JP2023062700A - Document analysis support system and method - Google Patents

Abstract

To provide a document analysis support system, a method, and a program, which support a user to comprehensively and accurately analyze multiple documents.SOLUTION: A document analysis support system extracts semantic structures from each of multiple documents, creates an index of semantic structures, accepts, from a user, a designation of a semantic structure condition based on the viewpoint that the user wants to know with reference to the index, or automatically designates it, displays to the user the documents that meet the designated semantic structure condition, accepts from the user selection of one or more documents and designation of classification tags that are tags to the documents, and associates the designated classification tags with the selected one or more documents. After that, the system searches for documents similar to at least one document associated with the designated classification tag from at least documents associated with no classification tag among the multiple documents, and if there is a similar document, displays it to the user and accepts document selection and classification tag designation for the similar document.SELECTED DRAWING: Figure 7

Description

The present invention relates generally to supporting document analysis.

Generally, documents contain text (strings of characters). For example, the technology disclosed in Patent Document 1 is known as a technology related to document analysis support. Also, non-patent document 1 is known at academic societies.

Japanese Patent Application Laid-Open No. 2015-88022

Wataru Sunayama, Masahiko Yachida, 2002, Observatory System for Extracting Important Sentences Based on Viewpoints and Its Implementation in a Search Engine https://www.jstage.jst.go.jp/article/tjsai/17/1/17_1_14 /_pdf

The technology disclosed in Patent Document 1 performs morphological analysis and dependency extraction for each of all documents that match the search conditions, and aggregates the documents based on the results of the morphological analysis and dependency extraction. (Paragraph 0010).

The technology disclosed in Non-Patent Document 1 is a method of presenting a word to the user that is likely to be the point of view that the user expects from the document, and searching for an important sentence based on that word.

However, with the technique disclosed in Patent Document 1, analysis accuracy is low if search conditions, morphological analysis, or dependency extraction are not appropriate. Further, the technique disclosed in Japanese Patent Laid-Open No. 2002-200022 merely aggregates and displays documents matching search conditions internally, and cannot comprehensively analyze a plurality of documents as a search range.

In addition, since the method of Non-Patent Document 1 focuses only on words, the point of view is unclear and the detection accuracy of important sentences is low.

A document analysis support system includes an indexing section and a user support section. An indexing unit extracts a semantic structure from each of a plurality of documents and creates an index of the semantic structure based on the number of cases for each extracted semantic structure. The User Support Department does the following:
(A) Receiving from the user a specification of a semantic structure condition based on the viewpoint that the user wants to know based on the index (or automatically specifying the semantic structure condition).
(B) If there is a document that matches the semantic structure conditions specified in (A) among the plurality of documents, at least a part or a summary of the document is displayed to the user.
(C) Accepting from the user selection of one or more documents from the displayed documents and specification of classification tags, which are tags for the one or more documents.
(D) Associate the classification tag specified in (C) with each of the one or more documents selected in (C).
(E) searching for a document similar to at least one document associated with a specified classification tag from among a plurality of documents not associated with at least a classification tag; Display at least part of or a summary of similar documents to the user, and perform (C).

According to the present invention, it is possible to assist a user in comprehensively and accurately analyzing a plurality of documents.

1 shows a configuration example of an entire system according to a first embodiment; 1 shows a logical configuration example of a document analysis support system according to the first embodiment; 4 shows a configuration example of a document DB according to the first embodiment; 4 shows a configuration example of a semantic structure index according to the first embodiment; An example of the flow of indexing processing is shown. A part of an example of the flow of classification support processing is shown. The rest of the example of the classification assistance process flow is shown. FIG. 7 shows an example of the UI displayed in S602 of FIG. 6. FIG. FIG. 7 shows an example of the UI displayed in S603 of FIG. 6. FIG. FIG. 7 shows an example of the UI displayed in S603 of FIG. 6. FIG. FIG. 8 shows an example of the UI displayed in S702 of FIG. 7. FIG. FIG. 8 shows an example of the UI displayed in S703 of FIG. 7. FIG. FIG. 8 shows an example of the UI displayed in S703 of FIG. 7. FIG. An example of classification tag input support is shown. Another example of classification tag input assistance is shown. An example of a semantic structure list (menu) UI is shown. An example of a UI of a list (menu) of classification tag types is shown. 8 shows a configuration example of part of a document DB according to the second embodiment; An example of a ranking list of subjects related to predicates (related to the particle "ga") is shown. An example of a ranking list in descending order of frequency of predicate-argument structures containing subjects is shown. 6 shows an example of the UI displayed in S602 of FIG. 6 in the second embodiment. 6 shows an example of the UI displayed in S603 of FIG. 6 in the second embodiment. 6 shows an example of the UI displayed in S602 of FIG. 6 in the second embodiment. 6 shows an example of the UI displayed in S603 of FIG. 6 in the second embodiment.

