, it may include a process to convert the expressions contained in the presentation text into neutral expressions. In the process of converting to neutral expressions, the information processing infrastructure 110 suppresses emotional, aggressive, or assertive expressions and prioritizes expressions that express reasons, etc., when constructing the presentation text.

[0087] The information processing infrastructure 110 may, in the process of converting to neutral expressions, take the words and phrases included in the query data and the reason sentences included in the reason data as input and execute a process to generate paraphrasing candidates. The information processing infrastructure 110 may employ a process of providing a natural language generation model with an instruction sentence that specifies the type of expression to be suppressed and the output style as the process of generating paraphrasing candidates. The information processing infrastructure 110 includes requirements in the instruction sentence for reducing aggressive expressions, mitigating assertive expressions, avoiding asserting the other party's intent, and clarifying reasons.

[0088] In the process of converting to a neutral expression, the information processing infrastructure 110 quotes or summarizes the reason statement included in the reason data as supporting information for the presentation statement. In the presentation statement, the information processing infrastructure 110 reconstructs the content of the reason statement from the perspectives of purpose, background, and constraints, and separates the assertion part and the reason part of the query data. In the presentation statement, in order to avoid definitive statements, the information processing infrastructure 110 adds conditions and assumed scope that serve as the basis for the recommendation.

[0089] For example, if the inquiry data includes target text related to interpersonal communication and the target text contains aggressive expressions, the information processing infrastructure 110 retains the assertive part of the target text while extracting the part corresponding to the reasons, etc., of the target text, and refers to the extraction result as reason data. The information processing infrastructure 110 reflects the reasons, etc., of the target text in the presentation text, generates a presentation text that has been converted into an expression that suppresses the aggressive expressions contained in the target text, and obtains the generated presentation text as response data.

[0090] For example, if the target text contains a definitive statement such as "The response is too slow, so operations have failed," the information processing infrastructure 110 extracts "delayed response time" and "difference from expected operational level" as elements corresponding to the reasons contained in the target text, and generates a statement consistent with the extraction results, such as "We would like to confirm that the delay in response time continues, resulting in a difference from the operational level expected by the user." In the statement, the information processing infrastructure 110 avoids making definitive statements and focuses on describing points to be confirmed and requests for improvement.

[0091] In step S706, the information processing infrastructure 110 outputs the response data. The information processing infrastructure 110 sends the response data to the client terminal 130. The client terminal 130 displays the response data received from the information processing infrastructure 110.

[0092] Figure 8 shows an example of a flowchart for the search process. The information processing infrastructure 110 executes the process shown in Figure 8 as the search process (step S703) in the query process shown in Figure 7.

[0093] In step S801, the information processing infrastructure 110 obtains the query vector. The information processing infrastructure 110 inputs the query reason statement as input text into the embedding model and obtains the embedding vector output from the embedding model as the query vector. The information processing infrastructure 110 obtains the query vector as a multidimensional numerical sequence.

[0094] For example, prior to obtaining the query vector, the information processing infrastructure 110 reads secret information used for communication authentication from environment variables. The information processing infrastructure 110 sends a request including the query reason statement to the embedding generator and obtains the numerical sequence included in the response received from the embedding generator as the query vector.

[0095] The information processing infrastructure 110 includes a model identifier that specifies the embedding model in the above requirements. The information processing infrastructure 110 specifies a text embedding model that outputs high-dimensional embedding vectors based on the meaning of the input text as the model identifier. The information processing infrastructure 110 may also use a large-scale text embedding model that outputs embedding vectors that reflect the semantic relationships between vocabulary words as the text embedding model. It is preferable to adopt a configuration in which the reason vector and the query vector are generated by the same embedding model. By using the same embedding model, the information processing infrastructure 110 can, for example, place the reason vector and the query vector on the same vector space, thereby improving the accuracy of similarity calculation.

[0096] In step S802, the information processing infrastructure 110 acquires a group of reason vectors. The information processing infrastructure 110 reads the group of reason vectors from the storage device 220 that stores the reason corpus. The information processing infrastructure 110 acquires multiple reason vectors as a group of reason vectors, which are included in the reason data managed according to the data item table 400 shown in Figure 4. The information processing infrastructure 110 acquires the group of reason vectors as the target for calculating similarity with the query vector.

[0097] In step S803, the information processing infrastructure 110 calculates the L2 norm for the query vector obtained in step S801. The information processing infrastructure 110 squares the value of each element of the query vector and calculates the sum of the squared values. The information processing infrastructure 110 calculates the square root of the sum of the squared values ​​and sets the calculation result as the L2 norm of the query vector. The information processing infrastructure 110 uses the calculated L2 norm in the subsequent normalization and similarity calculation processes.

