Summary creation method, summary creation system, and summary creation program

The method efficiently creates accurate summaries by decomposing sentences, assigning labels, and extracting summary elements from documents, enhancing understanding of complex content.

JP7578226B2Active Publication Date: 2024-11-06AIXS INC
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
JP2023135560
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2023-08-23
Publication Date
2024-11-06
Estimated Expiration
2039-05-17

AI Technical Summary

Technical Problem

Existing methods for creating summaries from documents, particularly academic papers, are inefficient and lack accuracy due to reliance on human analysis, which varies in quality, and existing automated methods fail to effectively extract the gist of complex documents without specialized knowledge.

Method used

A computer-based method that decomposes sentences into smaller units, assigns labels based on content, and extracts summary elements to create a highly accurate summary using a network of document similarity and centrality analysis.

Benefits of technology

Enables the creation of highly accurate summaries that efficiently convey the gist of documents, even for non-specialists, by leveraging AI to decompose and label sentences, improving summary accuracy and understanding.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007578226000001
    Figure 0007578226000001
  • Figure 0007578226000002
    Figure 0007578226000002
  • Figure 0007578226000003
    Figure 0007578226000003
Patent Text Reader

Abstract

To provide a summary creation method, a summary creation system and a summary creation program capable of creating a summary of high accuracy from one text.SOLUTION: In a summary creation system in which a document database, an information terminal and a server are connected through a communication network such as the Internet, an intranet, a VPN, etc., a summary creation method comprises: a text decomposition step (S6) of decomposing a text into sentences by the server; a labeling step (S7) of giving labels matching contents of the decomposed text; a summary element extraction step (S8) of extracting sentences given a predetermined label as summary elements; and a summary creation step (S9) of creating a summary based upon the summary elements.SELECTED DRAWING: Figure 5
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a method, a system, and a program for creating a summary from a text. [Background technology]

[0002] Traditionally, when analyzing a large number of academic papers, literature, and other documents, a person would read the documents and classify them by content or create a summary. Human analysis takes time, and when multiple people analyze the documents, the accuracy of the classification and summary creation tends to vary depending on the experience and knowledge of the workers.

[0003] In addition, while complex and highly specialized documents such as academic papers require advanced expertise to understand their contents, there is a demand for people without such expertise to be able to easily obtain, understand, and use the latest information.

[0004] For example, a cluster analysis method has been proposed in which morphological analysis is performed on technical documents retrieved by a concept search, weights are assigned to each word obtained from the analysis, each technical document is converted into a vector, and technical documents with vectors with similar directions are grouped together into a single cluster (for example, Patent Document 1).

[0005] In Patent Document 1, important words for each cluster are extracted and displayed in descending order of weight, allowing the gist of the cluster to be grasped. Here, a larger weight is assigned to a word that appears frequently in one cluster but appears less frequently in another cluster. However, when extracting words that appear frequently in documents that constitute a cluster in this way, noise increases, and words that indicate the gist of the cluster are not necessarily extracted. Also, if only words are listed, it is difficult for a person other than an expert in that field to understand the gist of the cluster. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2005-92443 A Summary of the Invention [Problem to be solved by the invention]

[0007] The present invention aims to provide a method, system, and program for creating a summary that can create a highly accurate summary from a single piece of text and can more efficiently understand the gist of the text. In this specification, a "document" is made up of "sentences," a "sentence" is made up of two or more "sentences," and a "sentence" is made up of two or more "words." In this specification, an "abstract" is a sentence that is placed at the beginning or end of a document and expresses the outline or characteristics of the document. In this specification, a "summary" is one or more sentences extracted from an object and expresses the outline or characteristics of the object. Here, the object may be a sentence, a document, or a cluster. [Means for solving the problem]

[0008] In other words, the present invention is a summary creation method in which a computer creates a summary from a piece of text, the summary creation method comprising: a sentence decomposition step of decomposing a sentence into sentences; a labeling step of assigning labels according to the contents of the decomposed sentences; a summary element extraction step of extracting the sentences to which predetermined labels have been assigned as summary elements; and a summary creation step of creating a summary based on the summary elements. Effect of the Invention

