Document analysis support device and document analysis support method

Abstract

It helps to accurately analyze the correspondence between documents. [Solution] A document analysis support device 10 comprising a storage device for storing a first document and a second document, a summarization process which generates a new document that summarizes the content of the documents by inputting each document into a trained model 23 that outputs text corresponding to the input text, a similarity determination process which calculates the similarity between the summary document relating to the first document and the summary document relating to the second document, and determines whether or not there is a correspondence between the first document and the second document based on the calculated similarity, and a result output process which outputs information representing the presence or absence or degree of the determined correspondence to an output device.

Description

Technical Field

[0001] The present invention relates to a document analysis support device and a document analysis support method.

Background Art

[0002] In system development, it is often necessary to go through various processes such as requirement definition, basic design, and detailed design. In this case, it is required that the documents created in each process accurately reflect the content of the documents created in the previous process. The management of such processes, that is, the management of traceability, has become increasingly important in recent years.

[0003] In the management of document traceability, it is important to confirm that the correspondence relationship between documents is established between processes. Therefore, it is conceivable to use large language models (LLMs) that have been developed in recent years. For example, as a technique for generating answers using an LLM, Patent Document 1 describes an additional sentence generation unit that generates an additional sentence related to a question sentence with a total number of characters that does not exceed the character limit that can be input to the answer generation unit of the large language model (LLM) in accordance with the number of characters of the question sentence input by the user, and a prompt generation unit that generates a prompt by adding, as reference information, the additional sentence generated by the additional sentence generation unit to the input question sentence, and generates an additional sentence so that the number of characters does not exceed the character limit that can be input to the LLM as a prompt, thereby generating a prompt in which an effective sentence is added as reference information to the input question sentence within the determined character limit range.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

[0005] However, as the content of documents became more complex, LLM processing was applied to matters unrelated to the essential content of the documents, potentially reducing the accuracy of comparisons.

[0006] This invention has been made in view of these circumstances, and its purpose is to provide a document analysis support device and a document analysis support method that can support the accurate analysis of correspondences between documents. [Means for solving the problem]

[0007] One of the present inventions for solving the aforementioned problems is a document analysis support device comprising: a storage device for storing a first document and a second document; a summarization process that generates a new document, which is a summary document, which is a summary of the contents of the first document, by inputting each of the first documents into a trained model that outputs text corresponding to the input text; a similarity determination processing unit that calculates the similarity between the summary document relating to the first document and the summary document relating to the second document, and determines whether or not there is a correspondence between the first document and the second document based on the calculated similarity; and a result output processing unit that outputs information indicating whether or not there is a correspondence determined to an output device. [Effects of the Invention]

[0008] According to the present invention, it is possible to support the accurate analysis of correspondences between documents. [Brief explanation of the drawing]

[0009] [Figure 1] This figure shows an example of the configuration of the document analysis support system according to this embodiment. [Figure 2] This figure shows an example of a settings management table. [Figure 3]This is a flowchart illustrating an example of document analysis support processing. [Figure 4] This is a flowchart illustrating an example of document analysis support processing. [Figure 5] This figure shows an example of the output network graph. [Figure 6] This figure shows an example of a higher-layer document with additional information added. [Figure 7] This figure shows an example of a list of potential omissions or omissions that may be created. [Modes for carrying out the invention]

[0010] Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[0011] The document analysis support system according to this embodiment is an information processing system that supports the analysis of the correspondence (consistency) of information between multiple types of document groups. Here, a document group is defined as a collection of documents created for each process of a business.

[0012] In this embodiment, the document set is assumed to be a set of documents created in each stage of system (program) development. The contents of the documents in the document set differ depending on the development stage. Specifically, the earlier the stage, the more abstract the description becomes (e.g., a specification), and the later the stage, the more specific the description becomes (e.g., a design document). Hereinafter, a set of documents that describe certain matters more abstractly, such as a specification, will be referred to as "higher-layer documents" or "first documents," and a set of documents that describe certain matters more specifically will be referred to as "lower-layer documents" or "second documents." In other words, the document analysis support system assists in analyzing the correspondence between each document in the higher-layer document set and each document in the lower-layer document set.