[0098] In step S804, the information processing infrastructure 110 calculates the L2 norm for each reason vector included in the reason vector group. For each reason vector extracted by the information processing infrastructure 110 as a comparison target, the information processing infrastructure 110 calculates the L2 norm of the reason vector by squaring the value of each element of the reason vector, calculating the sum of the squared values, and then calculating the square root of the sum of the squared values. The information processing infrastructure 110 uses the calculated L2 norm in the subsequent normalization and similarity calculation processes.

[0099] In step S805, the information processing infrastructure 110 calculates the cosine similarity between the query vector and the reason vector. The information processing infrastructure 110 selects one reason vector from the group of reason vectors included in the reason corpus and calculates the dot product of the selected reason vector and the query vector. In calculating the dot product, the information processing infrastructure 110 adds the product of each element of the query vector and each element of the reason vector for the number of dimensions and calculates the result of the addition as the dot product. The information processing infrastructure 110 calculates the cosine similarity between the query vector and the reason vector by dividing the calculated dot product by the product of the L2 norm of the query vector calculated in step S803 and the L2 norm of the reason vector calculated in step S804. The information processing infrastructure 110 repeats the above calculation for each reason vector included in the group of reason vectors and calculates the cosine similarity for each reason vector.

[0100] In step S806, the information processing infrastructure 110 performs ranking based on similarity. The information processing infrastructure 110 stores the cosine similarity calculated for each reason vector in step S805, associating it with an identifier. After the cosine similarity calculated for each reason vector included in the reason vector group is collected, the information processing infrastructure 110 sorts the search result candidates, including the identifier and cosine similarity, in descending order of cosine similarity. Based on the sorting result, the information processing infrastructure 110 selects the top-ranking search result candidates. The information processing infrastructure 110 obtains a predetermined number (one or more) identifiers as the top-ranking search result candidates.

[0101] In this embodiment, the information processing system 100 may normalize the reason vectors generated in the reason corpus construction process before storing them in the storage device 220. The information processing system 100 may calculate the L2 norm for the reason vectors obtained by the reason vectorization unit 222 and generate normalized reason vectors by dividing each element of the reason vectors by the L2 norm, thereby normalizing the reason vectors as unit vectors. The information processing system 100 may store the normalized reason vectors in the storage device 220 as components of the reason corpus using the storage unit 223. The information processing system 100 may normalize the query vector in the search process and calculate the dot product with the normalized reason vectors to perform similarity calculation and ranking based on a scale equivalent to cosine similarity. By storing the normalized reason vectors in the reason corpus, the information processing system 100 can omit the normalization process of the reason vector group when executing the search process, reduce the amount of computation, and simplify the processing procedure.

[0102] Figure 9 shows a specific example of the data handled by the information processing infrastructure 110 when constructing a reason corpus. In the example shown in Figure 9, the information processing infrastructure 110 extracts reason sentences from the target text 901 input as construction data, generates reason sentence data 902 containing the extracted reason sentences, and generates reason data 903 containing reason vectors based on the reason sentence data 902.

[0103] The information processing infrastructure 110 receives text included in the construction data as target text 901. The information processing infrastructure 110 receives, for example, an explanatory text about solar power generation, which is an example of renewable energy technology, as target text 901. If the target text 901 contains an explanation that represents the reason for introduction, the purpose of introduction, or the background of introduction, the information processing infrastructure 110 will treat the reason expression contained in the target text 901 as the target for extraction in the subsequent reason statement acquisition process.

[0104] The information processing infrastructure 110 generates reason statement data 902 based on the target text 901. The information processing infrastructure 110 generates data as reason statement data 902 that includes at least an identifier, classification information (category), and a reason statement (Why text). The information processing infrastructure 110 assigns a string that uniquely identifies the reason statement data 902 as an identifier. The information processing infrastructure 110 assigns information used for classifying the reason statement data 902 as classification information (category). In the example shown in Figure 9, the information processing infrastructure 110 assigns a string indicating automatically generated classification information as classification information (category).

[0105] The information processing infrastructure 110 obtains text containing expressions that represent the reasons, etc., included in the target text 901 as reason statements (Why text). The information processing infrastructure 110 obtains explanatory text as reason statements (Why text) that includes perspectives such as reducing electricity costs, reducing carbon dioxide emissions, addressing fuel supply risks, and using it as an emergency power source during natural disasters. The information processing infrastructure 110 outputs the reason statements (Why text) as JSON format data with a key "why", for example, and stores it in the reason statement data 902.

