Information processing methods, information processing systems, programs, artificial intelligence models, and data structures

Abstract

The objective is to provide an established information processing method for generating hypotheses or insights from large-scale text collections, going beyond mere information extraction and summarization. [Solution] A computer uses artificial intelligence with text generation capabilities to generate hypotheses or insights from a data set containing linguistic information. This method includes (a) a process of analyzing the data set and generating structured data that identifies conflicts or contradictions among multiple claims present therein, and (b) a process of using the structured data as input and causing the artificial intelligence to generate at least one candidate hypothesis, etc., that can explain, resolve, or reconcile the conflicts or contradictions. This intentionally utilizes the contradictory structures, etc., present in the text set as a source of creativity and supports the generation of hypotheses or insights, etc., that could not be easily predicted from a simple sum of information.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing and knowledge discovery using artificial intelligence. More specifically, it relates to an information processing method, an information processing system, a program, an artificial intelligence model, and a data structure for generating hypotheses, findings, etc. that cannot be predicted from the simple sum of information from a large-scale text group using artificial intelligence having text generation ability.

Background Art

[0002] In recent years, the capabilities of artificial intelligence having text generation ability, represented by large language models (LLMs), have made a leap forward, and various technologies for the purpose of knowledge discovery from text groups have been developed.

[0003] As a first technology, there is a technology for extracting entities and relationships from a text group and constructing a knowledge graph, represented by Watson Discovery of IBM. As a second technology, there is a hierarchical text classification technology used in, for example, the automatic classification of patent documents (see, for example, Patent Document 1). As a third technology, there are technologies developed from retrieval-augmented generation (RAG) and multi-document summarization. Open-source frameworks such as LangChain and systems developed by companies such as Google and Microsoft can generate consistent summaries and answers by referring to multiple documents.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, existing similar technologies, while each employing a different approach, all faced fundamental challenges in carrying out higher-level intellectual creative activities. The first technology excels at extracting and visualizing "known relationships" that exist explicitly or implicitly within text on a large scale. However, these technologies are limited to organizing and structuring information, and do not provide a systematic integration process for interpreting the extracted collection of information and generating hypotheses or insights that cannot be predicted from the simple sum of information, such as theories or conceptual frameworks.

[0006] The second technique assists in organizing information by assigning documents to existing hierarchical classification systems. However, "hierarchical" here refers to the hierarchical structure of classification labels, and does not mean that the analysis and integration process itself is hierarchical, generating higher-level integrated hypotheses from the bottom up.

[0007] The third type of technology is one in which the primary purpose of these technologies is the efficient extraction and summarization of information common to multiple documents. There was no established methodology for intentionally integrating different or contradictory findings to generate higher-level hypotheses, etc.

[0008] Thus, while existing technologies each streamlined specific aspects of intellectual production, they all failed to address the higher-order aspects inherent in or arising from the interaction of texts. It did not provide a reproducible methodology for systematically generating conceptual knowledge.

[0009] The background technologies described above are optimized for "processing," "searching," or "summarizing" existing information contained in text sets. However, there is a problem in that there is no established methodology for systematically integrating multiple text sets and generating hypotheses or insights that could not be easily predicted from the simple summation of that information. In addition, existing technologies tend to treat multiple contradictory claims and conflicting viewpoints contained in a set of texts as "noise" or "errors" to be processed, lacking a process for analyzing the conflicting structures between these claims and systematically generating candidate hypotheses or insights that can explain or integrate those conflicts.

[0010] This invention has been made in view of the aforementioned problems, and aims to provide an information processing method, information processing system, program, artificial intelligence model, and data structure for generating "hypotheses or knowledge, etc." such as theories, hypotheses, and strategies in a systematic and reproducible form by intentionally utilizing the conflicting and contradictory structures present in large-scale text collections. [Means for solving the problem]

[0011] To achieve the above objective, according to one aspect of the present invention, an information processing method is provided for generating hypotheses or insights from a data set including linguistic information, which is executed by a computer. This method is characterized by comprising the steps of: using artificial intelligence having text generation capabilities to analyze the data set related to the theme to be analyzed and generating structured data that identifies at least one conflict or contradiction among multiple claims inherent in the data set; and having the artificial intelligence generate at least one candidate hypothesis, etc., that can explain or integrate the conflict or contradiction, based on the structured data.