[0013] FIG. 1 is a diagram showing an example of the configuration of a document analysis support system 1 according to the present embodiment. The document analysis support system 1 includes a document analysis support device 10 and an LLM server 20.

[0014] The document analysis support device 10 is an information processing device used by a user who performs document analysis.

[0015] The LLM server 20 is an information processing device accessed by the document analysis support device 10.

[0016] The LLM server 20 stores a learned model 23. The learned model 23 is a mathematical model (for example, a large language model (LLM)) that receives a prompt of a natural language text and outputs a natural language text corresponding to the input prompt. The learned model 23 is constructed by, for example, BERT (Bidirectional Encoder Representations from Transformers), XLNet, or GPT (Generative Pre-trained Transformer). In the following, the learned model 23 may also be referred to as a generative AI.

[0017] The LLM server 20 also has functional units of a communication processing unit 21 and an LLM execution unit 22.

[0018] The communication processing unit 21 performs data communication with the document analysis support device 10.

[0019] The LLM execution unit 22 inputs the prompt received from the document analysis support device 10 to the learned model 23, thereby causing the learned model 23 to output corresponding data (specifically, as will be described later, data of a text representing a superordinate concept for the input). The LLM execution unit 22 transmits the output data to the document analysis support device 10.

[0020] The document analysis support device 10 and the LLM server 20 are connected by a wired or wireless communication network 5 such as the Internet, a LAN (Local Area Network), a WAN (Wide Area Network), or a dedicated line.

[0021] As shown in FIG. 1, the document analysis support device 10 includes an arithmetic device 11 such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), a memory 12 such as a RAM (Random Access Memory) or a ROM (Read Only Memory), a storage device 13 such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), an input device 14 composed of a keyboard, a mouse, a touch panel, etc., an output device 15 for performing screen display composed of a monitor (display), etc., and a communication device 16 composed of a NIC (Network Interface Card), a wireless communication module, a USB (Universal Serial Interface) module, a serial communication module, etc. Note that the LLM server 20 also includes similar hardware.

[0022] The document analysis support device 10 stores each data of the document DB 120, the setting value management table 130, the network graph 500, and the omission concern list 700.

[0023] The document DB 120 stores each document of each layer.

[0024] The network graph 500 is graph data representing the correspondence relationship of matters described in the document between layers. The omission concern list 700 is data storing a list of the correspondence relationships of matters described in the document between layers. The network graph 500 and the omission concern list 700 are generated by each process of the document analysis support device 10. Details of the network graph 500 and the omission concern list 700 will be described later.

[0025] The setting value management table 130 stores parameters related to the processing performed by the document analysis support device 10.

[0026] Furthermore, the document analysis support device 10 has the following functional units: a communication processing unit 101, a omission / missing data analysis processing execution unit 102, and a generation AI execution unit 103.

[0027] The communication processing unit 101 performs data communication with the LLM server 20. The omission / missing data analysis processing execution unit 102 analyzes the correspondence between documents in layers. The generation AI execution unit 103 obtains output data by inputting prompts to the trained model 23 on the LLM server 20.

[0028] The data omission and missing information analysis execution unit 102 includes the following functional units: file reading processing unit 104, text extraction processing unit 105, preprocessing unit 106, summarization processing unit 107, vectorization processing unit 108, similarity determination processing unit 109, and result output processing unit 110.

[0029] The file reading processing unit 104 reads the data for each document (the first document and the second document) from the document DB 120.

[0030] The text extraction processing unit 105 extracts the text portion from each document read by the file reading processing unit 104.

[0031] The preprocessing unit 106 performs preprocessing to remove predetermined types of text from each document from which the text portion has been extracted.

