JP2024091709A - Sentence creation device, sentence creation method, and sentence creation program - Google Patents

Abstract

The present invention aims to provide a sentence creation device, a sentence creation method, and a sentence creation program that enable cross-sectional access to a huge amount of specialized documents and pinpoint access to necessary specialized documents based on advanced analysis.
SOLUTION
The sentence creation device includes a receiving unit that receives document input, a division unit that divides the document into a plurality of sections, a keyword tag generation unit that generates keyword tags for each of the plurality of sections divided by the division unit, a determination unit that determines whether the plurality of sections contain a wording that suggests a reference, and an identification unit that identifies the referenced section when the wording that suggests a reference is contained.
[Selected figure] Figure 4

Description

The present invention relates to a sentence creation device, a sentence creation method, and a sentence creation program. Specifically, the present invention relates to a sentence creation device, a sentence creation method, and a sentence creation program that structure technical documents and facilitate reference between technical documents.

Conventionally, a sentence creation device, a sentence creation method, and a sentence creation program are known that classify multiple specialized documents into multiple categories to make it easier for a user who wants to search and refer to the specialized documents.

JP 2003-16055 A Patent No. 6469919

In the sentence creation device and sentence creation method of the above Patent Document 1, word or phrase data is input before the selection of a sentence template, and sentence candidates are generated using only sentence templates that describe the semantic categories of all the input word or phrase data and auxiliary words contained in the character string. In addition, in the sentence creation device, sentence creation method, and sentence creation program of the above Patent Document 2, the judgment is appropriately divided into short sentences, the short sentences are classified, and from these classified or categorized short sentences, other short sentences of the judgment related to the judgment, such as commentaries, are searched for, and sentences such as complaints are created from these classified or categorized short sentences.

Until now, legal textbooks and government documents were scattered in various places and media, and were often analog media that could not be searched. Even if some of them were digitized, they were often limited to being provided as printable PDF data with low viewability and searchability. Therefore, in order for legal practitioners to access the information they need, they had to identify documents, obtain documents, and use tables of contents and indexes to find the information they need from the documents, which required a huge amount of time and effort. In addition, while conventional e-book readers, such as PDF viewers, can perform full-text searches using a single keyword tag or simple transitions from the table of contents, they cannot narrow down chapters and sections using multiple keyword tags or smoothly and simultaneously access and view related descriptions inside and outside the book. The sentence creation device and sentence creation method of the above-mentioned Patent Document 1 can only create sentences within the same document, so it is not possible to pinpoint access to other documents. The sentence creation device, sentence creation method, and sentence creation program of the above-mentioned Patent Document 2 are technologies that only support judgment documents, so they cannot search and view all documents across the board. As a result, the above prior art technologies were unable to adequately meet the need to quickly and accurately identify and narrow down the necessary information in legal research, which requires advanced and complex information gathering and analysis.

Therefore, in recent years, there has been a demand for a text creation device, a text creation method, and a text creation program that have a highly accurate search function that exceeds general full-text searches by allowing users to move to another document with one click when there is a reference to another document in a document, such as a reference or note, and by allowing users to display the other document as a related document when a certain part of the document being viewed is referenced by another document, and by assigning keyword tags to each section of the document, such as chapters and paragraphs.

Therefore, the present invention aims to provide a text creation device, text creation method, and text creation program that electronically centralizes distributed specialized document resources, making it possible to search for information across the board and quickly access necessary information, and not only makes specialized documents available for electronic viewing, but also analyzes and organizes specialized documents using natural language processing technology, etc., and analyzes the cross-reference relationships of specialized documents and the use of legal-specific terms (keyword tags) in sentences to organize the relationships between specialized documents (for example, legal documents; by way of example, documents in other specialized fields where similar cross-references between documents are important may also be used), thereby enabling pinpoint access to necessary information from all specialized documents and thorough investigation of related information.

To solve the above problem, the sentence creation device of the present invention includes a receiving unit that receives document input, a division unit that divides the document into multiple sections, a keyword tag generation unit that generates keyword tags for each of the multiple sections divided by the division unit, a determination unit that determines whether the multiple sections contain a wording that suggests a reference, and an identification unit that identifies the referenced section when the wording that suggests a reference is contained.

The sentence creation device of the present invention may further include a link generation unit that generates a link from the document to the section of the reference destination identified by the identification unit, and a link output unit that outputs a link to the reference destination identified by the identification unit.

The determination unit may detect phrases that suggest references using at least one of a method for solving sequence labeling problems, such as pattern matching and conditional random fields, and a neural network, and identify the section to which each link refers.

The sentence creation device of the present invention may further include a related document acquisition unit that acquires information indicating related documents, which are reference data that are similar to the document and indicate the content of the text that suggests a reference, and a related document output unit that outputs the related documents.

The related document acquisition unit may acquire related documents from the section and the referenced section.

The related document acquisition unit may acquire related documents from documents other than the document.

The related document acquisition unit may acquire related documents by either network structuring, which uses graph theory to infer relevance, or by similarity calculation in a low-dimensional Euclidean space based on embedding.

In order to solve the above problem, the sentence creation method of the present invention includes a receiving step for receiving document input, a division step for dividing the document into a plurality of sections, a keyword tag generating step for generating keyword tags for each of the plurality of sections divided in the division step, a determination step for determining whether the plurality of sections contain a wording that suggests a reference, and an identification step for identifying the referenced section when the wording that suggests a reference is contained.

The sentence creation method of the present invention may further include a generating step of generating a link from the document to the section of the reference identified in the identifying step, and an output step of outputting a link to the reference identified in the identifying step.

The determination step may detect reference-indicating text using at least one of techniques for solving sequence labeling problems, such as pattern matching, conditional random fields, and neural networks, and identify the referenced section for each link.

The sentence creation method of the present invention may further include a related document acquisition step of acquiring information indicating a related document, which is data of a reference destination that indicates the content of a sentence similar to the document and suggests a reference, and a related document output step of outputting the related document.

The related document acquisition step may acquire related documents from the section and the referenced section.

The related document acquisition step may acquire related documents from other documents different from the document.

The relevant document acquisition step may acquire relevant documents by either network structuring, which uses graph theory to infer relevance, or by similarity calculation in a low-dimensional Euclidean space based on embedding.

In order to solve the above problem, the sentence creation program of the present invention has a computer realize a receiving function for receiving document input, a dividing function for dividing the document into multiple sections, a keyword tag generating function for generating keyword tags for each of the multiple sections divided by the dividing function, a determining function for determining whether a section contains a wording that suggests a reference, and an identifying function for identifying the section to which the reference is directed when the wording that suggests a reference is contained.

The sentence creation program of the present invention may further cause the computer to realize a link generation function that generates a link from the document to a section of the reference destination identified by the specific function, and a link output function that outputs a link to the reference destination identified by the specific function.