[0012] Furthermore, according to another aspect of the present invention, an information processing system is provided that performs the above-described information processing method. The system is characterized by comprising: a function that performs a process to generate structured data by analyzing a data set related to a theme to be analyzed and identifying at least one conflict or contradiction among a plurality of claims inherent in the data set; and a function that performs a process to generate at least one candidate hypothesis, etc., that can explain or integrate the conflict or contradiction, based on the structured data.

[0013] These functions may be executed integrally by a single processing module, or sequentially by multiple different processing modules. Furthermore, these functions may also be implemented, for example, as a process that transitions from a first data state representing a logical inconsistency within the data set to a second data state representing an interpretive model that explains or integrates the inconsistency.

[0014] Furthermore, according to another aspect of the present invention, a program is provided for causing a computer to function as the above-mentioned information processing system. Furthermore, according to another aspect of the present invention, a data structure generated by the above-described information processing method is provided. The data structure is characterized by associating and storing either or both of the following: information relating to at least one identified conflict or contradiction, and information relating to at least one candidate hypothesis, etc., that can explain or resolve the conflict or contradiction.

[0015] Furthermore, according to another aspect of the present invention, an artificial intelligence model configured to perform the above-described information processing method, wherein, upon input of a predetermined set of data, identifies conflicts or contradictions inherent in the set of data and generates hypotheses, etc., that can explain or integrate such conflicts or contradictions. An artificial intelligence model is provided, characterized in that its internal parameters are adjusted accordingly. [Effects of the Invention]

[0016] According to the present invention, by systematically utilizing the logical inconsistencies between multiple claims present in a group of texts as the target of structured information processing, it becomes possible to efficiently generate candidate hypotheses or insights that could not be easily predicted from the simple summation or combination of such information. While the prior art treated this as "noise" to be removed for information inconsistency, the present invention utilizes it as a positive "signal" for developing it into an interpretation model with higher explanatory power. This approach can be a fundamental technology that supports the generation of theories and hypotheses that may go beyond existing knowledge frameworks and accelerates the discovery process in various intellectual production fields.

Brief Description of Drawings

[0017] [Figure 1] It is a flowchart showing the basic flow of an information processing method for generating hypotheses or findings, etc., according to an embodiment of the present invention. [Figure 2] It is a block diagram showing the functional configuration of an information processing system according to an embodiment of the present invention.

Modes for Carrying Out the Invention

[0018] Hereinafter, embodiments of the present invention will be described in detail. However, the present invention is not limited to the following embodiments, and various modifications are possible without departing from the gist of the present invention. In this specification, "text corpus" or "data group containing linguistic information" refers to an aggregate of single or multiple documents, etc., containing linguistic information related to the theme to be analyzed. The format of the data is not particularly limited, and includes files in text format (e.g., TXT, HTML, XML), files in binary format (e.g., DOCX, PDF), files in image format (e.g., JPEG, TIFF), or character transcriptions of audio data, etc., including text information inside or from which text information can be extracted, and all forms of data.

[0019] [Information Processing Method] One aspect of the present invention is an information processing method for generating hypotheses, findings, etc. from a data group including linguistic information, which is executed by a computer. The case where a user such as a researcher or an analyst executes the information processing method of this aspect step by step via a general-purpose interactive artificial intelligence interface or the like will be described.

[0020] (a) Step of causing AI to create structured data First, the user defines a theme to be analyzed and prepares a data group (hereinafter also referred to as a "corpus") including relevant linguistic information. The corpus may be a single document (for example, a single book) or a collection of multiple documents. The user inputs this corpus into an artificial intelligence having a text generation ability (hereinafter referred to as "artificial intelligence") or designates it as a processing target, and gives an instruction (prompt) such as the following. "Analyze the following corpus, identify pairs or combinations of claims that are in an opposing or contradictory relationship from the main claims included therein, and create structured data listing the content and the context in which each claim is made."