[0032] The summarization processing unit 107 inputs each document preprocessed by the preprocessing unit 106 into the trained model 23, thereby generating new summary documents that summarize the content of the original documents.

[0033] The summarization processing unit 107 may, after determining whether or not to perform the summarization processing based on predetermined criteria such as the part of speech or sentence length handled in the document, summarize each document by converting it into an appropriate bulleted list or similar format.

[0034] Furthermore, the text extraction processing unit 105 or the summarization processing unit 107 may, before or after performing the summarization processing, divide the first document and the second document (and the resulting summarized documents) into multiple texts (hereinafter referred to as chunks) using a predetermined algorithm.

[0035] The vectorization processing unit 108 vectorizes each summary document (chunk).

[0036] The similarity determination processing unit 109 calculates the similarity between the summary document (chunk) related to the first document and the summary document (chunk) related to the second document based on the vector calculated by the vectorization processing unit 108, and determines whether or not there is a correspondence between the first document (chunk) and the second document (chunk) based on the calculated similarity.

[0037] The result output processing unit 110 outputs information to the output device 15 that indicates the presence or degree of a correspondence determined by the similarity determination processing unit 109. For example, the result output processing unit 110 generates a network graph 500 and a list of potential omissions and errors 700.

[0038] (Settings management table) Figure 2 shows an example of a setting value management table 130. The setting value management table 130 contains data for the memory size 131 that can be used for processing in the similarity determination processing unit 109, etc., a search parameter 132 related to how broadly the correspondence between documents (specifically chunks) is determined in the similarity determination processing unit 109 (how many chunks are involved), a similarity calculation method 133 used in the similarity determination processing unit 109, an upper limit 134 for the number of items to be determined to have a correspondence, a similarity threshold 135 for determining that a correspondence exists, a gravity value 136 in the network graph to be created, and a gravity attractor 137 in the network graph to be created. The similarity calculation method 133 is, for example, Euclidean distance, vector dot product, or cosine similarity.

[0039] The functions of each functional unit of the document analysis support device 10 described above are realized by the hardware of the document analysis support device 10, or by the arithmetic unit 11 of the document analysis support device 10 reading and executing programs stored in the memory 12 or storage device 13. These programs are stored in storage devices such as secondary storage devices, non-volatile semiconductor memory, hard disk drives, SSDs, or recording media readable by each device, such as IC cards, SD cards, or DVDs. Each device may be realized in whole or in part using virtual information processing resources provided using virtualization technology or process space isolation technology, such as virtual servers provided by a cloud system. Furthermore, all or part of the functions provided by these devices may be realized by services provided by a cloud system via an API (Application Programming Interface), for example. Next, the processing performed in the document analysis support system 1 will be described.

[0040] Figures 3 and 4 are flowcharts illustrating an example of document analysis support processing performed in the document analysis support system 1 (divided into two figures for space limitations). Document analysis support processing is initiated, for example, when a user makes a predetermined input to the document analysis support device 10.

[0041] First, as shown in Figure 3, the document analysis support device 10 reads each document in each layer and extracts the text portion from each read document (s11). For example, the document analysis support device 10 extracts the text portion of the document, excluding parts where images or videos are embedded, parts that define the document layout, etc.

[0042] The document analysis support device 10 analyzes each text extracted in s11, identifies predetermined types of text portions (words, phrases, or sentences) that are not essential to the content of the document, and performs preprocessing to remove the identified portions from the text (s13).

[0043] For example, the document analysis support device 10 may obtain text from which the non-essential parts have been removed by inputting a prompt to the generating AI that includes the target text and the specification of the non-essential parts mentioned above, or it may obtain text from which the non-essential parts have been removed by performing a text search on the target text for the non-essential parts mentioned above. The non-essential parts include, for example, numbers, times, amounts, fixed phrases such as greetings, prescribed symbols (such as (C)Copyright), or proper nouns such as names of people or places.