The determination function may detect reference-indicating text using at least one of the following techniques: pattern matching, methods for solving sequence labeling problems such as conditional random fields, and neural networks, and identify the referenced section for each link.

The sentence creation program of the present invention may further implement in the computer a related document acquisition function for acquiring information indicating related documents, which are reference data that indicate the content of a sentence that is similar to a document and suggests a reference, and a related document output function for outputting related documents.

The related document retrieval function may retrieve related documents from a section and its referenced sections.

The related document acquisition function may also acquire related documents from other documents different from the document.

The related document acquisition function may acquire related documents either by network structuring, which uses graph theory to infer relevance, or by calculating similarity in a low-dimensional Euclidean space based on embedding.

According to the sentence creation device, sentence creation method, and sentence creation program of the present invention, by electronically centralizing distributed technical document resources, it is possible to search for information across the board and quickly access necessary information. It not only makes technical documents available for electronic viewing, but also analyzes and organizes technical documents using natural language processing technology, etc., and analyzes the cross-reference relationships of technical documents and the use of unique terms (keyword tags) in sentences to organize the relationships between technical documents (for example, legal documents; this may also be an example, or documents in other specialized fields where similar cross-references between documents are important), making it possible to pinpoint access to necessary information from all technical documents and to investigate all related information without missing anything.

FIG. 2 is a schematic diagram for explaining an example of processing by a system including a sentence creation device. FIG. 1 is a diagram illustrating an example of a schematic configuration of a system including a sentence creation device. FIG. 2 illustrates an example of a schematic configuration of a server. FIG. 2 is a diagram showing components of a sentence creation device according to the present invention. FIG. 2 is a diagram showing an interface of the sentence creation device according to the present invention. FIG. 2 is another diagram showing an interface of the sentence creation device according to the present invention. FIG. 11 is yet another diagram showing an interface of the sentence creation device according to the present invention. 2 is a flowchart showing a sentence creation method according to the present invention. FIG. 1 is a diagram showing an overview of a method for implementing the present invention. FIG. 2 illustrates a first step in the method for implementing the present invention. FIG. 4 illustrates a second step in the method for implementing the present invention. FIG. 4 illustrates a third step in the method for implementing the present invention. FIG. 4 illustrates a fourth step in the method for implementing the present invention. FIG. 4 illustrates a fifth step in the method for implementing the present invention.

Below, a sentence creation device, a sentence creation method, and a sentence creation program according to one aspect of the present disclosure will be described with reference to the drawings. However, please note that the technical scope of the present disclosure is not limited to these embodiments, but extends to the inventions described in the claims and their equivalents.

(Outline of processing by sentence creation system 1)
FIG. 1 is a schematic diagram for explaining an example of processing by a sentence creation system 1. As shown in FIG.

The text creation system 1 receives a number of technical documents from one or more computers 2, stores them in a server 3, and provides the technical documents in response to a search request from a user's terminal device 4. Technical documents may include a wide range of documents such as legal documents, technical documents, papers, and technical magazines, and are usually composed of text data. These documents may also be image data (moving images, still images) or audio data, so long as they contain information that can be converted into text data in addition to text data.

The server 3 stores a technical document database 311. The technical document database 311 stores a plurality of technical documents together with technical document identification information (technical document ID). In this embodiment, the technical document database 311 is a server, but may be located outside the sentence creation system 1.

When a user wishes to search for technical documents stored in the server 3, the user uses the terminal device 4 to send a search request to the server 3. In response to the received search request, the server 3 extracts technical documents from the technical document database 311 of the server 3 and sends them to the user's terminal device 4.

The user can input document data into the interface of the sentence creation device 100 connected to the terminal device 4, and can search for information within the document that the user wants to view, or convert the document data into document data that is set to display an explanatory text of the document contents.

The above description of FIG. 1 is merely for the purpose of deepening understanding of the contents of the present invention. Specifically, the present invention is implemented in each of the embodiments described below, and may be implemented in various modified forms without substantially exceeding the principles of the present invention. All such modified forms are included in the scope of the present invention and the disclosure of this specification. For example, in this embodiment, a configuration for searching legal documents is described, but a configuration for searching technical documents based on contracts may also be used. Furthermore, the sentences in this embodiment do not need to be interpreted as being limited to legal documents. In other words, they can be used to search for sentences that are related to each other. Furthermore, the sentences of the present invention are not limited to legal documents and can be applied to other sentences. For example, they can be applied to papers, blogs, etc.

(Overall configuration of sentence creation system 1)
FIG. 2 is a diagram showing an example of a schematic configuration of the sentence creation system 1. As shown in FIG.

The sentence creation system 1 has multiple computers 2, 2, 2..., a server 3, and a user terminal device 4. In the following, the multiple computers 2 may be simply referred to as computers 2. The computers 2, 2, 2... and the server 3 are connected to each other via a communication network such as network 5. Furthermore, the user terminal device 4 and the server 3 are connected to each other via a communication network such as network 5. The sentence creation device 100 is connected to the server 3. In addition, although one network 5 is illustrated here, if the network 5 consists of multiple networks, a gateway (not shown) may be appropriately provided between each network. A program executed by the computer 2 (e.g., a browsing program) and a program executed by the server 3 (e.g., a management program) communicate using a communication protocol such as the hypertext transfer protocol (HTTP).

(Schematic configuration of server 3)
FIG. 3 is a diagram showing an example of a schematic configuration of the server 3.

The server 3 stores a database that stores multiple technical documents. The technical document database 311 stores multiple technical documents together with technical document identification information (technical document ID). These technical documents are received by the server 3 from the computer 2 and searched by the terminal device 4. The server 3 then searches for technical documents based on search conditions input by the terminal device 4 and transmits them to the terminal device 4. To realize the above functions, the server 3 further includes a server communication unit 314, a server memory unit 315, and a server processing unit 316.

The server communication unit 314 has a communication interface circuit for connecting the server 3 to the network 5. The server communication unit 314 receives document data of technical documents from the computer 2. Furthermore, the server communication unit 314 transmits data such as search conditions received from the terminal device 4 to the server processing unit 316. The server communication unit 314 then transmits the extraction results obtained from the server processing unit 316 and the generated screen configuration to the terminal device 4. The server communication unit 314 can transmit or receive information other than the above-mentioned information to the computer 2 or the terminal device 4.

The server storage unit 315 has, for example, at least one of a semiconductor memory, a magnetic disk device, and an optical disk device, and is connected to the server 3 via a bus. The server storage unit 315 stores driver programs, operating system programs, application programs, data, and the like used in processing by the server processing unit 316. For example, the server storage unit 315 stores, as driver programs, communication device driver programs that control the server communication unit 314, and the like. Computer programs may be installed in the server storage unit 315 from computer-readable portable recording media such as CD-ROMs and DVD-ROMs using a known setup program, and the like. The server storage unit 315 also stores a specialized document database 311, a contract database 312, and the like.