[0021] Based on this instruction, the artificial intelligence generates structured data that organizes the points of opposition between the claims existing in the corpus. The structured data referred to here includes any form interpretable by a computer in subsequent processing, such as, for example, a list form, a table form, or data composed of key-value pairs. As a preferred embodiment, in addition to the above, the user may instruct to summarize the main common views (consensus) in the corpus. This may be useful as context information for interpreting the identified contradictions and the like. Furthermore, in a preferred embodiment, the process may include the steps of: dividing the data set into multiple batches corresponding to the multiple subdomains and generating first-level structured data for each batch; and integrating the multiple first-level structured data generated for different batches to generate second-level structured data that identifies conflicts or inconsistencies between the subdomains.

[0022] (b) Steps to have artificial intelligence generate hypotheses, etc. Next, the user inputs the structured data generated in the step of generating the structured data into the artificial intelligence, or designates it as the target for processing. Then, they give instructions such as the following: "Generate at least one candidate hypothesis or similar statement that can explain or reconcile the conflicts or contradictions between the claims presented in this list."

[0023] Based on these instructions, the artificial intelligence generates candidate hypotheses and other outputs. These hypotheses and outputs may include various forms of intellectual output depending on the purpose of the analysis, such as scientific theories, definitions of product concepts, proposed research and development strategies, hypotheses regarding financial investments, and legal interpretation models. If, in step (a) above, the main common views (consensus) within the corpus are summarized, it is preferable to further instruct in step (b) to consider consistency with the main common views.

[0024] Figure 1 shows a flowchart of an example of an information processing method related to this embodiment. (i) Input step In the input step, a large amount of text (corpus), such as academic papers, reports, and news articles, which will be used as material for analysis, is prepared and entered into the information processing system. (ii) Steps to have the AI ​​create structured data In this step, prompts are created to identify inconsistencies and other issues within the input corpus, and these prompts are sent to the AI ​​along with the corpus. (iii) Processing with AI The AI ​​receives a prompt, reads the sent corpus, identifies points of inconsistency and contradictions in the claims, and organizes these inconsistencies into lists or tables (known as structured data). It then sends this structured data to an information processing system. (iv) Steps to have the AI ​​generate hypotheses, etc. The system receives structured data from the AI, creates prompts to explain inconsistencies and other issues within the structured data, and sends them to the AI ​​along with the structured data. (v) Processing in AI The AI ​​receives prompts, reads the sent structured data, and generates ideas that effectively explain inconsistencies within the structured data, or ideas that reconcile those inconsistencies. It then sends the generated ideas to the information processing system. (vi) Output step The AI ​​outputs ideas it has received. These ideas are candidates for new hypotheses or insights that are not written in the original corpus. In the output step, it is also preferable to display the generated ideas and the conflicts or contradictions that the ideas explain or resolve in relation to each other on a display device such as a user interface.

[0025] [Information Processing Systems] Another embodiment of the present invention is an information processing system that performs the above-described information processing method. This system can function as a module having the functions described below when a predetermined program is executed. Functionally, this system has the following configurations (I) to (IV). It has. It may also have other components.

[0026] (I) Corpus Input Unit: This unit has the function of collecting or receiving a data set (corpus) containing linguistic information to be analyzed from one or more sources, based on a theme specified by the user or direct input of a data set. The corpus input unit may be a keyboard or microphone for directly inputting a corpus into the system, or it may be a receiver that receives corpus input from an external source via a communication function.

[0027] (II) Conflict Structuring Processing Unit: This unit receives a corpus input from the corpus input unit, constructs a prompt containing instructions for identifying conflicts or contradictions between multiple claims present in the corpus, and transmits this prompt to an artificial intelligence capable of text generation. It also has the function of receiving structured data in which the conflicts or contradictions have been identified as a reply from the artificial intelligence. Here, the artificial intelligence may be provided as an external API or may be a model built into this system.