[0044] The document analysis support device 10 generates new texts (summary texts) that summarize the content of each text obtained in s13 (s15).

[0045] For example, the document analysis support device 10 generates a prompt requesting the generating AI to summarize the target text, and inputs the generated prompt into the generating AI. Additional requests regarding the summarization may be attached to this prompt. These additional requests may include, for example, instructions to summarize within a predetermined number of characters, instructions to convert the text into a number of bullet points corresponding to the number of characters in the target text, or instructions to minimize the creativity or unpredictability of the output data.

[0046] Alternatively, the document analysis support device 10 may generate a summarized text of the target text by performing a predetermined lexical analysis on the target text, rather than using a generation AI. In this case, for example, the document analysis support device 10 may obtain a predetermined dictionary data and, based on the obtained dictionary data, replace each word or phrase constituting the target text with a word or phrase of a higher-level concept.

[0047] The document analysis support device 10 divides each summary text obtained in s15 into multiple text parts (chunks), and converts each divided chunk into a numerical value (in this embodiment, it is an embedded representation (vector), but other numerical formats may also be used) (s17).

[0048] For example, the document analysis support device 10 may divide the summary text into sentence units or into paragraph units. Alternatively, the document analysis support device 10 may divide the text into units with a number of characters less than or equal to a predetermined number of characters. Furthermore, the document analysis support device 10 may refer to predetermined layout information contained in the document corresponding to the summary text and generate chunks in layout units (e.g., cells, sheets, blocks, objects, etc.) indicated by the layout information. Alternatively, the document analysis support device 10 may perform the division into chunks in advance, before summarizing, during processing s11 or s13, rather than in this process.

[0049] The document analysis support device 10 calculates the similarity between each chunk (specifically, the similarity between chunks in different layers, and the similarity between chunks within the same layer) according to the memory size 131 of the setting value management table 130 (s19).

[0050] Specifically, for example, the document analysis support device 10 calculates the similarity between a vector of a certain chunk selected from the first layer and a vector of a certain chunk selected from the second layer, according to the method shown in the similarity calculation method 133 of the setting value management table 130.

[0051] Based on the similarity scores calculated in s19, the document analysis support device 10 identifies combinations of chunks from the first document (upper layer) and chunks from the second document (lower layer) that correspond to each other, starting from the upper layer chunks, under the constraints indicated by the upper limit value 134 of the number of entries in the setting value management table 130 (s21).

[0052] Specifically, for example, the document analysis support device 10 identifies chunks in lower layers whose similarity to each chunk in the upper layer is greater than or equal to the threshold 135 in the setting value management table 130.

[0053] Furthermore, for example, the document analysis support device 10 may identify other chunks in the upper layer whose similarity to each chunk in the upper layer is greater than or equal to the threshold 135 of the setting value management table 130, and identify chunks in the lower layer whose similarity to the identified upper layer chunk is greater than or equal to the threshold 135 of the setting value management table 130. In addition, for example, the document analysis support device 10 may identify other chunks in the lower layer whose similarity to the identified lower layer chunk is greater than or equal to the threshold 135 of the setting value management table 130. In this way, the document analysis support device 10 determines the correspondence between upper layer chunks and lower layer chunks via one or more chunks.

[0054] The document analysis support device 10 stores, for each of the chunks in the upper layer and the combination of chunks in the upper layer that it determined to have a correspondence in s21, that there is a correspondence between each chunk (s21: Yes, s23). After that, the process in s27 is performed.

[0055] The document analysis support device 10 adds information to the list of potential omissions and missing data 700 (s21: No, s25) indicating that there is no correspondence between each of the upper layer chunks and combinations of upper layer chunks other than those determined to have a correspondence in s21. After that, the process in s27 is performed.