The server processing unit 316 divides the sentences received from the computer 2 into chapters and sections, associates the sentences, chapters, and sections, and stores them in the server storage unit 315. Furthermore, the server processing unit 316 extracts search results in response to various search requests received from the user's terminal device 4, generates a screen configuration for transmitting the extracted search results to the user's terminal device 4, and sends the generated screen configuration to the server communication unit 314. These functions are functional modules realized by a program executed by a processor provided in the server processing unit 316. Alternatively, each of these units may be implemented in the server 3 as an independent integrated circuit, microprocessor, or firmware.

[Sentence creation device]
The sentence creation device 100 according to the present embodiment will be described with reference to the drawings. The sentence creation device 100 is an annotation device that accepts input of a technical document and performs a conversion process to add an explanation to the technical document. Adding an explanation here means providing a link to a sentence in a reference document for a reference phrase included in a sentence of the technical document, providing document data that explains technical terms and the like included in the technical document, or providing a link to another document that explains the technical term.

As shown in FIG. 4, the sentence creation device 100 includes a reception unit 10, a division unit 20, a keyword tag generation unit 30, a control unit 40, a link generation unit 70, a link output unit 80, a related document acquisition unit 90, and a related document output unit 110.

In the sentence creation device 100, technical documents are structured. Specifically, first, a first stage of structuring (hereinafter referred to as structuring level 1) is performed to structure unstructured data such as PDF information (hereinafter referred to as structuring level 0) by syntactic analysis. In this embodiment, hierarchical structures of technical documents such as chapter structures and section structures (hereinafter, chunks in the document hierarchical structure such as chapters, sections, and paragraphs will be called sections) as well as headers, annotations, etc. are extracted from unstructured data such as PDF information of technical documents. This enables browsing suitable for browsing on the web, and a table of contents linked to the main text can be displayed for all technical documents. Next, a second stage of structuring (hereinafter referred to as structuring level 2) is performed to structure the document data that has been syntactically structured in the first stage of structuring by semantic analysis. In this embodiment, three things are done. First, keyword tags (e.g., important phrases, legal names, and document names) that accurately represent the contents are assigned to each section of all legal documents, thereby describing the relationship between the technical documents themselves and the sections and the tags. This makes it possible to pinpoint and search for the technical document you want to view using keyword tags. Secondly, it extracts explicit reference relationships between technical documents. This makes it possible to follow reference links in technical documents and view cited documents. In addition, it makes it possible to suggest sections of other documents that refer to the technical document you are viewing. Thirdly, it uses the keyword tags added in the first step to estimate the relationships between the sections of all technical documents. In order to improve the accuracy of this estimation, it is possible to use the explicit reference relationships added in the second step, and in addition, it is possible to estimate the mutual relationships between keyword tags when generating keyword tags or at a time independent of their generation. This makes it possible to suggest related sections of other documents that do not explicitly reference the document you are viewing.

The reception unit 10 is a communication interface that receives document input from a user's terminal (an information processing device such as a PC, a smartphone, or a tablet terminal) via the network 5.

The document accepted by the accepting unit 10 may be a legal document or other specialized document. At this stage, the document is unstructured data and may be in the form of a PDF file, text data, or Word data. These are merely examples, and the document can be realized in other forms. This stage is structuring level 0.

The division unit 20 divides the document received by the reception unit 10 into multiple sections. As shown in the left part of FIG. 5, the data of the unstructured PDF file has a structure in which the title, the hierarchical structure of the document such as the chapter structure and the section structure, the header, the main text, and the annotations are integrated without distinction. For example, the title and the main text may be integrated, or the page number may be integrated with the text. In particular, when an annotation is included for a word or phrase in the middle of a sentence, a line break may be inserted in the middle of the sentence, and the annotation may begin. In such a case, a page break is inserted, and the continuation of the sentence is written on the next page in the middle of the sentence. In order to eliminate such inconvenience, in the sentence creation device 100 in this embodiment, the division unit 20 divides the document into multiple sections, with each section being a meaningful part that is a whole, such as the title part, the part of the sentence up to the annotation, and the annotation part. For example, as shown in the right part of FIG. 5, if there is a document with the title "Third Party Allocation", it is divided into multiple sections, such as "(f) Third Party Allocation" as one section, "Third Party Allocation is ... Transfer of Control" as one section, and "Note 17" ... Corporate Disclosure Guidelines C III (1) (4)" as one section. In other words, "Transfer of Control" and "Third Party Allocation with" that were separated by a page break are recognized as a series of sentences called "Third Party Allocation with Transfer of Control". This stage is structuring level 1. For keyword tags such as this "third party allocation", in the case of legal documents, the target is the wording of the clause, technical terms, and legal article names (for example, Article 272, paragraph 2 of the Insurance Business Act). In the case of technical terms, the target is technical terms and JIS terms published in technical books in each technical field. In the case of papers, the target is the journal name, title, author name, etc. In the case of technical journals, the target is the journal name, title, author name, technical terms and JIS terms in each technical field. By searching documents using these keyword tags, you can view the chapters and sections to which the keyword tags are attached, improving search accuracy compared to regular keyword searches.

In this method of achieving syntactic structuring, the format for recording data may be an HTML-based data format. The HTML format is highly machine-readable, making it ideal for operation and viewing on the web. However, the HTML-based data format is just one example, and other data formats such as XML may also be used. Conversion to this data format may be performed fully automatically or semi-automatically by a learning model obtained by machine learning using a supervised model, and the segmentation results obtained by syntactic structuring performed semi-automatically by this learning model may be corrected by a human, or may be performed entirely by a human.

Next, we will explain semantic structuring at structuring level 2 using Figures 9 to 14.

Figure 9 is a flowchart showing a method for realizing semantic structuring at the structuring level 2. First, a keyword tag set is generated, and based on the set, keyword tags are assigned to each section (hereinafter, referred to as the section-keyword tag relationship), and the hierarchical relationship between the keyword tags is analyzed (hereinafter, referred to as the keyword tag-keyword tag relationship). Also, a reference relationship between the sections is detected (hereinafter, referred to as the section-section explicitly existing relationship). Then, based on the section-keyword tag relationship, the relevance between the sections is analyzed (hereinafter, referred to as the section-section true relationship). Also, optionally, the relationship between the sections may be analyzed based on the section-keyword tag relationship based on the keyword tag-keyword tag relationship, or the relationship between the sections may be analyzed based on the section-keyword tag relationship based on the section-section explicitly existing relationship. The generation of the keyword tag set, the section-keyword tag relationship, and the keyword tag-keyword tag relationship may be performed step by step as independent methods, or may be performed simultaneously.