[0106] The information processing infrastructure 110 generates reason data 903 based on reason statement data 902. The information processing infrastructure 110 generates reason data 903 that includes at least classification information (category), an identifier, a reason vector, and a reason statement (Why text). The information processing infrastructure 110 stores the reason data 903 in the storage device 220 as vector information used for searching in the reason corpus. The information processing infrastructure 110 may store the reason statement (Why text) in the reason data 903 as JSON format data having, for example, the key "why_text".

[0107] The information processing infrastructure 110 uses a multidimensional numerical sequence as the reason vector, which is generated by inputting the reason statement (Why text) as input text into the embedding model and obtaining the embedding vector output from the embedding model. In the example shown in Figure 9, the information processing infrastructure 110 stores the reason vector in the reason data 903 as an array consisting of multiple real numbers. The information processing infrastructure 110 normalizes the reason vector when executing the search process and uses it to calculate the cosine similarity with the query vector.

[0108] The information processing infrastructure 110 manages the association between reason vectors and reason statements (Why text) using identifiers contained in the reason data 903. In the search process, the information processing infrastructure 110 selects the top identifiers by calculating similarity based on the reason vectors, and obtains reason data to be used to generate response data for query data by referring to the reason statements (Why text) corresponding to the selected identifiers.

[0109] Figure 10 shows a specific example of the data handled by the information processing infrastructure 110 when it performs search processing and response data generation processing on query data. In the example shown in Figure 10, the information processing infrastructure 110 receives a query statement 1001, and as a result of the search processing, obtains reason data 1002 which includes an identifier, classification information (category), reason vector, and reason statement (Why text), obtains the reason statement (Why text) contained in the reason data 1002 as reason statement 1003, and generates response data 1004 based on reason statement 1003.

[0110] The information processing infrastructure 110 receives the target text included in the inquiry data entered by the user as an inquiry statement 1001. In the example shown in Figure 10, the information processing infrastructure 110 receives an inquiry statement 1001 that asks for the reasons for introducing solar power generation.

[0111] The information processing infrastructure 110 executes a search process based on the query statement 1001. The information processing infrastructure 110 obtains the query reason statement from the query statement 1001, inputs the query reason statement as input text into the embedding model, and obtains the embedding vector output from the embedding model as the query vector. The information processing infrastructure 110 calculates the similarity between the query vector and the reason vectors included in the reason corpus, and selects identifiers with high similarity. The information processing infrastructure 110 obtains the record corresponding to the selected identifier as reason data 1002.

[0112] The information processing infrastructure 110 handles the reason data 1002 as a data structure that conforms to the data item table 400 shown in Figure 4. The information processing infrastructure 110 includes an item indicating classification information (category), an item indicating an identifier, an item indicating a reason vector, and an item indicating a reason statement (Why text) for the reason data 1002. When the information processing infrastructure 110 represents the reason statement (Why text) as an external representation in JSON format, it may use the key "why_text" as the key to store the value corresponding to the reason statement (Why text). In the example shown in Figure 10, the reason data 1002 is represented in JSON format having the keys "category", "id", "vector", and "why_text".

[0113] The information processing infrastructure 110 refers to the reason statement (Why text) contained in the reason data 1002 and retrieves the reason statement (Why text) as reason statement 1003. In the example shown in Figure 10, reason statement 1003 includes an explanatory text that includes perspectives such as reducing electricity costs, suppressing carbon dioxide emissions, addressing fuel supply risks, and securing power during natural disasters and power outages as reasons for introducing solar power generation.

[0114] The information processing infrastructure 110 generates response data 1004 based on the query statement 1001 and the reason statement 1003. In generating response data 1004, the information processing infrastructure 110 may employ a process in which the query statement 1001 and the reason statement 1003 are given as input to a natural language processing model, and an explanatory text consistent with the reason statement 1003 is obtained as output. In the example shown in Figure 10, the information processing infrastructure 110 generates response data 1004 that includes an explanatory text stating that the introduction of solar power generation may be related to the reduction of greenhouse gas emissions and the stable securing of energy.

[0115] According to this embodiment, even if the words included in the query data do not match the words included in the construction data, a reason statement can be searched based on semantic similarity, and response data with justification regarding the reason, intention, background, or purpose can be output for the query data.

[0116] (II) Addendum An information processing system is a term that describes a conceptual framework that includes computing resources for performing arithmetic operations, memory resources, and communication resources. An information processing system can be configured as a single computer, as a distributed system using multiple computer devices, or as a virtualized computing environment. An information processing system can also be configured as a system including server devices and terminal devices.

[0117] Construction data is a term that describes the data concept of input for generating a reason corpus, which may include expressions that represent reasons, intentions, background, or purposes. Construction data may include text equivalent to documents, meeting minutes, explanatory texts, proposals, specifications, posts, or comments. Information processing systems may receive construction data from terminal devices, read it from storage devices, and retrieve it from external storage areas.