[0056] In s27, the document analysis support device 10 checks whether there are any chunk combinations for which similarity has not been determined by the processes in s19 and s21. If there are chunk combinations for which similarity has not been determined (s27: Yes), the document analysis support device 10 repeats the process in s19 for those combinations. If there are no chunk combinations for which similarity has not been determined (s27: Yes), the document analysis support device 10 executes the process in s29.

[0057] Furthermore, the document analysis support device 10 may add chunk combinations for which similarity was not calculated to a list of potential omissions or missing information, treating them as having no correspondence.

[0058] Next, as shown in s29 of Figure 4, the document analysis support device 10 creates a network graph representing the correspondence between each chunk in the layers, which was identified in the processing up to s27, based on the gravity value 136 and gravity attractor 137 of the setting value management table 130, and displays it on the screen.

[0059] Furthermore, the document analysis support device 10 searches for combinations of chunks between different layers whose similarity has been determined in the processing up to s27, in which there is no correspondence or the correspondence is insufficient (e.g., similarity is less than a predetermined value) (s31).

[0060] If there are any chunk combinations that do not have a correspondence or have insufficient correspondences (s33:Yes), the document analysis support device 10 executes the process in s35. If there are no chunk combinations that do not have a correspondence or have insufficient correspondences (s33:No), the document analysis support device 10 executes the process in s37.

[0061] In s35, the document analysis support device 10 adds information (hereinafter referred to as additional information) indicating the lack of correspondence or insufficient correspondence (chunk content) in the corresponding document for each combination of chunks that does not have a correspondence or has insufficient correspondence (hereinafter referred to as non-corresponding combination) (s35). After that, the processing in s37 is performed.

[0062] For example, the document analysis support device 10 identifies the layer on which the corresponding chunk does not exist among the incompatible combinations (hereinafter referred to as the "missing layer," for example, a lower layer), and adds additional information to the document relating to the identified missing layer, in the form of comments or the like, indicating that the information for the corresponding chunk is missing (or that such a chunk does not exist). The additional information may be, for example, information for the corresponding chunk, or information about the location of the document in which the original text of the corresponding chunk is written. This location is identified, for example, based on the document's layout information.

[0063] Furthermore, for example, the document analysis support device 10 may add additional information to documents in layers that are not missing (e.g., documents in higher layers) indicating the presence (or absence) of missing chunks. The additional information may be added, for example, to the location of the corresponding chunk in the document.

[0064] In s37, the document analysis support device 10 checks whether there are any chunk combinations that have not been processed since s31. If there are chunk combinations that have not been processed since s31 (s37: No), the document analysis support device 10 repeats the processing from s31 onwards for those chunk combinations. If there are no chunk combinations that have not been processed since s31 (s37: Yes), the document analysis support device 10 terminates the document analysis support process.

[0065] (Network graph) Here, Figure 5 shows an example of the output network graph. This network graph 500 is a network graph that shows the correspondence between chunks across multiple layers. This network graph 500 represents each chunk of each document as a node 501, and the presence or degree of correspondence between chunks of different layers is represented as a link 502. Specifically, each node 501 is represented by a type of shape corresponding to its layer. The links 502 may be represented by shapes corresponding to the level of similarity or the presence or absence of a correspondence (whether the similarity is above a predetermined threshold).

[0066] For example, as shown in the enlarged portion 510 of the network graph 500, node 501A represents a chunk in the upper layer, and node 501B represents a chunk in the lower layer. Node 501C is connected to node 501A in the upper layer and node 501B in the lower layer by link 502 (they are in a corresponding relationship or have a similarity above a predetermined threshold), so it can be said that there are descriptions of related documents from the upper layer down to the lower layer, ensuring traceability. On the other hand, node 501D is connected to node 501A in the upper layer, but there is no connection to a node in the lower layer, resulting in a gap in traceability.

[0067] Furthermore, as shown in the other enlarged section 520, node 501E represents a higher layer, and node 501F represents a chunk of node 501E that is at the same level as or in a different layer. Node 501G is linked to node 501E of the higher layer, and is also linked to node 501F, which is at the same level as or in a different layer. Therefore, traceability is ensured between these nodes.