[0009] According to the present invention, a highly accurate summary can be created from a text, and the gist of the text can be understood more efficiently. [Brief description of the drawings]

[0010] [Figure 1]1 is a diagram showing the overall configuration of a summary creation system according to an embodiment of the present invention; [Diagram 2] FIG. 13 is an explanatory diagram showing a state in which a sentence is decomposed and labeled. [Diagram 3] 13 is a diagram showing an example of a display of a cluster analysis result and a summary creation result displayed on an output unit of an information terminal. [Figure 4] FIG. 4 is a partially enlarged schematic diagram of the cluster analysis results and summary creation results shown in FIG. 3. [Diagram 5] 4 is a flowchart showing a summary creation control routine executed by a server of the summary creation system in one embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0012] 1 is a diagram showing the overall configuration of a summary creation system according to one embodiment of the present invention, the configuration of this embodiment will be described with reference to this figure.

[0013] As shown in Fig. 1, in the summary creation system 1 according to this embodiment, a document database 2 (hereinafter, database will be abbreviated as "DB"), an information terminal 3, and a server 4 are connected via a communication network N. The communication network N is, for example, the Internet, an intranet, a VPN (Virtual Private Network), etc., and is a communication network capable of transmitting information bidirectionally using wired or wireless communication means. Also, in Fig. 1, for the sake of simplicity of explanation, one document DB 2 and one information terminal 3 are connected to one server 4, but the server 4 can be connected to multiple document DBs and multiple information terminals 3.

[0014] The document DB2 is a database that stores information on documents such as academic papers, patent documents, magazines, books, and newspaper articles, and the stored documents are made available to limited or unlimited persons. In this embodiment, the document DB2 is described as an example of a document DB that stores information on medical literature. However, there is no limit to the contents, fields, and types of documents that can be stored in the document DB of the present invention. In this embodiment, the information on medical literature includes bibliographic items such as the author's name, publication year, and the author's affiliated institution, content items of the paper such as the title, summary, and main text of the paper, citation and cited information such as the number of citations and cited documents, and publication information such as the name of the academic society, magazine, or publisher in which the document was published.

[0015] The information terminal 3 is, for example, a personal computer (hereinafter, referred to as a “PC”), a smartphone, a tablet PC, or a mobile terminal such as a mobile phone.

[0016] The output unit 10 is, for example, a device such as a display or a printer, and is capable of visually displaying the display data generated by the server 4.

[0017] The input unit 11 is a device such as a keyboard or a mouse, and is capable of inputting and operating information. The output unit 10 and the input unit 11 may be integrated together to form, for example, a touch panel.

[0018] A person (user) using the information terminal 3 can check the display data generated by the server 4 on the output unit 10 and can also issue various instructions to the server 4 via the input unit 12.

[0019] The server 4 is composed of one or more servers (computers) that break down one or more sentences contained in one or more documents and generate summaries (summary sentences). The server 4 has various calculation units and storage units, such as a document storage unit 20, a similarity calculation unit 21, a cluster classification unit 22, an index calculation unit 23, a document selection unit 24, a sentence breaker unit 25, a label assignment unit 26, a summary element extraction unit 27, a summary creation unit 28, a display data generation unit 29, and a learning system 30.

[0020] In detail, the document storage unit 20 is a storage unit that is connected to the document DB2 via the communication network N and acquires and stores information on necessary documents from the document DB2. For example, in this embodiment, medical literature is acquired from the document DB2 and stored. The document storage unit 20 also has a search function, and is capable of acquiring a population of documents to be classified into clusters by extracting documents containing specific keywords from the stored documents, or by extracting documents specified by date. The document storage unit 20 also has a function of automatically updating documents in the document storage unit 20 in synchronization with updates such as addition or deletion of documents in the document DB2.