[0118] Query data is a term that describes the data concept of text, including question or request statements entered by a user, which is input for search and answer generation processes. Query data may include natural language sentences, text consisting of multiple sentences, and target text as supplementary information. Information processing systems may receive query data from terminal devices, application programs, and communication networks.

[0119] Response data is a term that represents the data concept of information output in response to inquiry data. Response data may include explanatory text, summary text, recommendation text, confirmation text, or presentation text, may include text consisting of multiple sentences, and may include expressions that quote or summarize supporting reasoning. The information processing system may transmit the response data to a terminal device, store it in a storage device, and convert it into a format for screen display.

[0120] A reason statement is a term that describes text obtained from the text contained in the construction data, specifically the portion containing expressions that represent reasons, intentions, background, or purposes. A reason statement may include text extracted from the construction data containing reason expressions, text obtained by summarizing the construction data and converting it into a form that includes reason expressions, and text that includes purpose or background expressions derived from the construction data. An information processing system may generate a reason statement in a form expressed using different vocabulary than that used in the construction data. Even if the construction data does not explicitly contain reason expressions, a reason statement may be obtained as text containing expressions that represent reasons, intentions, background, or purposes, generated by a large-scale language model based on the semantic content of the construction data.

[0121] The query reason statement is a term that describes text extracted from the text contained in query data, specifically the portion containing expressions that represent the reason, intention, background, or purpose. The query reason statement may include text extracted from the query data containing reason expressions, text converted from the query data into a summarized form containing reason expressions, and text containing purpose or background expressions derived from the query data. An information processing system may generate the query reason statement in a form expressed using different vocabulary than that used in the query data. Even if the query data does not explicitly contain reason expressions, the query reason statement can be obtained as text containing expressions representing the reason, intention, background, or purpose, generated by a large-scale language model based on the semantic content of the query data.

[0122] The term "acquisition unit" represents the concept of a processing function that obtains predetermined information from input data. The term "first acquisition unit" represents the concept of a processing function that obtains reason statements from construction data. The term "second acquisition unit" represents the concept of a processing function that obtains query reason statements from query data. The information processing system can implement the first and second acquisition units through rule-based extraction processing, through statistical language processing-based extraction processing, and through large-scale language model-based extraction processing.

[0123] A prompt is a term that represents a combination of an instruction and input text that is input to a language processing model. A large-scale language model is a term that represents the concept of a language processing model that outputs text based on input text. An information processing system may insert construct data into a prompt, may insert query data into a prompt, and may include instructions in a prompt that include expressions representing reasons, intentions, backgrounds, or purposes as the target of extraction. An information processing system may parse the text output from a large-scale language model and adopt it as a reason statement or query reason statement.

[0124] An embedding model is a term that describes the concept of a model that outputs a multidimensional sequence of numbers that reflects the meaning of the input text. A multidimensional embedding vector is a term that describes the concept of a multidimensional sequence of numbers generated by an embedding model. A reason vector is a term that describes a vector obtained by converting a reason statement into a multidimensional embedding vector using an embedding model. A query vector is a term that describes a vector obtained by converting a query reason statement into a multidimensional embedding vector using an embedding model. An information processing system may use a neural network-based text representation model as the embedding model, and may use a model that outputs a distributed representation that reflects the semantic relationships between words in the vocabulary.

[0125] Reason data is a term that represents the concept of a data structure that associates reason statements with reason vectors. Reason corpus is a term that represents the concept of a data set that holds multiple reason data for searching. Identifier is a term that represents the concept of information for uniquely identifying reason data. Classification information is a term that represents the concept of information for classifying reason data. An information processing system may assign strings or numerical sequences as identifiers, and may assign strings equivalent to field names, topic names, or tags as classification information. An information processing system may assign classification information by rules, by statistical classification processing, or by generating it using a language processing model.

[0126] The search unit is a term that represents the concept of a processing function that selects reason data from a reason corpus based on the relationship between the query vector and the reason vector. Similarity is a term that represents the concept of a measure of the closeness between the query vector and the reason vector. Geometric closeness is a term that represents a concept that includes closeness based on distance, angle, or dot product between vectors in a multidimensional space. The information processing system may use cosine similarity, a measure based on Euclidean distance, a measure based on dot product, or a measure based on normalized dot product as the similarity measure. The predetermined number is a term that represents the concept of the number of reason data obtained as search results. The information processing system may set the predetermined number as a fixed value, set it variably according to the length of the query data, or set it variably according to the classification information.