[0068] Such a network graph 500 allows users to determine whether there are correspondences (consistencies, etc.) between documents in different layers.

[0069] (Document with added information) Figure 6 shows an example of a higher-layer document to which additional information has been added. This document 600 is, for example, a specification, and in the area 601 where the text corresponding to the missing chunks is written, a comment 603 has been added as additional information stating that information regarding the missing chunks (text such as chunks separated from the summary document 602) is not included in the lower-layer document (design document, etc.).

[0070] (List of potential omissions / omissions) Figure 7 shows an example of a list of potential omissions and missing data that may be created. For each mismatched combination, this list of potential omissions and missing data 700 is configured with the following information: the document chunk 701 of the layer that is not the missing layer, the location 702 of that document, and a message 703 ("no correspondence") indicating that there is no chunk corresponding to the missing layer.

[0071] As described above, the document analysis support device 10 of this embodiment inputs the first document (a group of upper-layer documents) and the second document (a group of lower-layer documents) into the trained model 23, generates summary documents that summarize the content of each document, determines whether there is a correspondence between the first document and the second document, or the degree of such a correspondence, based on the similarity between the summary document for the first document and the summary document for the second document, and outputs information representing the determined presence or degree of the correspondence to the output device (for example, a list of potential omissions and missing information 700).

[0072] In other words, the document analysis support device 10 of this embodiment can extract important parts of documents in terms of their substantive content rather than their format (word level, etc.) by summarizing documents in different layers, and by determining the similarity between these summarized documents, it can accurately determine the correspondence between documents in different layers according to their substantive content.

[0073] Thus, the document analysis support device 10 of this embodiment can help accurately analyze the correspondence between documents.

[0074] In this embodiment, the document analysis support device 10 divides the summary document related to the first document and the summary document related to the second document into multiple texts (chunks), calculates the similarity between each chunk related to the first document and each chunk related to the second document, and determines, based on the calculated similarity, whether or not there is a correspondence between each chunk of the first document and each chunk of the second document.

[0075] In this way, by dividing each document into multiple chunks and calculating the similarity for each chunk, the correspondence between documents can be determined with an appropriate level of information granularity.

[0076] Furthermore, the document analysis support device 10 of this embodiment removes a predetermined type of text from each document and inputs each document from which the predetermined type of text has been removed into the trained model 23.

[0077] In this way, by pre-removing certain types of text from each document (for example, text that clearly does not need to be summarized), it is possible to output an appropriate summary document.

[0078] Specifically, the document analysis support device 10 of this embodiment removes at least one of the following texts from each document: a predetermined proper noun, a numerical value, or a fixed phrase.

[0079] This allows you to remove information that is not important to the document's content and output a suitable summary document.

[0080] Furthermore, the document analysis support device 10 of this embodiment outputs a network graph 500 in which the first document and the second document are represented as nodes 501, and the presence or degree of correspondence between the first document and the second document is represented as a link 502.

[0081] This makes it easy to grasp the overall correspondence between the first document and the second document.

[0082] Furthermore, the document analysis support device 10 of this embodiment adds information (additional information) to the first document that it has determined has no correspondence with the second document or has insufficient correspondence with the second document, indicating the parts that have no correspondence or are insufficient with the first document.

[0083] In this way, by adding information about missing or insufficient correspondences to the documents being analyzed, it becomes possible to clearly understand the correspondences for each document.

[0084] Furthermore, the document analysis support device 10 of this embodiment uses prompts to input to the trained model 23 that include each document and at least one of the following instructions: an instruction to summarize each document within a predetermined number of characters, an instruction to summarize each document by converting it into a number of bullet points corresponding to its character count, or an instruction to minimize the creativity or unpredictability of the summarized document.