Figure 10 specifically explains a method for generating a keyword tag set. Using words from indexes of legal books, dictionaries, Wikipedia, etc., and a large amount of specialized documents such as legal books, laws, precedents, and government documents as input, a keyword tag set is output using algorithms such as Textrank and Rapid Automatic Keyword Extraction. Figure 6 shows an interface showing document data semantically structured by structuring level 2. For example, it is possible to create data that recognizes "Note 17" as a reference to an annotation and recognizes the section below it as an annotation section, or discolors the name of a legal clause and allows the user to move to another section of another document by selecting the discolored part of the legal clause name. The keyword tag generating unit 30 generates keyword tags for each of the multiple sections divided by the dividing unit 20.

Another method for generating a set of keyword tags is to consider the similarity of independent words and select the word that is most similar to the set of independent words that appear in the document (hereinafter referred to as method X). In this method, the similarity of independent words is considered based on word co-occurrence, and specifically, word2vec is used. Word2vec is expressed by a matrix of the hidden layer of a machine learning model used to predict word co-occurrence. In addition, the word that is most similar to the most words can be selected by considering the page rank of a directed graph with the cosine distance of word2vec as a weight.

A specific method for generating a set of keyword tags is to have algorithms such as Textrank and Rapid Automatic Keyword Extraction select keyphrases, which are combinations of keyword tags, from documents, and then select characteristic words from among them. More specifically, machine learning is an effective method for selecting keyphrases, and learning and prediction are performed in which the input is the entire document and the output is the keyphrase.

A specific method for generating a set of keyword tags is to represent the set of words as a directed graph, with nodes representing independent words that appear in a document and branches representing co-occurrences of words within n words, with the order of appearance being the direction, and to calculate some kind of score for each node and select the node with the highest score. The PageRank algorithm is a typical algorithm for calculating the score of each node.

Next, the keyword tag generating unit 30 generates a set of keyword tags, and outputs a tree-structured set of keyword tags using the generated set of keyword tags and a set of sections created by the dividing unit 20 by dividing the document data as input. Specifically, a keyword tag that semantically represents each section is assigned to each section using algorithms such as Textrank, Rapid Automatic Keyword Extraction, Word2Vec, and Topic Model. FIG. 7 shows a user interface in which the assigned keywords are displayed as tags under each chapter and each section. As an example, one section has an item "(f) Third Party Allocation" and under that item, "Third Party Allocation" and "Regulation" are assigned as tags, and the word "third party allocation" is highlighted in the text as shown by symbol T1, and the legal clause indicating the content of "regulation" is highlighted as shown by symbol T2.

In addition to the above specific method in which the keyword tag generator 30 generates keyword tags, other possible methods for assigning tags include a method in which keyword tags are assigned directly from the results of executing the above specific method, and a method in which a set of keyword tags with weighting for each keyword tag (using tf-idf, etc.) is generated from the results of the above specific method, and only the top n keyword tags are assigned to each section as keyword tags. In addition, there is a method in which these two methods are performed in parallel, and the union of the selected sets of keyword tags is used as the keyword tag, or the intersection of the selected sets of keyword tags is used as the keyword tag.

Figure 12 shows a method for analyzing the hierarchical relationships between keyword tags, that is, for showing the relationships between keyword tags. With this method, a set of keyword tags and a set of sections are input, and a set of keyword tags in the form of a tree is output. Analysis of the relationships between keyword tags means analyzing whether two selected keyword tags are similar to each other, and if so, the degree of similarity, and also analyzing whether the concepts indicated by the selected keyword tags are in an inclusive relationship, and if so, which one encompasses which. Specifically, there are two methods for evaluating similarity: a method that recognizes that keyword tags that are simultaneously assigned to more sentences are highly similar when the keywords are assigned, and a method that recognizes that pairs with a high degree of co-occurrence calculated when the keyword tag set is generated are highly similar. As a method of evaluating inclusiveness, for example, when considering the conditional probability that a and b co-occur when a appears in a certain range, and the conditional probability that b and a co-occur when b appears in a certain range, if the latter is greater than the former, a is considered to be a keyword tag that includes b. Similarly, if the former is greater than the latter, b can be considered to be a keyword tag that includes a.

Furthermore, using the same method as method X, if a is found to have similarities with more words than b, it is possible to analyze the inclusion relationship between keyword tags by considering a to include b. Similarly, if b is found to have similarities with more words than a, it is possible to consider b to include a. Conversely, if b is found to have similarities with more words than a, it is possible to analyze the inclusion relationship between keyword tags by considering a to include b. Similarly, if a is found to have similarities with more words than b, it is possible to consider b to include a.

Also, in the logical structure of a book's text, keyword tags that appear in lower-level chapter headings can be considered to be included in keyword tags that appear in higher-level chapter headings.

Or, for example, if a book index has two items, "prospectus" and "false statements in a prospectus," it can be inferred that the latter is attached to the former string using the particle "no" to explain a subordinate concept, and it can be seen that "false statements" is a subordinate concept to "prospectus." In this way, the index can be used to infer the relationships between keyword tags from the substring relationships between items. Similarly, the hierarchical structure of the table of contents can also be used.

Alternatively, it is also possible to analyze the dependencies of sentences and determine the hierarchy between keyword tags.

Figure 13 shows a method for detecting explicit reference relationships between sections (explicitly existing relationships between sections). The control unit 40 includes a determination unit 50 and an identification unit 60. The determination unit 50 determines whether or not multiple sections contain wording suggesting a reference. Specifically, when words such as "Note:", "See," "Example," legal provisions (for example, Article 2, paragraph 1 of the Patent Act; this is an example, and other examples such as Article 272, paragraph 2 of the Insurance Business Act) and industry jargon are included, they are recognized as a set of character strings known to be reference notations. In addition, it is determined whether or not a set of sections that is not known to be reference notations contains wording suggesting a reference.

In this method of detecting explicit reference relationships between sections (explicitly existing relationships between sections), the determination unit 50 may use pattern matching, or may model the problem as a sequence labeling problem and then use a conditional random field or a neural network, or may use a neural network.

Methods for analyzing explicit reference relationships between sections based on a large collection of sections of legal documents that are divided into sections and contain strings of characters that are not known to be references include estimating the location of the reference from the location of the reference in the PDF data of a book, etc., mechanically extracting expressions that are thought to be references by pattern matching, modeling the problem as a "sequence labeling" problem and solving it using conditional random fields, and using neural networks.

Methods for analyzing reference relationships between sections based on a set of strings that contain strings known to be reference notations include modeling the problem as a "sequence labeling" problem and solving it using conditional random fields, or using neural networks.