[0021] The similarity calculation unit 21 has a function of calculating the similarity between the contents of one document and the contents of another document for a population of documents stored in the document storage unit 20 or generated documents. For example, TF-IDF or cosine similarity can be used to calculate the similarity. That is, the similarity calculation unit 21 extracts words used in the contents of each document, weights each word based on the product of the frequency of occurrence in the document (TF: Term Frequency) and the rarity of the word used in other documents (IDF: Inverse Document Frequency), and vectorizes the document. Then, the similarity calculation unit 21 calculates the cosine (cos) value between the vectorized documents as the value of the similarity between the documents. For example, the similarity between the first document and the second document is 0.856, and the similarity between the first document and the third document is 0.732. The similarity is expressed as a value between 0 and 1, and the closer to 1 the value is, the more similar the documents are.

[0022] The cluster classification unit 22 generates a network including each document connected by lines (hereinafter referred to as "edges") based on the similarity calculated by the similarity calculation unit 21, and classifies similar documents into clusters (document groups). The cluster classification algorithm is not particularly limited, but for example, an algorithm (the so-called Girvan-Newman algorithm) that specifies clusters by iterative calculations in which connectivity between nodes is maintained as much as possible even if edges are cut off can be used.

[0023] The index calculation unit 23 has a function of calculating a first index indicating the centrality of each document in the network generated by the cluster classification unit 22. The algorithm for calculating centrality is not particularly limited, and for example, eigenvector centrality, PageRank, betweenness centrality, degree centrality, etc. can be used. In this embodiment, eigenvector centrality is used. Eigenvector centrality is indicated by the probability of passing through an arbitrary node in a network when edges are repeatedly traced starting from the arbitrary node in the network.

[0024] The document selection unit 24 has a function of selecting documents having high centrality in a cluster from the document storage unit 20 using the centrality index calculated by the index calculation unit 23 .

[0025] The text decomposition unit 25 has a function of decomposing the text of the abstract contained in the document selected by the document selection unit 24 into sentences. For example, FIG. 2 illustrates a case where the text of the abstract is "This study is a cohort study on the effectiveness of drug therapy for XX disease. XX disease is an intractable disease for which no definitive treatment is known. However, in recent years, there has been remarkable progress in drug therapy to slow the progression of the disease. However, it is known that drug therapy is divided into patients for whom it is effective and patients for whom it shows little effect. Therefore, this study focuses on several physical and lifestyle factors of patients and evaluates the influence of these factors on the effectiveness of drug therapy through a five-year cohort analysis. The subjects of the analysis were five patients with XX disease, and the cohort analysis was performed according to the XX method. As a result of the analysis, the following two factors were identified as factors that affect the effectiveness of drug therapy: □□ and △△. By clarifying these previously unknown factors, this study makes it possible to predict in advance the effectiveness of drug therapy for XX disease for each patient." The sentence decomposition unit 25 decomposes the sentence into sentence units such as "This study is a cohort study on the effectiveness of drug therapy for XX disease.", "XX disease is an intractable disease for which no definitive treatment is known.", and "By clarifying these previously unknown factors, this study makes it possible to predict in advance the effectiveness of drug therapy for XX disease for each patient." The sentence decomposition unit 25 can decompose the sentence into sentences using a learning system 30 described later.

[0026] The labeling unit 26 has a function of assigning a predetermined label to each sentence decomposed by the sentence decomposition unit 25. For example, in the sentence described in FIG. 2, the labels of sentence 1 to sentence 7 are assigned. FIG. 2 shows labels that are predetermined to be assigned to the gist of a research paper, and sentences 1 to 7 correspond to the contents of "position of research," "background of research," "issues of previous research," "objective of research," "method of research," "contribution of research," and "significance of research." The labeling unit 26 can assign labels by associating the sentences decomposed by the sentence decomposition unit 25 with each label using a learning system 30 described later.