In the following description, an "interface device" may be one or more interface devices. The one or more interface devices may be at least one of the following:
- One or more I/O (Input/Output) interface devices. An I/O (Input/Output) interface device is an interface device for at least one of an I/O device and a remote display computer. The I/O interface device to the display computer may be a communications interface device. The at least one I/O device may be any of a user interface device, eg, an input device such as a keyboard and pointing device, and an output device such as a display device.
- One or more communication interface devices. The one or more communication interface devices may be one or more of the same type of communication interface device (e.g., one or more NICs (Network Interface Cards)) or two or more different types of communication interface devices (e.g., NIC and It may be an HBA (Host Bus Adapter).

Also, in the following description, "memory" may be one or more memory devices, typically a main memory device. At least one memory device in the memory may be a volatile memory device or a non-volatile memory device.

Also, in the following description, a "persistent storage device" is one or more persistent storage devices. A permanent storage device is typically a non-volatile storage device (for example, an auxiliary storage device), and more specifically, for example, an HDD (Hard Disk Drive) or SSD (Solid State Drive).

Also, in the following description, "storage" may be at least memory of memory and persistent storage.

Also, in the following description, a "processor" is one or more processor devices. The at least one processor device is typically a microprocessor device such as a CPU (Central Processing Unit), but may be another type of processor device such as a GPU (Graphics Processing Unit). At least one processor device may be single-core or multi-core. At least one processor device may be a processor core. At least one processor device may be a broadly defined processor device such as a hardware circuit (for example, FPGA (Field-Programmable Gate Array) or ASIC (Application Specific Integrated Circuit)) that performs part or all of processing.

In addition, in the following description, the function may be described using the expression “kkk unit”, but the function may be realized by executing one or more computer programs by a processor, or may be realized by executing one or more computer programs. It may be realized by the above hardware circuits (FPGA or ASIC, for example). When a function is realized by executing a program by a processor, the defined processing is performed using a storage device and/or an interface device as appropriate, so the function may be at least part of the processor. good. A process described with a function as the subject may be a process performed by a processor or a device having the processor. Programs may be installed from program sources. The program source may be, for example, a program distribution computer or a computer-readable recording medium (for example, a non-temporary recording medium). The description of each function is an example, and multiple functions may be combined into one function, or one function may be divided into multiple functions.

Several embodiments of the present invention will be described below with reference to the drawings.
[First embodiment]

FIG. 1 shows a configuration example of the entire system according to the first embodiment.

The document analysis support system 13 and the user system 11 of the user using the document analysis support system 13 communicate via a communication network (for example, the Internet) 160 . A “user” may be an organization such as a company, a member of the organization (e.g., an employee), or a general consumer.

The user system 11 is a physical computer system (for example, a personal computer or a smartphone), and includes, for example, an input device 113 (for example a keyboard or mouse), an output device 114 (for example a display device), an interface device 111, a storage device 112 and A processor 115 is provided. Integrated devices such as input device 113 and output device 114 may also be used. An input device 113 and an output device 114 are connected to the interface device 111 , and communication with the document analysis support system 13 is performed through the interface device 111 . A processor 115 is connected to the interface device 111 and the storage device 112 . The user system 11 may be a virtual computer system (for example, a virtual machine on a server) instead of such a physical computer system.

The document analysis support system 13 comprises an interface device 131, a storage device 132 and a processor 133 connected thereto. Communication with the user system 11 is performed via the interface device 131 . The storage device 132 stores computer programs executed by the processor 133 and data referenced or updated by the processor 133 . Processor 133 executes computer programs stored in storage device 132 . In this embodiment, the document analysis support system 13 is a physical computer system composed of one or more physical computers. infrastructure) based virtual computer system (for example, cloud computing service).