Sequence labeling is a problem setting in which each element that makes up a sequence, such as a string of words or a string of tokens, is labeled while taking into account the overall optimality of the sequence. Here, a set of sections with links to other sections is output from a set of sections and a set of strings. Methods for solving this sequence labeling problem include a method that uses conditional random fields and a method that uses a neural network-based method, and in some cases it may be possible to identify the link destination at the same time as a by-product.

Pattern matching is another method for analyzing explicit reference relationships. Pattern matching is a method for solving sequence labeling based on information about what expressions are likely to be used as links in books, etc. In other words, by preparing a pattern such as "If it says "See XX for details," this part is likely to be a link to XX," explicit reference relationships can be analyzed automatically, semi-automatically, or manually.

In addition, if there is no reference section candidate group that is probabilistically identified using the above-mentioned method for analyzing explicit reference relationships between sections and that has a probability higher than a threshold value of being a reference section, the candidate group is recorded in the technical document database 311, and the reference candidates are presented to the user for selection, and the selection is stored, thereby improving search accuracy.

A set of sections with links to other sections is output by estimating the location of the reference from the location of the reference notation on the PDF data of the book, etc., mechanically extracting expressions that are thought to be references by pattern matching, modeling it as a "sequence labeling" problem and solving it using conditional random fields, etc., and using neural networks. These methods may be used in combination, or any one method may be used alone.

In this method of detecting explicit reference relationships between sections (explicit relationships that exist between sections), the identification unit 60 identifies the referenced section when the determination unit 50 determines that a sentence that suggests a reference is included.

In this method of detecting explicit reference relationships between sections (explicit relationships that exist between sections), the link generation unit 70 generates a link from the document to the referenced section identified by the identification unit 60.

In this method of detecting explicit reference relationships between sections (explicit relationships that exist between sections), the link output unit 80 outputs a link to the reference destination identified by the identification unit 60.

Figure 14 shows a method for analyzing the relevance between sections (the true relationships between sections). In this analysis method, a set of sections to which keyword tags that semantically describe each section have been added, a set of sections to which links to other sections have been added, and a set of keyword tags arranged in a tree structure are input, and the relevance (similarity, etc.) for each section is output.

In this method of analyzing the relevance between sections (the true relationship between sections), the related document acquisition unit 90 acquires information indicating related documents similar to a certain document. The related document acquisition unit 90 acquires related documents from a section and a referenced section. The related document acquisition unit 90 also acquires related documents by either network structuring, which uses graph theory to infer relevance, or by similarity calculation in a low-dimensional Euclidean space based on embedding.

Specifically, sections are given as input, and each section is given links and keyword tags. One method is to roughly structure the sections into a network using links, keyword tags, section positions, and vocabulary that appears in the sections as clues, and then further refine the relationships based on the rough information in the structure. For example, if sections A and B contain links to the same section C, then first A and C, and B and C are connected in a rough network using the information that "there is a link." After constructing a rough network in this way, it is inferred that A and B are also related to some extent based on the fact that A and C and B and C are directly connected, and A and B are also connected, and so on. In computer science, this relationship network is one of the data structures called "graphs," and can be formulated and analyzed as a problem in graph theory. Another method is based on embedding, in which each section has multiple clues, such as links and keyword tags, as in the above network structuring, and these can be used to associate each section with a vector in a low-dimensional Euclidean space. This allows us to calculate section-to-section similarity using known techniques for calculating vector-to-vector similarity.

In this method of analyzing the relevance between sections (the true relationship between sections), the related document output unit 110 outputs the related documents acquired by the related document acquisition unit 90.

With this configuration, the text creation device of this embodiment electronically centralizes distributed technical document resources, allowing for cross-sectional information searches and quick access to required information. It not only makes technical documents electronically viewable, but also analyzes and organizes technical documents using natural language processing technology, etc., and analyzes the cross-reference relationships of technical documents and the use of unique terms (keyword tags) in sentences to organize the relationships between technical documents (e.g., legal documents; this is just an example, and documents in other specialized fields where similar cross-references between documents are important may also be used), making it possible to realize pinpoint access to required information from all technical documents and to conduct thorough investigations of related information.

[How to write a sentence]
The sentence creation method according to the present invention will now be described with reference to FIG.

The sentence creation method includes a receiving step S10, a dividing step S20, a keyword tag generating step S30, a determining step S40, a detecting step S41, an identifying step S42, a specifying step S50, a link generating step S60, a link output step S70, a related document acquiring step S80, and a related document output step S90. Each of these steps will be described in detail below.

In the reception step S10, the reception unit 10 is a communication interface that receives document input via communication from a user's terminal (an information processing device such as a PC, smartphone, or tablet terminal) via the network 5.

The document accepted by the accepting unit 10 may be a legal document or other specialized document. At this stage, the document is unstructured data and may be in the form of a PDF file, text data, or Word data. These are merely examples, and other forms may also be used.

In the division step S20, the division unit 20 divides the document received by the reception unit 10 into a plurality of sections. As shown in the left part of FIG. 5, the data of the unstructured PDF file has a structure in which the title, chapter structure, section structure, header, main text, and annotations are integrated without distinction. For example, when the title and the main text are integrated, or when an annotation is added to a word or phrase in the middle of a sentence, a line break may be inserted in the middle of the sentence, and the annotation may begin. In such a case, a page break is inserted, so that the continuation of the sentence is written on the next page in the middle of the sentence. In order to eliminate such inconvenience, in the sentence creation device 100 in this embodiment, the division unit 20 divides the document into a plurality of sections, with each section being a meaningful part that is a whole, such as the title part, the part of the sentence up to the annotation, and the annotation part. For example, as shown in the right part of Figure 5, if there is a document with the title "Third Party Allotment", it will be divided into multiple sections, with "(f) Third Party Allotment" as one section, "Third Party Allotment ...Transfer of Control" as one section, and "Note 17" ...Corporate Disclosure Guidelines C III (1) (4)" as one section. In other words, "Transfer of Control" and "Third Party Allotment with" that were separated by a page break are recognized as a single sentence, "Third Party Allotment with Transfer of Control". This is called syntactic structuring. By searching a document with this keyword tag, it is possible to view the chapters and sections to which the keyword tag is attached, improving search accuracy compared to normal keyword searches.

In this method of achieving syntactic structuring, the format for recording data may be an HTML-based data format. The HTML format is highly machine-readable, making it ideal for operation and viewing on the web. However, the HTML-based data format is just one example, and other data formats such as XML may also be used. Conversion to this data format may be performed fully automatically or semi-automatically by a learning model obtained by machine learning using a supervised model, and the segmentation results obtained by syntactic structuring performed semi-automatically by this learning model may be corrected by a human, or may be performed entirely by a human.