[0127] Cosine similarity is a term that describes the concept of a measure of the angular closeness between a query vector and a reason vector. Dot product is a term that describes the concept of an operation that calculates the sum of the products of each element of a vector. L2 norm is a term that describes the concept of a magnitude calculated as the square root of the sum of the squares of each element of a vector. An information processing system may perform normalization using the L2 norm of the query vector and the L2 norm of the reason vector, and calculate cosine similarity using the normalized dot product. To reduce numerical calculation errors, an information processing system may use operations based on floating-point arithmetic or operations based on fixed-point arithmetic.

[0128] The output unit is a term that represents the concept of a processing function that generates response data based on reason data obtained as search results and outputs the response data externally. In generating response data, an information processing system may employ a method in which query data and reason statements are given as input to a language generation model, and text containing descriptions of reasons, intentions, backgrounds, or purposes is obtained as output. In generating response data, an information processing system may employ a method of quoting reason statements, a method of summarizing reason statements, or a method of integrating multiple reason statements.

[0129] The term "Target Text" refers to text that may be included in the query data, corresponding to user-handled text, posts, or comments. The term "Presentation Text" refers to text that may be generated as a result of transforming the expression of the Target Text. "Emotional Expression" refers to the concept of an expression that contains words or phrases that evoke strong emotions. "Aggressive Expression" refers to the concept of an expression that contains words or phrases that may blame the other party. "Assertive Expression" refers to the concept of an expression that contains words or phrases that establish a conclusion without indicating assumptions or conditions. "Neutral Expression" refers to the concept of an expression in which the intensity of emotional, aggressive, or assertive expressions has been mitigated. In generating the presentation text, the information processing system may incorporate descriptions of reasons, intentions, background, or purposes contained in the reason statement as a basis, may add recommended conditions or assumptions, and may express them in the form of confirmations.

[0130] The identifiers of components described herein (e.g., prefixes and symbols such as "First" and "Second") are for convenience only and do not limit the number, order, function, arrangement, etc., of the components. The same identifier may refer to different components in different embodiments, and one component may also perform the function of another component. Therefore, the identifiers of components described herein are not intended to limit the technical scope, functional scope, or scope of rights of the components, and each component should be interpreted flexibly according to the context of its embodiment.

[0131] In this specification, "interface device" means a component that may include one or more interface devices. Such interface devices may include, but are not limited to, I / O (Input / Output) interface devices, communication interface devices, or combinations thereof. For example, an I / O interface device may be configured to function as a user interface and may include at least one input device (e.g., a keyboard, a pointing device) and / or an output device (e.g., a display). These I / O interface devices may be configured to have communication functions that allow connection to remote computing devices, in which case the I / O interface device can also operate as a communication interface device. Furthermore, the communication interface device may include identical communication means (e.g., multiple NICs (Network Interface Cards)) or a combination of different types of communication means (e.g., a NIC and an HBA (Host Bus Adapter)). This enables a flexible configuration that ensures connectivity with heterogeneous systems. Interface devices configured in this way are not limited to a specific hardware configuration and can accommodate future technological advancements and diversification of embodiments.

[0132] In this specification, “storage device” means a component that may include at least one storage device. Depending on the intended use and system configuration, such storage devices may be classified, for example, into “memory” which temporarily holds data during operation and “persistent storage device” which retains data even after power is lost. Memory can function as a temporary storage medium accessible by the processor and may include volatile, non-volatile, or a combination thereof memory devices. Specifically, examples include, but are not limited to, volatile memory such as DRAM (Dynamic Random Access Memory) and SRAM (Static RAM), and non-volatile memory such as MRAM (Magnetoresistive RAM) and ReRAM (Resistive RAM). Persistent storage devices are components intended for long-term data storage and include devices using non-volatile storage media. Specifically, these may include HDD (Hard Disk Drive), SSD (Solid State Drive), NVMe (Non-Volatile Memory Express) drives, etc. Next-generation storage technologies such as phase-change memory may also be included as examples of storage device configurations. A storage device configured in this way is not limited by the type or architecture of the storage medium, and can accommodate future technological advancements and diversification of implementations.

[0133] In this specification, "processor" means a component that may include an arithmetic unit or circuit capable of performing at least one processing function. Depending on the application, such a processor may include, for example, a microprocessor device such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), or it may include implementations using dedicated circuits such as an FPGA (Field-Programmable Gate Array), a CPLD (Complex Programmable Logic Device), or an ASIC (Application Specific Integrated Circuit). A processor may consist of a single core, a multi-core, or a single processor core. A processor may be configured to implement processing functions using a computer program, directly implemented using hardware circuits, or a hybrid configuration combining these. When processing is performed by a program, the processor may perform processing in cooperation with other components such as memory devices or interface devices. In this specification, a particular function may be described as a "part," but such a function can be realized by a program executed by the processor, an implementation using circuits, or a combination thereof. Therefore, such a function can be considered to be at least a part of the processor. The program may be supplied from an external program source. Examples of program sources include, but are not limited to, network-connected program distribution servers and computer-readable non-temporary storage media. Processors configured in this way are not limited to specific hardware configurations or implementation forms, and can adapt to future technological advancements and diversification of implementations.