FIG. 2 shows a logical configuration example of the document analysis support system 13. As shown in FIG.

A document DB (database) 200 , a semantic structure index 201 , and a learned language model 203 are stored in the storage device 132 . The indexing unit 211, the user support unit 212, and the model learning unit 213 are implemented by the processor 133 executing the computer program. The trained language model 203 is a model trained by the model learning unit 213 (typically, a machine learning model such as a neural network).

FIG. 3 shows a configuration example of the document DB 200. As shown in FIG.

The document DB 200 is an example of information containing multiple documents. The document DB 200 has an entry for each document. Each entry contains information such as document ID 1200 , respondent attributes 1201 , document 1202 , semantic structure 1203 , classification tag 1024 and document vector 1205 . Take a document as an example.

A document ID 1200 represents the ID of the document of interest. The respondent attribute 1201 represents the attribute (for example, name, age or sex) of the respondent who has input the document of interest as an answer. A document 1202 represents the document itself (text (character string)).

The semantic structure 1203 represents a set of semantic structures (one or more semantic structures) extracted from the document. The "semantic structure" referred to in this embodiment will be described in detail later.

Classification tags 1204 represent one or more classification tags associated with the document. A "classification tag" is a tag to a document and is a tag used to classify the document. If no classification tag is associated with the document of interest, the classification tag 1204 does not contain any classification tag (indicated by "-" in FIG. 3). When a document is associated with multiple tags, the classification tags 1204 are separated by commas in the following structure.
(Example) When there are two tags. ['Ingenuity of application utilization', 'Utilization of video content']

A document vector 1205 is a value expressed by quantifying the document, and represents an N-dimensional vector of the document of interest. The document vector 1205 is used in searching for similar documents.

FIG. 4 shows a configuration example of the semantic structure index 201. As shown in FIG.

The semantic structure index 201 has an entry for each semantic structure, and each entry includes information such as a semantic structure 1300 , frequency count 1301 and document ID 1302 . Take one semantic structure as an example.

Semantic structure 1300 represents a semantic structure of interest extracted from one or more documents. The frequency number 1301 represents the number of extracted semantic structures. Document ID 1302 represents the document ID of each of the one or more documents having the semantic structure of interest.

An example of processing performed in this embodiment will be described below.

FIG. 5 shows an example of the flow of indexing processing.

The indexing unit 211 reads all documents 1202 from the document DB 200 (S500). The indexing unit 211 performs morphological analysis on each document 1202 (S501), extracts a semantic structure from the plurality of morphemes extracted in S501, and adds the extracted semantic structure to the semantic structure feature 1203 corresponding to the document. (S502).

The indexing unit 211 aggregates the frequency count (the number of semantic structures) and the document ID (the ID of the document from which the semantic structure is extracted) for each semantic structure, and creates the semantic structure index 201 based on the aggregate result. and stores the created semantic structure index 201 in the storage device 132 (S503).

In this embodiment, text as "document" includes at least one sentence or document. A "sentence" refers to one or more consecutive sentences.

In this embodiment, a "semantic structure" is a combination of semantic elements in a sentence. A "semantic element" is an element that constitutes a meaning, and the "element" referred to here may be any unit such as a morpheme, a word, or a phrase. Specifically, for example, the "semantic structure" may be a predicate-argument structure (e.g., subject-predicate combination, predicate-object combination), or a noun (e.g., content word or substance word) and a verb (e.g., function word), or a co-occurrence relationship of nouns in the same sentence (or sentence) (for example, "schedule" and "scheduler" and "meeting" and "scheduler" in the same sentence (or sentence)). If there is, it is possible to presume that it is the topic of the meeting or schedule setting with the scheduler software).

6 and 7 show an example of the flow of classification support processing.

The user support unit 212 reads the semantic structure index 201 (S600), and displays to the user a semantic structure list (for example, a list of semantic structures in descending order of frequency 1301) as at least part of the index 201. FIG. “Displaying 〇〇〇 to the user” means outputting information for displaying 〇〇〇 (for example, providing the UI (User Interface) of 〇〇〇 to the user), and this implementation In the form, as a result, 〇〇〇 is displayed on the output device 114 of the user system 11 of the user.