The above describes the syntactic structuring technique, and next we will explain semantic structuring as the second structuring level. As shown in FIG. 6, for example, data can be created that recognizes "Note 17" as an annotation, discolors the legal clause name with the section below it as an annotation section, and enables moving to another section of another document by selecting the discolored legal clause name. The keyword tag generation step S30 generates keyword tags for each of the multiple sections divided by the division unit 20. Specifically, a keyword tag set is generated using algorithms such as Textrank and Rapid Automatic Keyword Extraction from words in indexes, dictionaries, Wikipedia, etc. of legal books, and a large amount of specialized documents such as legal books, laws, precedents, and government documents.

Another method for generating a set of keyword tags is to consider the similarity of independent words and select the word that is most similar to the set of independent words that appear in the document (hereinafter referred to as method X). In this method, the similarity of independent words is considered based on word co-occurrence, and specifically, word2vec is used. Word2vec is expressed by a matrix of the hidden layer of a machine learning model used to predict word co-occurrence. In addition, the word that is most similar to the most words can be selected by considering the page rank of a directed graph with the cosine distance of word2vec as a weight.

Another method for generating keyword tag sets is to use algorithms such as Textrank and Rapid Automatic Keyword Extraction to select keyphrases, which are combinations of keyword tags, from documents, and then select characteristic words from among them. More specifically, machine learning is a powerful method for selecting keyphrases, which involves learning and prediction in which the input is the entire document and the output is the keyphrase. tf-idf, for example, is used to select characteristic words.

A specific method is to represent the independent words that appear in a document as nodes, co-occurrences of words within n words as branches, and a directed graph with the order of appearance as the direction, then calculate the score of the nodes and select the node with the highest score. The PageRank algorithm is a typical algorithm for calculating the score of each node.

As a specific method, the score is calculated by dividing the co-occurrence degree of an independent word by its co-occurrence frequency, and the permutation of the independent words that exists in the corpus and maximizes the score is considered. More specifically, a square matrix is considered in which the independent words that appear in the document correspond to each matrix, and each element is the number of times the corresponding word co-occurs within n words, where the co-occurrence degree is the sum of the row and column, and the co-occurrence frequency is the maximum value within the row and column.

Next, the keyword tag generation step S30 generates keyword tags from the keyword tag set generated by the keyword tag generation unit 30. Specifically, from the keyword tag set generated by the keyword tag generation unit 30 and the set of sections created by division by the division unit 20, keyword tags that semantically represent the sections are assigned to each section using algorithms such as Textrank, Rapid Automatic Keyword Extraction, Word2Vec, and Topic Model. FIG. 7 shows a user interface in which the assigned keyword tags are displayed under each chapter or section. As an example, one section has an item "(f) Third Party Allocation," and under that item, "Third Party Allocation" and "Regulation" are assigned as keyword tags, and the word "third party allocation" is highlighted in the text as indicated by the symbol T1, and the legal clause indicating the content of "regulation" is highlighted as indicated by the symbol T2.

In addition to the specific method described above in which the keyword tag generator 30 generates keyword tags, other possible methods for assigning tags include a method in which keyword tags are assigned directly from the results of executing the specific method described above, and a method in which a set of keyword tags with weighting for each keyword tag (using tf-idf, etc.) is generated from the results of the specific method described above, and only the top n keyword tags are assigned to each section as keyword tags. In addition, there is a method in which these two methods are performed in parallel, and the union of the selected sets of keyword tags is used as the keyword tag, or the intersection of the selected sets of keyword tags is used as the keyword tag.

It is also possible to analyze the hierarchical relationships between keyword tags, that is, to analyze the relationships between keyword tags. With this method, a set of keyword tags and a set of sections are input, and a set of keyword tags arranged in a tree is output. Analyzing the relationships between keyword tags means analyzing whether two selected keyword tags are similar to each other, and if so, what the degree of similarity is. It also means analyzing whether the concepts indicated by the selected keyword tags are in an inclusion relationship, and if so, which one encompasses which. Specifically, there are two methods for evaluating similarity: one in which keyword tags that are simultaneously assigned to more sentences are deemed to be highly similar when the tags are assigned, and another in which pairs with a high degree of co-occurrence calculated when the keyword tag set is generated are deemed to be highly similar. As a method of evaluating inclusiveness, for example, when considering the conditional probability that a and b co-occur when a appears in a certain range, and the conditional probability that b and a co-occur when b appears in a certain range, if the latter is greater than the former, a is considered to be a keyword tag that includes b. Similarly, if the former is greater than the latter, b can be considered to be a keyword tag that includes a.

In addition, using the same method as method X, if a is found to have more similarities with words than b, then a is considered to contain b. Similarly, if b is found to have more similarities with words than a, then b is considered to contain a. In the book structure, keyword tags that appear in lower-level chapter headings are considered to be contained in keyword tags that appear in higher-level chapter headings.

In addition, using the same method as method X, if a is found to have similarities with more words than b, it is possible to analyze the inclusion relationship between keyword tags by considering a to include b. Similarly, if b is found to have similarities with more words than a, it is possible to consider b to include a. Conversely, if b is found to have similarities with more words than a, it is possible to analyze the inclusion relationship between keyword tags by considering a to include b. Similarly, if a is found to have similarities with more words than b, it is possible to consider b to include a.

Also, in the logical structure of a book's text, keyword tags that appear in lower-level chapter headings can be considered to be included in keyword tags that appear in higher-level chapter headings.

Or, for example, if a book index has two items, "prospectus" and "false statements in a prospectus," it can be inferred that the latter is attached to the former string using the particle "no" to explain a subordinate concept, and it can be seen that "false statements" is a subordinate concept to "prospectus." In this way, the index can be used to infer the relationships between keyword tags from the substring relationships between items. Similarly, the hierarchical structure of the table of contents can also be used.

Alternatively, it is also possible to analyze the dependencies of sentences and determine the hierarchy between keyword tags.

Next, in judgment step S40, it is determined whether the multiple sections contain any wording that suggests a reference. Specifically, if the multiple sections contain words such as "Note:", "See," or "Example," or legal provisions (e.g., Article 2, paragraph 1 of the Patent Act), or industry jargon, they are recognized as a set of character strings that are known to be reference notations. In addition, it is determined whether the set of sections that are not known to be reference notations contains any wording that suggests a reference.

In the determination step S40, a detection step S41 detects text suggesting a reference using at least one of a method for solving sequence labeling problems such as pattern matching and conditional random fields, and a neural network, and an identification step S42 identifies the referenced section for each link.

Sequence labeling is a problem setting in which each element that makes up a sequence, such as a string of words or a string of tokens, is labeled while taking into account the overall optimality of the sequence. Here, a set of sections with links to other sections is output from a set of sections and a set of strings. Methods for solving this sequence labeling problem include a method that uses conditional random fields and a method that uses a neural network-based method, and in some cases it may be possible to identify the link destination at the same time as a by-product.