[0134] In this specification, "system" means a set of components that may include at least one computing resource. Such a system may consist of one or more physical computers (dedicated hardware, on-premises servers, etc.) or virtualized computing resources (cloud infrastructure, virtual machines, containers, etc.). A system may include, but is not limited to, cloud computing systems, cluster configurations, serverless environments, etc. A system may be configured within a single device or in a configuration in which multiple computing resources cooperate via a network. Each component may be logically integrated or physically separated. Such a system may also include components such as processors, storage devices, and interface devices, and these components may be implemented as physical devices or realized as virtual configurations. A system configured in this way is not limited to a specific hardware configuration, implementation form, deployment form, etc., and can accommodate future technological advancements and diversification of embodiments.

[0135] The embodiments described above have, for example, the following features.

[0136] (1) An information processing system (e.g., information processing system 100) that outputs response data to user inquiry data (e.g., response data generated by output unit 226 and sent to client terminal 130), comprising: a first acquisition unit (e.g., acquisition unit 221, information processing infrastructure 110) that acquires content including expressions representing reason, intention, background or purpose from construction data (e.g., construction data sent from management terminal 120 to information processing infrastructure 110) as reason statements (e.g., reason statements acquired by acquisition unit 221 by extraction, generation, etc. from construction data), and the first acquisition unit A reason vectorization unit (for example, reason vectorization unit 222, information processing infrastructure 110) obtains a reason vector (for example, an embedding vector corresponding to the reason statement obtained by the reason vectorization unit 222) by converting the reason statement obtained by the first acquisition unit into a multidimensional embedding vector (for example, a multidimensional sequence of numbers output from the embedding model) using an embedding model (for example, an embedding model that takes the reason statement as input and outputs a multidimensional sequence of numbers) and associates the reason statement obtained by the first acquisition unit with the reason vector obtained by the reason vectorization unit A storage unit (e.g., storage unit 223, information processing infrastructure 110) generates data (e.g., reason data including reason statements and reason vectors associated with identifiers and classification information), stores the generated reason data, and constructs a reason corpus (e.g., a collection of multiple reason data stored in the storage device 220), and an acquisition unit (e.g., acquisition unit 221 extracts, generates, etc., from user query data (e.g., query data sent from client terminal 130 to information processing infrastructure 110) and retrieves content including expressions representing reasons, intentions, backgrounds, or purposes into query reason statements (e.g., acquisition unit 221 extracts, generates, etc., from query data). A second acquisition unit (e.g., acquisition unit 221, information processing infrastructure 110) acquires the query reason statement (which is advantageous), and a query vectorization unit (e.g., query vectorization unit 224, information processing infrastructure 110) converts the query reason statement acquired by the second acquisition unit into a multidimensional embedding vector using the embedding model to obtain a query vector (e.g., an embedding vector corresponding to the query reason statement acquired by the query vectorization unit 224), and the similarity between the query vector converted by the query vectorization unit and the reason vectors included in the reason corpus (e.g.,The system comprises a search unit (e.g., search unit 225, information processing infrastructure 110) that calculates the cosine similarity between the query vector and the reason vector, and searches for a predetermined number of reason data (e.g., the number of top search results) from the reason corpus based on the calculated similarity, and an output unit (e.g., output unit 226, information processing infrastructure 110) that outputs response data to the query data based on the reason sentences associated with the reason data retrieved by the search unit.

[0137] According to the above configuration, for example, even if the words included in the query data do not match the words included in the construction data, the accuracy of the process of extracting supporting information based on semantic similarity regarding reasons, intentions, background, or objectives and generating response data that aligns with the user's motivations or values ​​can be improved.

[0138] (2) The first acquisition unit generates a prompt (for example, an instruction statement including instructions to insert sentences contained in the construction data and extract expressions representing reasons, intentions, backgrounds, or purposes) that includes the construction data (for example, construction data transmitted from the management terminal 120 to the information processing infrastructure 110), inputs the prompt into a large-scale language model (for example, a language processing model that outputs text based on input text), and obtains text containing expressions representing reasons, intentions, backgrounds, or purposes from the large-scale language model as a reason statement (for example, a reason statement obtained as a string corresponding to the key "why"). The second acquisition unit generates a prompt (for example, an instruction statement including instructions to insert sentences contained in the query data and extract expressions representing reasons, intentions, backgrounds, or purposes) that includes the query data (for example, query data transmitted from the client terminal 130 to the information processing infrastructure 110), inputs the prompt into a large-scale language model, and obtains text containing expressions representing reasons, intentions, backgrounds, or purposes from the large-scale language model as a query reason statement (for example, a query reason statement obtained as a string corresponding to the key "why").