The order of the displayed semantic structures may be in order according to other rules instead of the descending order of the semantic structure frequency count 1301 . For example, the user support unit 212 accepts input of one or more desired semantic elements (for example, one or more words) from the user, and the semantic structure list includes sentences including the one or more semantic elements. One or more semantic structures may be arranged in descending order of suitability.

The user support unit 212 receives from the user a selection of a semantic structure desired by the user from the displayed semantic structure list (S601). The user support unit 212 searches the document DB 200 for a document having the specified semantic structure (the document 1202 corresponding to the semantic structure feature 1203 including the specified semantic structure) based on the document ID 1302 of the semantic structure index 201, and finds the document. The document obtained is displayed to the user (S602). The document displayed here can be all, part or a summary of the document.

The user support unit 212 accepts from the user a selection of one or more documents out of the documents displayed in S602 and designation of classification tags for the one or more documents, and the selected one or more documents are displayed. A designated classification tag is associated with each document (S603). This collectively associates the classification tag with one or more of the selected documents. Associating the designated classification tag with the document means that the designated classification tag is added to the classification tag 1204 corresponding to the document in this embodiment.

The user support unit 212 determines whether or not to end the classification support process (S604).

For example, if an instruction to end the classification support processing is received from the user, S604 becomes YES and the classification support processing ends.

On the other hand, for example, when an instruction to continue the classification support process (for example, an instruction to display existing classification tags) is received from the user, S604 becomes NO, and the user support unit 212 performs S700 in FIG. That is, the user support unit 212 displays to the user one or more classification tags in the column of classification tags 1204 in the document DB 200 (S700). The one or more classification tags displayed here may be the classification tags associated with the last time, the classification tags arranged in descending order of the number of associated documents, or the classification tags specified by the user. It may be a classification tag that satisfies the condition (or classification tags arranged in descending order of degree of conformity to the classification tag condition).

The user support unit 212 receives from the user a specification of one of the classification tags displayed in S700 (S701).

The user support unit 212 identifies from the document DB 200 documents associated with the classification tag designated in S701 (the document 1202 corresponding to the classification tag 1204 containing the designated classification tag), and for each identified document The document DB 200 is searched for documents similar to the document, and the found similar documents are displayed to the user (S702). The similar documents displayed here may be all, part or a summary of the similar documents. A similar document search may be performed by a known method. For example, the user support unit 212 finds, as a similar document, the document 1202 corresponding to the document vector 1205 that satisfies a predetermined condition with the document vector 1205 of the document associated with the classification tag specified in S701.

The user support unit 212 receives from the user a selection of one or more similar documents from among the similar documents displayed in S702 and designation of a classification tag for the one or more similar documents, and A designated classification tag is associated with each of the one or more similar documents (S703). As a result, one or more selected similar documents are collectively associated with the classification tag.

The user support unit 212 determines whether or not to end the classification support process (S704). For example, if an instruction to end the classification support processing is received from the user, S704 becomes YES and the classification support processing ends. On the other hand, for example, when an instruction to continue the classification support process (for example, an instruction to display an existing classification tag) is received from the user, S704 becomes NO, and the process returns to S700.

8 to 13 show examples of UIs displayed to the user in the classification support process. Note that the UI referred to here is typically a GUI (Graphical User Interface), but may be a UI other than the GUI.

FIG. 8 shows an example of the UI displayed in S602 of FIG.

Before describing the UI 400 in the state illustrated in FIG. 8, the configuration of the UI 400 will be described. In the processing part illustrated in FIG. 6, a UI 400 is displayed. The UI 400 has UI components 401-408, for example. For example, the UI components 401 to 408 are as follows.
The display column 401 displays the semantic structure specified in S601 from the semantic structure list. • The button 402 is operated to display a semantic structure list based on the semantic structure index 201 . When this button 402 is pressed, for example, a menu screen as shown in FIG. 16 is displayed and S601 is executed. As a result 401 the selected semantic structure is displayed.
• A button 403 is operated to display a document list 410 that is a list of documents containing the semantic structure displayed in the text box 401 .
A button 404 is operated to display the document in units of sentences (only one sentence is displayed in the document).
- The pull-down menu 405 is operated to display a menu (list) of classification tag types. For example, a list such as that shown in FIG. 17 is displayed.
A user-desired classification tag (specifically, a character string representing the classification tag) is entered in the text box 406 . If there is no applicable existing classification tag, the user selects "new" and directly inputs a new classification tag into the text box 406 using the keyboard or the like.
- The button 407 is operated to associate the classification tag entered in the text box 406 .
- The button 408 is operated for S604 to determine the end of the work on the UI 400 .