[0085] By structuring the prompts in this way, it becomes possible to create appropriate summary documents that are suitable for determining the correspondence between documents.

[0086] The above description of embodiments is for the purpose of facilitating understanding of the present invention and does not limit it. The present invention can be modified and improved without departing from its spirit, and equivalents thereof are included.

[0087] For example, some of the hardware components of each device in this embodiment may be provided in other devices.

[0088] Furthermore, each program of each device may be provided in other devices, a program may consist of multiple programs, or multiple programs may be integrated into a single program.

[0089] Furthermore, in this embodiment, the documents being compared are assumed to be documents created in the preceding and succeeding processes, respectively, but the relationship between any documents whose textual content corresponds may also be that of any two documents.

[0090] Furthermore, in this embodiment, a network graph is shown as an example of a graph representing the correspondence between documents, but other types of graphs may be used. [Explanation of Symbols]

[0091] 10 Document analysis support device, 102 Analysis processing execution unit, 106 Preprocessing unit, 107 Summarization processing unit, 108 Vectorization processing unit, 109 Similarity determination processing unit, 110 Result output processing unit

Claims

1. A storage device for storing the first document and the second document, and A summarization process is performed in which each of the aforementioned documents is input into a trained model that outputs text corresponding to the input text, thereby generating a new document, which is a summary document, that summarizes the content of the aforementioned document. A similarity determination process that calculates the similarity between the summary document relating to the first document and the summary document relating to the second document, and determines whether or not there is a correspondence between the first document and the second document based on the calculated similarity, A computing device that performs result output processing, which outputs information indicating the presence or degree of the determined correspondence to an output device. A document analysis support device equipped with the following features.

2. The aforementioned computing device is A preprocessing step is performed to remove a predetermined type of text from each of the aforementioned documents. In the summarization process, each document from which the predetermined type of text has been removed is input to the trained model. The document analysis support device according to claim 1.

3. The aforementioned computing device is In the aforementioned preprocessing, at least one of the following texts is removed from each document: The document analysis support device according to claim 2.

4. The aforementioned computing device is In the result output processing described above, the first document and the second document are represented as nodes, and a graph is generated in which the presence or degree of correspondence between the first document and the second document is represented as a link, and the generated graph is output to the output device. The document analysis support device according to claim 1.

5. The aforementioned computing device is In the result output process described above, if it is determined that there is no correspondence with the second document or that there is insufficient correspondence with the second document, information is added to the first document that has no correspondence or insufficient correspondence with the first document. The document analysis support device according to claim 1.

6. The aforementioned computing device is A splitting process is performed on the summary document relating to the first document and the summary document relating to the second document, respectively, to divide them into multiple texts using a predetermined algorithm. In the similarity determination process, the similarity between each of the divided texts in the first document and each of the divided texts in the second document is calculated, and based on the calculated similarity, the existence or degree of correspondence between each of the divided texts in the first document and each of the divided texts in the second document is determined. The document analysis support device according to claim 1.

7. The aforementioned computing device is In the summarization process, the trained model is input a prompt that includes each of the documents and at least one of the following instructions: an instruction to summarize each of the documents within a predetermined number of characters, an instruction to convert each of the documents into a number of bullet points corresponding to the number of characters to constitute a summary, or an instruction to minimize the creativity or unpredictability of the summarized documents. The document analysis support device according to claim 1.

8. A document analysis support method using an information processing device comprising a storage device for storing a first document and a second document, and a computing device, wherein the computing device is A summarization process is performed in which each of the aforementioned documents is input into a trained model that outputs text corresponding to the input text, thereby generating a new document, which is a summary document, that summarizes the content of the aforementioned document. A similarity determination processing unit calculates the similarity between the summary document relating to the first document and the summary document relating to the second document, and determines whether or not there is a correspondence between the first document and the second document based on the calculated similarity; The system performs a result output process that outputs information to an output device indicating the presence or degree of the determined correspondence relationship. Document analysis support method.

Images