[0139] In the above configuration, a prompt containing query data is generated and input into a large-scale language model, and text containing expressions representing reasons, intentions, backgrounds, or purposes is obtained from the large-scale language model as a query reason statement. For example, even if the expressions of reasons, intentions, backgrounds, or purposes included in the construction data and query data are diverse, the reason statement and query reason statement can be reliably obtained through context-aware extraction processing, thereby improving the quality of input information for search processing related to reasons, etc.

[0140] (3) The search unit calculates a similarity (for example, a measure representing the closeness between a query vector and a reason vector) based on the geometric closeness (for example, closeness based on the distance, angle, or dot product between vectors in a multidimensional space) between the query vector converted by the query vectorization unit (for example, a query vector obtained by converting a query reason statement into a multidimensional embedding vector using an embedding model) and multiple reason vectors contained in the reason corpus (for example, a reason corpus stored in the memory device 220) (for example, reason vectors obtained by converting reason statements into multidimensional embedding vectors using an embedding model), and retrieves reason data (for example, reason data associating reason statements with reason vectors) in descending order of the calculated similarity.

[0141] In the above configuration, for example, the information processing system calculates similarity based on the geometric proximity between the query vector and multiple reason vectors included in the reason corpus, and retrieves reason data in order of the calculated similarity. According to the above configuration, for example, even if the words included in the query data do not match the words included in the construction data, it is possible to search for reason sentences with similar intent based on their semantic proximity, thereby improving the accuracy of retrieving reason sentences that serve as the basis for the answer data to the query data.

[0142] (4) The search unit calculates the dot product (for example, the dot product calculated by adding the products of each element of the query vector and each element of the reason vector equal to the number of dimensions) of the query vector (for example, the query vector obtained by converting the query reason statement into a multidimensional embedding vector using the embedding model) and the reason vectors contained in the reason corpus (for example, the reason corpus stored in the memory device 220) and calculates the L2 norm of the query vector (for example, the query vector The L2 norm of the reason vector (for example, the magnitude calculated as the square root of the sum of the squares of each element of the reason vector) is calculated, and the cosine similarity (for example, a measure representing the closeness of the angle between the query vector and the reason vector) is calculated by dividing the calculated dot product by the product of the calculated L2 norms, and a predetermined number of reason data (for example, reason data associating reason sentences with reason vectors) are retrieved from the reason corpus based on the calculated cosine similarity (for example, the number of results obtained as top search results).

[0143] In the above configuration, the information processing system calculates the dot product of the query vector and the reason vector, calculates the L2 norm of the query vector and the L2 norm of the reason vector, and calculates the cosine similarity by dividing the calculated dot product by the product of the calculated L2 norms. With the above configuration, for example, it is possible to suppress the variation in similarity caused by differences in the magnitude of the vectors, to prioritize the search for reason statements whose reason vectors are in similar directions, and to improve the accuracy of selecting reason data that serves as the basis for the answer data to the query data.

[0144] (5) The output unit generates response data (for example, response data generated by providing the query data and the reason text as input to a language processing model) based on the reason data obtained by the search unit (for example, reason data that associates reason texts with reason vectors), the reason texts associated with the reason texts (for example, reason texts included in the reason data (Why text)), and the query data (for example, query data received from the client terminal 130), and outputs the generated response data. This response data includes descriptions of the reasons, intentions, backgrounds, or objectives included in the reason texts (for example, descriptions of reducing electricity costs, reducing carbon dioxide emissions, addressing fuel supply risks, and securing power during natural disasters).

[0145] In the above configuration, response data is generated and output that includes descriptions of the reasons, intentions, background, or objectives contained in the reason statement. For example, compared to the presentation of search results which tend to be biased towards presenting objective information, it is possible to present response data that includes explanations consistent with the motivation or value judgments sought by the user, thereby improving the suitability of the response to the inquiry data.

[0146] (6) The output unit generates a presentation sentence (for example, a presentation sentence generated by converting the target text (for example, a target text corresponding to a post or comment that may be included in the query data) related to the query data (for example, query data received from the client terminal 130) into a neutral expression (for example, an expression in which the intensity of emotional, aggressive, or assertive expressions is mitigated) that suppresses emotional expressions (for example, expressions containing words that evoke strong emotions), aggressive expressions (for example, expressions containing words that can blame the other party), or definitive expressions (for example, expressions containing words that determine a conclusion without showing estimations or conditions), and outputs the generated presentation sentence as the response data (for example, response data sent to the client terminal 130).