For example, when the button 402 is operated, S600 of FIG. 6 is performed (a semantic structure list such as that shown in FIG. 16 is displayed), and in S601 of FIG. . In this case, the user support unit 212 searches the document DB 200 (see FIG. 3) for a document 1202 corresponding to the semantic structure feature 1203 including the semantic structure "want: consideration", and in S602 of FIG. is displayed as shown in FIG. The document list 410 includes, for each found document, a check box column, at least a part (or summary) column of the document, and an existing classification tag associated with the document (classification tag corresponding to the document). each classification tag in 1204) and a column for additional classification tags additionally associated with the document. Note that the existing classification tag column and the additional classification tag column may be common. In this case, existing classification tags are displayed in the common field and new classification tags are added.

In S603 of FIG. 6, the user selects several documents as shown in FIG. ” is selected, the category tag “Problem with home network environment” is entered in the text box 406 , and finally the button 407 is operated. In this case, in response to the operation, the user support unit 212 displays "home network environment problem" in the additional tag field of the document selected by the user, as shown in FIG. A category tag “home network environment problem” is added to the category tag 1204 of the document.

After that, in the processing illustrated in FIG. 7, by designating the work end button 408, instead of the UI 400, the UI 700 illustrated in FIGS. 11 to 13 is displayed.

FIG. 11 shows an example of the UI displayed in S702 of FIG.

Before describing the UI 700 in the state illustrated in FIG. 11, the configuration of the UI 700 will be described. The UI 700 has, for example, UI components 701-708. For example, the UI components 701 to 708 are as follows.
- The pull-down menu 701 is operated to display a menu (list) of existing classification tags (classification tags in the column of classification tags 1204).
A button 702 is operated to display a similar document list 710 that is a list of documents similar to the document associated with the classification tag displayed in the pull-down menu 701 . That is, when this button 702 is pressed, for example, the similar document search in S702 is executed.
A button 703 is operated to display a list of similar document search options and specify (select) the options.
A button 704 is operated to display the similar document in units of sentences (only one sentence is displayed in the similar document).
- The pull-down menu 705 is operated to display a menu of classification tag types.
A user-desired classification tag (specifically, a character string representing the classification tag) is entered in the text box 706 .
A button 707 is operated to associate the classification tag input in the text box 706 (or the classification tag displayed in the pull-down menu 701).
- A button 708 is operated for S704 to determine whether the work on the UI 700 is finished.

For example, in S700 and S701 of FIG. 7, assume that the user selects the previously tagged category tag “home network environment problem” from the pull-down menu 701 and operates the button 702 as an existing category tag. In this case, the user support unit 212 executes S702 in FIG. 7 in response to the operation. That is, the user support unit 212 searches the document DB 200 for a document 1202 similar to the document 1202 corresponding to the classification tag 1204 including the classification tag "home network environment problem". A similar document list 710, which is a list of found similar documents, is displayed. The similar document list 710 includes, for each found similar document, a check box column, at least a part (or summary) column of the similar document, and an existing classification tag associated with the similar document (the similar document and a column for additional category tags additionally associated with the similar document. Note that the existing classification tag column and the additional classification tag column may be common. In this case, existing classification tags are displayed in the common field and new classification tags are added.

In S702 of FIG. 7, as shown in FIG. 12, the user selects several documents (puts a check mark in the check box corresponding to the document desired by the user), and selects the classification tag type "same" from the pull-down menu 706. ” is selected and the button 708 is operated. The classification tag type “same” means the same classification tag as the classification tag “home network environment problem” displayed in the pull-down menu 701 . In this case, in response to the operation, the user support unit 212 displays the column of additional tags of the similar document selected by the user (the similar document corresponding to the check box with a check mark) as shown in FIG. , and adds the classification tag ``home network environment problem'' to the classification tag 1204 of the similar document.