[0027] The summary element extraction unit 27 has a function of extracting, as summary elements, sentences with specific labels from among the sentences decomposed by the sentence decomposition unit 25. For example, in the sentence described in FIG. 2, the sentences labeled as sentences 4 and 7, "Therefore, this study focused on several physical and lifestyle factors of the affected patients and evaluated their effects on the effectiveness of drug therapy through a five-year cohort analysis," and "By clarifying these previously unknown factors, this study makes it possible to predict in advance the effectiveness of drug therapy for XX disease for each patient," are extracted as summary elements. Note that specific labels can be set individually depending on the field of the document.

[0028] The summary creation unit 28 creates a summary from the summary elements extracted by the summary element extraction unit 27. For example, in the document described in FIG. 2, the summary elements "In this study, we focused on several physical and lifestyle factors of the affected patient and evaluated their effects on the effectiveness of drug therapy through a five-year cohort analysis" and "By clarifying these previously unknown factors, this study makes it possible to predict the effectiveness of drug therapy for XX disease in advance for each patient" are used to create a one-sentence summary: "By evaluating the effects of factors on the effectiveness of drug therapy through cohort analysis, it becomes possible to predict the effectiveness of drug therapy for XX disease in advance for each patient." The summary creation unit 28 can create a summary from the summary elements using a learning system 30 described later.

[0029] The display data generating unit 29 has a function of generating display data showing the network by displaying each document according to its centrality index, according to the type of cluster, and according to the degree of similarity between each document. The display data generating unit 29 also has a function of generating data for displaying summaries of documents with high centrality in each cluster. The display data generating unit 29 transmits the display data, etc. to the information terminal 3 connected via the communication network N.

[0030] Fig. 3 shows an example of the cluster analysis results and summary creation results displayed on the output section of an information terminal. As shown in the figure, the cluster analysis results are displayed in area A, and the summary creation results are displayed in area B.

[0031] Specifically, as shown in FIG. 4, the display data generation unit 29 displays one document on the network (hereinafter referred to as a "node") as one circle, represents the centrality index by the size of the circle, represents the type of cluster by color, and represents the degree of similarity by the thickness of the edge.

[0032] In Fig. 4, ten nodes 40a to 40j (hereinafter collectively referred to as "nodes 40") are displayed, with the four nodes 40a to 40d at the top left belonging to a first cluster, and the six nodes 40e to 40j at the bottom right belonging to a second cluster. The first cluster and the second cluster can be shown in different colors. In Fig. 4, the difference in color is shown by different hatching.

[0033] The size of the node 40 indicates the degree of centrality, and in FIG. 4 it can be seen that nodes 40a and 40e are documents with high centrality.

[0034] Furthermore, the thickness of the edge 41 connecting the nodes 40 indicates the degree of similarity between the documents connected by the edge 41. Therefore, in Fig. 3, since the edge 41 between node 40a and node 40c and between node 40e and node 40h is thick, it can be seen that the similarity between these nodes is high.

[0035] As shown in range B, the color of each cluster and a further shortened summary of the cluster created by the summary creation unit 28 are displayed as a legend. Note that the summary displayed as the legend may be the summary created by the summary creation unit 28 itself.

[0036] The learning system 30 has a function of supplying an AI for breaking down each document into each item in the text breaking down unit 25, an AI for assigning labels according to the contents of the sentences in the label assignment unit 26, and an AI for creating a summary sentence based on the summary elements in the summary creation unit 28. In detail, the learning system 30 generates an AI (trained model) for breaking down a document into its constituent items by performing machine learning (e.g., deep learning or neural network) on data obtained by breaking down a document into sentences of each item in advance as learning data (teaching data). Then, the learning system 30 generates an AI (trained model) for assigning labels according to the contents of the broken down sentences by performing machine learning (e.g., deep learning or neural network) on data for assigning labels according to the contents of the broken down sentences as learning data (teaching data). Furthermore, the learning system 30 generates an AI (trained model) for creating a summary sentence based on the summary elements by performing machine learning (e.g., deep learning or neural network) on summary sentences using the summary elements in advance as learning data.