[0028] (III) Hypothesis generation processing unit: This unit has the function of using the structured data received by the contradiction structuring processing unit to construct a prompt requesting the generation of hypotheses that can explain or integrate the conflicts or contradictions contained in the structured data, and sending the prompt to the artificial intelligence. It also has the function of receiving a response from the artificial intelligence which will be a candidate for the final knowledge.

[0029] (IV) Knowledge Output Unit: This unit has the function of presenting candidate hypotheses, etc., received from the AI ​​by the Hypothesis Generation Processing Unit in a format that is easy for the user to interpret. The knowledge output unit only needs to be able to display the received candidate hypotheses, etc., and may be a display device such as a display or an output device such as a printer.

[0030] (V) AI with text generation capabilities: The AI ​​may be built into the information processing system or exist as an API outside the information processing system. A general-purpose generation AI can be applied.

[0031] In this embodiment, the system may execute the functions of the contradiction structuring processing unit and the hypothesis generation processing unit sequentially or integrally in response to a single instruction from the user. These processes, as a whole, can be viewed as an information processing process that generates structured data that recognizes logical problems (inconsistencies, etc.) from the data set under analysis, and then generates new ideas (hypotheses, insights, etc.) that explain or integrate those problems, thereby transitioning to a higher-order second data state. Furthermore, if an artificial intelligence model is incorporated into this system, the model may be trained or fine-tuned using the information processing method of this embodiment (for example, paired data of numerous contradictions and their resolution hypotheses) to acquire the ability to generate contradiction-resolving hypotheses for specific inputs. The generated data and ideas may include additional information to facilitate infringement detection or to aid user understanding.

[0032] For example, generated hypotheses are clearly associated with corresponding contradiction information in structured data by linkage information such as identifiers or pointers that indicate which conflicts or contradictions the hypothesis was generated to explain or reconcile. Furthermore, identified conflicts or inconsistencies, as well as generated hypotheses, may be accompanied by citation information (source information) indicating which document and which part of the input data set they rely on (e.g., title, page, paragraph, highlighted text, etc.).

[0033] The knowledge output unit preferably provides a display device, such as a user interface, that utilizes this linkage information and citation information to enable users to intuitively verify the validity of hypotheses. This display device may graphically arrange and display conflicting source texts and hypotheses that integrate them, or make them mutually referable via hyperlinks.

[0034] Furthermore, a program for operating a computer as part of the above-mentioned information processing system is also one embodiment of the present invention. Another embodiment of the present invention is an artificial intelligence model configured to receive a data set containing linguistic information as input and output hypotheses or insights, etc. This artificial intelligence model is characterized by having its internal parameters adjusted so that, upon input of a predetermined set of data, it performs the following processes: identifying conflicts or contradictions inherent in the data set, and generating hypotheses or the like that can explain or resolve the identified conflicts or contradictions.

[0035] Furthermore, a data structure generated by the above-described information processing method is also an embodiment of the present invention. Specifically, the data structure is characterized by associating and storing either or both of the following: information relating to at least one identified conflict or contradiction, and information relating to at least one candidate hypothesis, etc., that can explain or resolve the conflict or contradiction. Preferably, the data structure further stores linkage information that associates the hypothesis, etc. with the conflict or contradiction that the hypothesis, etc. explains or resolves. Furthermore, it is preferable that the data structure also stores citation information indicating which part of the data set the conflict or contradiction or hypothesis, etc., is based on.