[0147] In the above configuration, the information processing system generates a presentation statement based on multiple reason data obtained by the search unit. With this configuration, for example, compared to relying on a single reason statement, it is possible to generate a presentation statement that reflects the background, purpose, or intent of the target text related to the query data from multiple perspectives, thereby improving the comprehensiveness and persuasiveness of the content of the presentation statement.

[0148] In the above configuration, the information processing system generates a presentation text that converts the expression of the target text into a neutral expression that suppresses emotional, aggressive, or assertive expressions. According to the above configuration, for example, it is possible to mitigate conflicts of expression in the dialogue space and support smooth communication while preserving the points the user intends to make. [Explanation of Symbols]

[0149] 100... Information processing system, 110... Information processing infrastructure, 120... Management terminal, 130... Client terminal.

Claims

1. An information processing system that outputs response data to user inquiry data, A first acquisition unit obtains content from construction data that includes expressions representing reasons, intentions, backgrounds, or purposes as reason statements, A reason vectorization unit obtains a reason vector by converting the reason statement obtained by the first acquisition unit into a multidimensional embedding vector using an embedding model, A storage unit generates reason data by associating the reason sentences acquired by the first acquisition unit with the reason vectors obtained by the reason vectorization unit, stores the generated reason data, and constructs a reason corpus. A second acquisition unit obtains content from user inquiry data that includes expressions representing the reason, intention, background, or purpose as an inquiry reason statement, A query vectorization unit obtains a query vector by converting the query reason statement obtained by the second acquisition unit into a multidimensional embedding vector using the embedding model, A search unit calculates the similarity between the query vector converted by the query vectorization unit and the reason vectors included in the reason corpus, and searches for a predetermined number of reason data from the reason corpus based on the calculated similarity. An output unit outputs response data to the query data based on the reason sentence associated with the reason data retrieved by the search unit, An information processing system equipped with the following features.

2. The information processing system according to claim 1, The first acquisition unit generates a prompt including the construction data, inputs the prompt to a large-scale language model, and acquires text from the large-scale language model as a reason statement, which includes expressions representing reason, intention, background, or purpose. The second acquisition unit generates a prompt including the query data, inputs the prompt into a large-scale language model, and acquires text from the large-scale language model as a query reason statement, which includes expressions representing the reason, intention, background, or purpose.

3. The information processing system according to claim 1, The search unit is an information processing system that calculates similarity based on the geometric proximity between the query vector converted by the query vectorization unit and a plurality of reason vectors included in the reason corpus, and retrieves reason data in order of the calculated similarity.

4. The information processing system according to claim 3, The search unit calculates the inner product of the query vector and the reason vectors included in the reason corpus, calculates the L2 norm of the query vector and the L2 norm of the reason vector, calculates the cosine similarity by dividing the calculated inner product by the product of the calculated L2 norms, and searches for a predetermined number of reason data from the reason corpus based on the calculated cosine similarity.

5. The information processing system according to claim 1, The output unit is an information processing system that generates response data including descriptions of the reasons, intentions, backgrounds, or purposes contained in the reason sentences based on the reason sentences associated with the reason data obtained by the search unit and the query data, and outputs the generated response data.

6. The information processing system according to claim 1, The output unit generates a presentation sentence based on a plurality of reason data obtained by the search unit, converting the expression of the target text relating to the inquiry data into a neutral expression that suppresses emotional, aggressive, or assertive expressions, and outputs the generated presentation sentence as the response data.

7. An information processing method that outputs response data to user inquiry data, The first acquisition unit acquires content from the construction data that includes expressions representing reasons, intentions, backgrounds, or purposes as reason statements, The reason vectorization unit converts the reason statement obtained by the first acquisition unit into a multidimensional embedding vector using an embedding model to obtain a reason vector, The storage unit generates reason data by associating the reason sentences acquired by the first acquisition unit with the reason vectors obtained by the reason vectorization unit, and stores the generated reason data to construct a reason corpus. The second acquisition unit acquires content from the user's inquiry data that includes expressions representing the reason, intention, background, or purpose as an inquiry reason statement. The query vectorization unit converts the query reason statement obtained by the second acquisition unit into a multidimensional embedding vector using the embedding model to obtain a query vector, The search unit calculates the similarity between the query vector converted by the query vectorization unit and the reason vectors included in the reason corpus, and searches for a predetermined number of reason data from the reason corpus based on the calculated similarity. An information processing method comprising: an output unit outputting response data to query data based on a reason statement associated with reason data retrieved by the search unit.

Images