Pattern matching is another method for analyzing explicit reference relationships. Pattern matching is a method for solving sequence labeling based on information about what expressions are likely to be used as links in books, etc. In other words, by preparing a pattern such as "If it says "See XX for details," this part is likely to be a link to XX," explicit reference relationships can be analyzed automatically, semi-automatically, or manually.

In addition, if there is no reference section candidate group that is probabilistically identified using the above-mentioned method for analyzing explicit reference relationships between sections and that has a probability higher than a threshold value of being a reference section, the candidate group is recorded in the technical document database 311, and the reference candidates are presented to the user for selection, and the selection is stored, thereby improving search accuracy.

A set of sections with links to other sections is output by estimating the location of the reference from the location of the reference notation on the PDF data of a book, etc., mechanically extracting expressions that are thought to be references by pattern matching, modeling it as a "sequence labeling" problem and solving it using conditional random fields, etc., and using a neural network. These four methods may be used in combination, or any one method may be used alone.

In the identification step S50, if the determination unit 50 determines in the determination step S40 that the text contains a phrase that suggests a reference, the identification unit 60 identifies the referenced section.

The link generation step S60 generates a link from the document to the reference section identified by the identification unit 60.

The link output step S70 outputs a link to the reference destination identified by the identification unit 60.

In the related document acquisition step S80, the related document acquisition unit 90 generates information indicating related documents similar to a certain document. The related document acquisition unit 90 generates related documents from a section and a referenced section. The related document acquisition unit 90 also acquires related documents by either network structuring, which infers relevance using graph theory, or by similarity calculation in a low-dimensional Euclidean space based on embedding.

Specifically, sections are given as input, and each section is given links and keyword tags. One method is to roughly structure the sections into a network using links, keyword tags, section positions, and vocabulary that appears in the sections as clues, and then further refine the relationships based on the rough information in the structure. For example, if sections A and B contain links to the same section C, then first A and C, and B and C are connected in a rough network using the information that "there is a link." After constructing a rough network in this way, it is inferred that A and B are also related to some extent based on the fact that A and C and B and C are directly connected, and A and B are also connected, and so on. In computer science, this relationship network is one of the data structures called "graphs," and it is possible to formulate and solve it as a problem in graph theory. Another method is based on embedding, in which each section has multiple clues, such as links and keyword tags, as in the above network structuring, and these can be used to associate each section with a vector in a low-dimensional Euclidean space. This allows us to calculate section-to-section similarity using known techniques for calculating vector-to-vector similarity.

In the related document output step S90, the related document output unit 110 outputs the related documents acquired by the related document acquisition unit 90.

By including these steps, the text creation method of this embodiment electronically centralizes distributed technical document resources, making it possible to search for information across the board and quickly access necessary information. It not only makes technical documents available for electronic viewing, but also analyzes and organizes technical documents using natural language processing technology, etc., and analyzes the cross-reference relationships of technical documents and the way unique terms (keyword tags) are used in sentences to organize the relationships between technical documents (e.g., legal documents; this is just an example, and documents in other specialized fields where similar inter-document cross-references are important may also be used), making it possible to pinpoint access to necessary information from all technical documents and to conduct thorough investigations of related information.
[Sentence creation program]

The sentence creation program of the present invention causes a computer to realize the functions of the sentence creation device of the present invention.

The various modifications described above may be selected and combined as appropriate, or other modifications may be made.

<Additional Information>
Although not shown in the above embodiment, a function realized by using the data created by the sentence creation device may include a function of reverse-searching related documents. When the entire document or a specific part of the document being viewed is referenced by another document, there may be a document that references the entire document or a specific part of the document being viewed. Conventionally, it was impossible to find other documents that reference the document being viewed. This is because all books in the world may reference that part, so in order to comprehensively find other documents that refer to a certain document, it is necessary to check all documents in the world, which is impossible due to realistic time constraints. In the invention according to this embodiment, the explicit reference relationships between the sections of all documents are extracted by the above-mentioned semantic structuring, so other documents that reference a certain document are displayed on the product screen as "documents that reference this document," and it is possible to move to the corresponding parts of those documents. Specifically, the referenced document is similarly structured and divided into sections by the method shown in the above embodiment, and each section is tagged, and a link is placed to the related section of the referenced document. This makes it possible to refer to related parts of related literature, discover other documents that refer to a certain document, which was previously impossible, and move to a section in the referenced book that contains a reference to the document in question. For example, if a document that references the judgment of "Osaka District Court..." that you are viewing is displayed, a "Backlink Search" button can be provided in the UI shown in Figure 7, and pressing this button will display a popup that says "Documents that refer to this document." By selecting one of the books, you can move to the section where the precedent is referenced.

In addition, the functions realized in the above embodiment may include a function to correct spelling variations. Conventionally, in books, "Article 272, Paragraph 1 of the Insurance Business Law" is written as "Article 272, Paragraph 1 of the Law" by abbreviating the name of the law, and even legal terms have spelling variations such as "third party allocation", "third party allocation", and "third party allotment". In such cases, if you search for "Article 272, Paragraph 1 of the Law", all documents containing the character string will be found, and references to Article 272, Paragraph 1 of laws other than the Insurance Business Law will be extracted. In the invention according to this embodiment, when there is a reference to a legal provision or legal term, it is automatically recognized as a keyword tag by the above-mentioned semantic structuring. In addition, since the relationship between keyword tags is also extracted, for example, it is also extracted that "Article 272, Paragraph 1 of the Law" has the same meaning as "Article 272, Paragraph 1 of the Insurance Business Law", and that "third party allocation", "third party allocation", and "third party allotment" have the same meaning. The sentence creation device stores the extracted terms in this way in association with each other, absorbing variations in spelling, and when the expressions are different but indicate the same content, links can be added to refer to the same section or explanation. This makes it possible to search for all documents related to "Article 272, Paragraph 1 of the Insurance Business Act" by performing a search using a keyword tag, and all sections of documents containing sentences with keyword tags are hit and displayed as a list of search results sorted by relevance, making it possible to search for all documents related to "Article 272, Paragraph 1 of the Insurance Business Act" for example. If the results of a search using the keyword tag "Article 272 of the Insurance Business Act" are displayed as images, the section containing the sentence with the keyword tag "Chapter 5: Regulation and Supervision of Small Amount and Short-Term Insurance Companies" in "Detailed Explanation of the Insurance Business Act" will be hit. This is because the reference to "Article 272, Paragraph 1 of the Act" included in the section has been recognized in advance as the keyword tag "Article 272 of the Insurance Business Act" and the keyword tag has been added. Previously, if a book simply contained the words "Article 272 of the Law," it was not clear whether it referred to the Insurance Business Law or the Commercial Code, and even if you searched for "Article 272 of the Insurance Business Law," it would not show up in the search results. However, by recognizing in advance that the character string "Article 272 of the Law" in the book is the keyword tag "Article 272 of the Insurance Business Law" based on its context and other information, it becomes possible to display in the search results list all sections to which the keyword "Article 272 of the Insurance Business Law" has been assigned, and documents that contain these sections (including those that only contain the character string "Article 272 of the Law"), based on context and other information, for search results with the keyword tag "Article 272 of the Insurance Business Law."