[0037] 5 is a flowchart showing a summary creation control routine executed by the server 4 of the summary creation system 1. The summary creation method of this embodiment will be described in detail below with reference to this flowchart.

[0038] When the server 4 receives input information such as a specific disease name from the information terminal 3, the server 4 acquires a population of documents corresponding to the input information from the document storage unit 20 in step S1.

[0039] In the next step S2, the similarity calculation unit 21 calculates the similarity between the documents that make up the population.

[0040] In step S3, the cluster classification unit 22 generates a network between documents based on the similarity calculated in step S2, and classifies the documents so that a set of similar documents constitutes a cluster.

[0041] In step S4, the index calculation unit 23 calculates an index indicating the centrality of the document in the network generated in step S3.

[0042] In step S5, the document selection unit 24 acquires documents from the document storage unit 20 based on the indexes calculated by the index calculation unit 23.

[0043] In step S6, the text dividing unit 25 divides the text contained in the acquired document into sentences.

[0044] In step S7, the labeling unit 26 assigns a label to each of the decomposed sentences.

[0045] In step S8, summary element extraction section 27 selects a label and extracts, from the decomposed sentences, sentences to which the selected label has been added, as summary elements.

[0046] In step S9, the summary creating unit 28 creates a summary based on the extracted summary elements.

[0047] In step S10, the server 4 determines whether or not the summary creation has been completed for all clusters. If the server 4 determines that the summary creation has not been completed (No), the server 4 returns the process to step S5. If the server 4 determines that the summary creation has been completed (Yes), the server 4 ends the routine.

[0048] As described above, the summary creation system 1 in this embodiment can create a highly accurate summary from a single piece of text by breaking down a sentence into sentences, assigning labels according to the content of the broken down sentences, extracting the sentences with prescribed labels as summary elements, and creating a summary based on the extracted summary elements.

[0049] Here, the summary creation method of the present invention will be compared with other methods. As another method, the automatic summarization algorithm LexRank will be described as an example. LexRank is an algorithm that breaks down multiple sentences included in a set (cluster) of related documents from multiple documents into sentences, generates a set of sentences, and uses the above-mentioned TF-IDF and cosine similarity to identify sentences that are highly similar to many sentences in the set of sentences, and extracts them as summary sentences. However, especially in academic papers, the summary of a paper expresses the outline and characteristics of the paper as a whole, and a specific sentence that constitutes the summary does not express the outline and characteristics of the entire paper. Therefore, the creation of a summary using LexRank may lack accuracy as a summary of the cluster.

[0050] In contrast, the summary creation method of the present invention forms a network based on the similarity of content between documents, classifies them into clusters composed of similar documents, extracts sentences from the clusters according to centrality indices that indicate centrality, and creates a summary from one of the extracted sentences.As a result, a summary can be created from the gist of a document that is typical or representative of the cluster, and the accuracy of the summary for the entire cluster can be improved.

[0051] In addition, because the label assignment unit 26 functions using AI that assigns labels, it can assign appropriate labels to appropriate items even if, for example, the arrangement of sentences that make up the abstract varies depending on the author of the document.

[0052] In this manner, according to the present embodiment, a highly accurate summary can be created from a sentence that has a high centrality within a cluster, and this allows a user to more efficiently understand the gist of the sentences in the cluster.

[0053] Although one embodiment of the present invention has been specifically described above, it will be understood that the present invention is not limited to this embodiment, and various changes and modifications therein can be made by those skilled in the art without departing from the scope or spirit of the present invention defined in the appended claims.

[0054] In the above embodiment, names such as sentence 1 and sentence 2 are given as labels, but the expression of the labels is not limited to this. For example, they may be expressed in hexadecimal numbers such as 0x01 and 0x02, or names may be given using letters such as the alphabet, numbers, or symbols.