[0036] The following shows some examples of anticipated deliverables and other items as specific application examples of this embodiment. However, the scope of application of the present invention is not limited in any way to these examples. [Application Example 1: Science, Technology, and Medical Fields] Nature of the subject of analysis and type of output: Proposal of scientific and theoretical frameworks, or support for the creation of hypotheses regarding drug targets and new treatments, through integrated analysis of a vast amount of academic papers, experimental data, clinical reports, etc. [Application Example 2: Manufacturing and Technology Development Sector] Nature of the subject of analysis and type of deliverables: Identifying untapped technological areas through integrated analysis of patent information and technical literature from around the world, formulating research and development strategies accordingly, or defining concepts for next-generation products. [Application Example 3: Finance and Economics Sector] Nature of the subject of analysis and type of deliverables: Constructing financial investment hypotheses that suggest market inefficiencies, or supporting the generation of predictive models regarding economic trends, by integrating qualitative data such as corporate financial reports, market news, and analyst reports. [Application Example 4: Policy Making and Law Fields] Nature of the subject of analysis and type of deliverables: Proposal of legal interpretation models or support for formulating evidence-based policy options through integrated analysis of past case precedents, bills, socioeconomic data, etc. [Industrial applicability]

[0037] This invention can be used in an extremely wide range of industrial fields where it is necessary to create value from any intellectual resources that exist as text. For example, in the field of scientific and technological research. This technology can contribute to the efficiency of research and development by comprehensively analyzing vast amounts of academic papers and technical reports, and supporting the creation of research hypotheses and theoretical models. In the fields of manufacturing and technology development, it can be used to analyze patent information and market research reports from around the world to understand the trends of competitors and to gain ideas for discovering untapped technological areas and future product concepts. In the field of finance and investment, it can be used to integrate qualitative text data such as corporate financial reports, earnings presentation materials, and news articles to identify market inefficiencies and risk factors, and to provide information for building unique investment strategies. Furthermore, in various industries where sophisticated judgment and creativity based on verbalized information are required, such as law, policy making, medicine, marketing, creative activities, pharmaceutical development, and financial investment strategies, this invention has high applicability as a foundational technology for improving intellectual productivity.

Claims

1. An information processing method for generating hypotheses or insights from a set of data containing linguistic information, which is performed by a computer. The process involves using artificial intelligence with text generation capabilities to analyze the data set related to the theme to be analyzed, and generating structured data that identifies at least one conflict or contradiction among multiple claims inherent in the data set. The steps include: causing the artificial intelligence to generate at least one candidate hypothesis that can explain or integrate the conflicts or contradictions based on the structured data; Includes, The aforementioned data set includes text that can be classified into multiple subdomains, The step of generating the structured data includes dividing the data set into multiple batches corresponding to the multiple subdomains, and generating first-level structured data for each batch, The steps include: integrating multiple first-tier structured data generated for different batches to generate second-tier structured data that identifies conflicts or inconsistencies between the subdomains; An information processing method characterized by including

2. The information processing method according to claim 1, wherein the step of generating the structured data further includes identifying key commonalities inherent in the data set.

3. The information processing method according to claim 2, wherein the step of causing the artificial intelligence to generate the hypothesis further includes instructing the artificial intelligence to take into consideration consistency with the identified main common views.

4. The information processing method according to claim 1, further comprising the step of visually relating and displaying the generated hypothesis and the conflict or contradiction that the hypothesis explains or resolves on a user interface.

5. An information processing system that generates hypotheses or insights from a data set containing linguistic information, A structured processing unit that analyzes the data set related to the theme to be analyzed and generates structured data that identifies at least one conflict or contradiction among multiple claims inherent in the data set, A hypothesis generation processing unit generates at least one candidate hypothesis that can explain or resolve the conflict or contradiction based on the structured data, Equipped with, The aforementioned data set includes text that can be classified into multiple subdomains, The structuring processing unit divides the data group into multiple batches corresponding to the multiple subdomains, and generates first-level structured data for each batch. This process integrates multiple first-tier structured data generated for different batches to generate second-tier structured data that identifies conflicts or inconsistencies between the subdomains. An information processing system characterized by the following:

6. A program for causing a computer to function as the information processing system described in claim 5.

7. A step of performing the information processing method described in any one of claims 1 to 4, A step in which, using the candidate hypotheses or findings generated in the above step, processes are carried out to achieve at least one objective selected from the group consisting of constructing scientific theories, formulating research and development strategies, formulating financial investment strategies, defining product concepts, developing pharmaceuticals, and formulating policies. A method characterized by including the following.

Images