The above embodiment realizes a function to display the number of hits for each keyword by section. Conventionally, when a document is searched for by keyword, which is a simple exact match search rather than a keyword tag search, the keyword is generally highlighted in the text, and the required information is searched for one by one by looking at the highlighted keywords. In the invention according to this embodiment, the above-mentioned syntactic structuring is used to display how many times the keyword is present in which section on the table of contents (TOC) displayed on the left side of the document text. This allows the user to instantly determine whether a section should be read from the wording of the table of contents and the number of hits. For example, when a search is performed in the text using two words, "demand for sale of stocks, etc., coercive nature," each keyword is highlighted in a different color in the text, making it easy to see. In addition, a number is displayed to the right of each heading in the TOC (Table of contents) displayed on the left side. This number indicates how many times the keyword is present in the page to which the heading is attached. With this function, when searching for a keyword within the text, you can instantly identify places that contain a lot of that keyword and move to those places.

The sentence creation device may also link "kosoado" words as phrases suggesting references, or link the branching destination of a condition to a condition clause such as "more than XX," "less than XX," or "other than △△." "Kosoado" words are demonstrative nouns such as this or that, and by searching for such words in a sentence and analyzing the context in that case, it is possible to identify what the demonstrative noun refers to and create a link. Similarly, for condition clauses, it is possible to identify what is referred to and create a link by identifying the content that is the source of the condition, identifying the branching destination depending on the condition in that case, and identifying which of the conditions it refers to.

Reference Signs List 1 Sentence creation system 2 Computer 3 Server 4 Terminal device 5 Network 100 Sentence creation device 10 Reception unit 20 Division unit 30 Keyword tag generation unit 40 Control unit 50 Determination unit 51 Detection unit 52 Identification unit 60 Identification unit 70 Link creation unit 80 Link output unit 90 Related document acquisition unit 110 Related document output unit

Claims (21)

a reception unit for receiving an input of a document;
a division unit for dividing the document into a plurality of sections;
a keyword tag generating unit that generates a keyword tag for each of the plurality of sections divided by the dividing unit;
a determination unit that determines whether or not the plurality of sections includes a phrase that suggests a reference;
A sentence creation device comprising: an identification unit that identifies a referenced section when a sentence suggesting the reference is included. a link generating unit that generates a link from the document to the section of the reference destination identified by the identifying unit;
a link output unit that outputs a link to the reference destination identified by the identification unit and enables movement to the reference destination by selecting the link;
The sentence creation device of claim 1 further comprising: The sentence creation device according to claim 2, characterized in that the determination unit detects the text suggesting the reference using at least one of a method for solving sequence labeling problems such as pattern matching and conditional random fields, and a neural network, and identifies the section to which the reference is directed for each of the links. a related document acquisition unit that acquires information indicating a related document that is similar to the document and indicates the content of a sentence that suggests the reference;
a related document output unit for outputting the related document;
The sentence creation device according to any one of claims 1 to 3, further comprising: The sentence creation device according to claim 4, characterized in that the related document acquisition unit acquires the related documents from the section and the referenced section. The sentence creation device according to claim 4, characterized in that the related document acquisition unit acquires the related documents from other documents different from the document. The sentence creation device according to claims 4 to 6, characterized in that the related document acquisition unit acquires the related documents by either network structuring, which uses graph theory to infer relevance, or by similarity calculation in a low-dimensional Euclidean space based on embedding. A receiving step of receiving an input of a document;
a dividing step of dividing the document into a plurality of sections;
a keyword tag generating step of generating a keyword tag for each of the plurality of sections divided by the dividing step;
determining whether the plurality of sections includes a reference-indicating phrase;
A sentence creation method comprising a step of identifying the referenced section when the sentence contains a wording that suggests the reference. a generating step of generating a link from the document to the section of the reference identified in the identifying step;
an output step of outputting a link to the reference destination identified in the identification step;
The method of claim 8 further comprising: The sentence creation method according to claim 9, characterized in that the determination step detects the phrase suggesting the reference using at least one of a method for solving sequence labeling problems such as pattern matching and conditional random fields, and a neural network, and identifies the section to which the reference is directed for each of the links. a related document acquisition step of acquiring information indicating a related document which is similar to the document and indicates the content of a sentence suggesting the reference;
a related document output step of outputting the related document;
The sentence creation method according to any one of claims 8 to 10, further comprising: The sentence creation method according to claim 11, characterized in that the related document acquisition step acquires the related documents from the section and the referenced section. The sentence creation method according to claim 11, characterized in that the related document acquisition step acquires the related document from another document different from the document. The sentence creation method according to claims 11 to 13, characterized in that the related document acquisition step acquires the related documents by either network structuring that infers relevance using graph theory, or similarity calculation in a low-dimensional Euclidean space based on embedding. On the computer,
A reception function for receiving document input;
a division function for dividing the document into a plurality of sections;
a keyword tag generating function for generating a keyword tag for each of the plurality of sections divided by the dividing function;
a determining function for determining whether the section contains a reference-indicating phrase;
A text creation program characterized by realizing a specific function of identifying a referenced section when a text suggesting a reference is included. The computer includes:
a link generating function for generating a link from the document to the referenced section identified by the identifying function;
a link output function for outputting a link to the reference destination identified by the identification function;
16. The sentence creation program according to claim 15, further comprising: The sentence creation program according to claim 16, characterized in that the determination function detects the text suggesting the reference using at least one of a method for solving sequence labeling problems such as pattern matching and conditional random fields, and a neural network, and identifies the section to which the reference is directed for each of the links. The computer includes:
a related document acquisition function for acquiring information indicating a related document which is data of a reference destination indicating the contents of a sentence similar to the document and which suggests the reference;
a related document output function for outputting the related document;
18. The sentence creation program according to claim 15, further comprising: The sentence creation program according to claim 18, characterized in that the related document acquisition function acquires the related documents from the section and the referenced section. The sentence creation program according to claim 18, characterized in that the related document acquisition function acquires the related document from a document other than the document. The sentence creation program according to claims 18 to 20, characterized in that the related document acquisition function acquires the related documents by either network structuring that infers relevance using graph theory, or similarity calculation in a low-dimensional Euclidean space based on embedding.