After that, the user selects a desired existing classification tag from the pull-down menu 701 and operates a button 702 to search for further similar documents. That is, S700 and S701 of FIG. 7 are performed again.

The above description can be summarized, for example, as follows. The following summary may include explanations of supplements and variations of the above.

The indexing unit 211 extracts a semantic structure from each of the plurality of documents 1202 and creates an index 201 of the extracted semantic structure. The user support unit 212 performs the following (A) to (E).
(A) Receive a selection of a semantic structure on the index 201 from the user.
(B) If the plurality of documents 1202 includes a document that matches the semantic structure selected in (A), at least a part or a summary of the document 202 is displayed to the user.
(C) Selecting one or more documents 1202 from the displayed documents 1202 and specifying classification tags, which are tags for the one or more documents, are received from the user.
(D) Associate the classification tag specified in (C) with each of the one or more documents selected in (C).
(E) Search for documents similar to at least one document associated with a designated classification tag from among documents not associated with at least a classification tag among the plurality of documents 1202, and if there is such a similar document, Display at least part of or a summary of the similar document to the user, and perform (C).

Accordingly, it is possible to assist the user in comprehensively and accurately analyzing the plurality of documents 1202 . Specifically, for example, it is as follows.

That is, if all the semantic structures extracted from all the documents 1202 are correct, and if the user can confirm all the semantic structures extracted from each sentence of each document and select a desired semantic structure, then the plurality of documents 1202 It is believed that exhaustive analysis will be realized. However, it is not always possible to accurately extract the semantic structure from the document due to so-called notation variations and the like. It is conceivable to express the meaning of a document by combining multiple semantic structures extracted from one document, but there are many combination patterns and it is difficult to realize. Moreover, if the user confirms all the semantic structures extracted from each sentence of each document, the user essentially sees all the documents, and the user's analysis burden is not reduced.

According to this embodiment, a document having a representative semantic structure selected by the user is searched, the document selected by the user among the found documents is associated with the classification tag desired by the user, and then the user-assigned receiving a selection of a user-desired classification tag from among the existing classification tags; searching for documents similar to the document having the selected existing classification tag; Associating user-desired classification tags with the document is repeated as necessary. In other words, it is possible to comprehensively analyze the documents using the document selected by the user among the documents found by the document search using the semantic structure as a key.

Selection of a semantic structure is an example of specifying a semantic structure condition based on the semantic structure. A "semantic structure condition" may be one semantic structure as a condition, or a combination of two or more semantic structures (for example, a combination of a semantic structure of a document to be adopted and a semantic structure of a document to be excluded). It's okay.

In addition, the present invention can be applied to support analysis of any document containing text. For example, the documents include documents as questionnaires, documents in which voice data of online meetings and radio programs are converted into text, comments on sites such as SNS (Social Networking Service), and messages between users (exchanges between users). A document or the like may be employed as a message that has been sent.

Further, according to the present embodiment, (a) displaying the semantic structure, the document and the classification tag to the user, (b) receiving selection by the user from the display in (a), (c) (b) ) is repeated as a result of processing with the semantic structure, document or taxonomy tag selected by the user in (a). That is, according to the present embodiment, documents are searched using the semantic structure selected by the user as a key, and the user associates classification tags with the found documents to add semantic elements to the documents. There is an iterative selection and retrieval and display of documents similar to the document with which the classification tag is associated. In other words, the system according to this embodiment is a system for so-called human augmentation, not a system for automating document classification.

In the UI for the user, the following may be displayed for each document 1202 to be displayed. This assists the user in document selection.
A character string containing at least semantic elements (for example, morphemes and words) that constitute the semantic structure extracted from the document 1202 .
• If a classification tag is associated with the document, the classification tag.