[0055] In the above embodiment, the learning system 30 generates an AI that breaks down a document into items and an AI that assigns labels according to the contents of the broken down sentences, but an AI that breaks down a document into sentences and assigns labels according to the contents of each broken down sentence may be generated. This AI is generated by performing machine learning (e.g., deep learning or neural network) on data of a document that has been previously divided and labeled, as learning data (teacher data). Then, the sentence breaker and labeling unit may break down a sentence into sentences and assign labels according to the contents of the broken down sentences using an AI that assigns labels according to the contents of the broken down sentences. [Explanation of symbols]

[0056] 1. Summary creation system 2 Document DB 3. Information terminal 4 Server 10 Output section 11 Input section 20 Document storage unit 21 Similarity calculation part 22 Cluster Classification Section 23 Indicator calculation section 24 Document Selection Section 25 Sentence analysis section 26 Labeling section 27 Summary element extraction unit 28 Summary Creation Department 29 Display data generation unit 30 Learning System N Communication Network

Claims

1. A method for creating a summary from a sentence related to a research paper by a computer, comprising the steps of: A sentence decomposition step of decomposing a text into sentences; a labeling step of assigning a label according to the content of the decomposed sentences using a trained model generated by machine learning data that assigns a label according to the content of each pre-decomposed sentence from among predetermined labels including at least a label corresponding to the purpose or significance of the research, which are expected to be attached to the research paper; a summary element extraction step of extracting, as a summary element, a sentence to which a predetermined label corresponding to a purpose or significance of the research is assigned from among the predetermined labels; a summary creation step of creating a summary based on the summary elements extracted in the summary element extraction step, using a trained model generated by machine learning a summary sentence using the summary elements in advance as training data; A method for creating a summary comprising the steps of:

2. a similarity calculation step of calculating a similarity between the contents of one document and the contents of another document among the plurality of documents; a cluster classification step of generating a network including each document based on the calculated similarity and classifying similar documents into clusters; a centrality index calculation step of calculating a centrality index indicating the centrality of each document in the network; Further comprising:

2. The method according to claim 1, wherein the sentences decomposed in said sentence decomposition step are included in documents extracted from the sentences included in the clusters using a centrality index.

3. The summary element extraction step extracts a plurality of sentences that are to become summary elements, 3. The method according to claim 1, wherein said summarizing step comprises creating a one-sentence summary based on a plurality of sentences that are summary elements.

4. A method according to any one of claims 1 to 3, wherein the research paper is a medical article.

5. A summary creation system for creating a summary from a single sentence related to a research paper, comprising: a sentence decomposition unit that decomposes a sentence into sentences; a label assignment unit that assigns a label according to the content of the decomposed sentences using a trained model generated by machine learning data that assigns a label according to the content of each pre-decomposed sentence from among predetermined labels including at least a label corresponding to the purpose or significance of the research, which are assumed to be attached to the research paper; and a summary element extraction unit that extracts, from the predetermined labels, sentences to which a predetermined label corresponding to a purpose of the research or a significance of the research has been assigned as a summary element; a summary creation unit that creates a summary based on the summary elements extracted by the summary element extraction unit, using a trained model that has been generated by machine learning using summary sentences that use summary elements in advance as training data; A summary creation system comprising:

6. A program for causing a computer to create a summary from a single sentence of a research paper, comprising: A sentence decomposition step of decomposing a text into sentences; a labeling step of assigning a label according to the content of the decomposed sentences using a trained model generated by machine learning data that assigns a label according to the content of each pre-decomposed sentence from among predetermined labels including at least a label corresponding to the purpose or significance of the research, which are expected to be attached to the research paper; a summary element extraction step of extracting, as a summary element, a sentence to which a predetermined label corresponding to a purpose or significance of the research is assigned from among the predetermined labels; a summary creation step of creating a summary based on the summary elements extracted in the summary element extraction step, using a trained model generated by machine learning a summary sentence using the summary elements in advance as training data; A program that causes a computer to execute the following.

Citation Information

Patent Citations

  • Method and device for summarizing text

    JP1990297157A

  • Cluster analyzing device and cluster analyzing method

    JP2005092443A

  • Search method and device, and storage medium

    WO2015181962A1