The user support unit 212 may determine whether or not the similar document retrieval stop condition is satisfied each time (D) is performed, and (E) may be performed when the result of the determination is false. The stopping condition may be any of the following. Such determination may be the determination in S604 of FIG. 6 or S704 of FIG. 7, for example. By providing such a stop condition, even if the number of documents 1202 is enormous, the processing can be finished at an appropriate timing.
- The proportion of documents with no classification tag associated with the plurality of documents 1202 is less than a certain proportion. In other words, it is a condition corresponding to the fact that the association of classification tags is deemed to have been sufficiently performed.
- The number of executions of (E) has reached a predetermined number.

In the similar document search, for example, when the user operates the button 703 of the UI 700 illustrated in FIG. 11, the user support unit 212 may present the following options to the user as similar document search options. (a) is the aforementioned similar document retrieval (similar document retrieval using the document vector 1205).
(a) Cosine similarity search (b) Classification model training and search (Tagged documents are used as learning data to learn a predictive classification model for document tags, and the classification model is used to classify untagged documents. Similar document search for searching documents on the street)
(c) Feature quantity option (selection of feature quantity vector calculation method used for classification) Known method such as FastText, BERT, LDA (d) Revolutional mode (classification model is updated when a classification tag is newly associated, increases the accuracy of classification predictions for tags in unknown documents)
(e) Input support for classification tags

For example, in a similar document search according to the option (e) or a document search using a user-selected semantic structure as a key, the user support unit 212 converts at least part or a summary of the classification tags to be associated with the document into the learned language. The model 203 may be used to identify and display to the user. Inputs to the trained language model 203 may include at least portions or summaries of documents. The output from the trained language model 203 may include at least some or summaries of the classification tags.

As such classification tag input support, for example, any of the following may be adopted.

For example, according to the classification tag input support illustrated in FIG. 14, the efficiency of reconfirmation is improved by automatically visualizing the basis after the classification tag is attached to the representative document.

Further, for example, according to the classification tag input assistance illustrated in FIG. 15, appropriate words to be included in the classification tag can be found.
[Second embodiment]

Features of the second embodiment are as follows.
(1) The semantic structure can be made more detailed than in the first embodiment.
(2) Notation variations can be absorbed.
(3) It is also possible to retrieve documents starting from predicates instead of or in addition to predicate terms.

A second embodiment will be described below. At that time, differences from the first embodiment will be mainly described, and descriptions of common points with the first embodiment will be omitted or simplified.

<Regarding “(1) The semantic structure can be made more detailed than in the first embodiment.”>

FIG. 18 shows a configuration example of part of the document DB 200 according to the second embodiment.

Document 1202 may be an entire document, but in this embodiment represents a sentence of the document. Therefore, although not shown, the document ID 1200 may be a combination of the ID of the document and the ID of the sentence in the document. That is, in this specification, a "document" may be an element including a set of sentences in a narrow sense, or may be a part of a set of sentences (for example, individual sentences) in a broad sense.

Semantic features 1203 differ from the configuration illustrated in FIG. Specifically, the semantic structure feature 1203 has information such as predicate11, predicate_hyouki12, negation13, pred_last14, arg_subj15, arg_obj16, arg_when17, arg_where18 and arg_other19. Take one sentence as an example (“object sentence” in the description of FIG. 18).

Predicate 11 represents the predicate (original form) in the target sentence. predicate_hyouki12 represents a predicate (notation) in the target sentence.

Negation 13 is a flag indicating whether or not the target sentence is a negative sentence. pred_last14 is a flag indicating whether or not the end of the sentence is a predicate.

Information 15 to 19 are examples of information for accurately representing the semantic structure. arg_subj15 represents the subject when the target sentence is a sentence related to the predicate by the subject. arg_obj 16 represents the object when the target sentence is a sentence related to the predicate by the object. arg_when 17 represents the time case when the target sentence is a sentence related to the predicate in the time case (time case). arg_where 18 represents the local case when the target sentence is a sentence related to the predicate by the local case (locative case). arg_other19 represents the modification when the target sentence is a sentence related to the predicate by another modification (for example, clauses such as reason and condition).

According to the example of FIG. 18, there is a dependency relationship between a predicate and its related target word (predicate term), and by analyzing the dependency relationship, elements of the semantic structure can be replaced with or added to words. can be phrases or clauses. By using such a semantic structure, it is possible to search for documents centering on predicates and target words related to the predicates.

FIG. 19 shows an example of a ranking list of subjects related to predicates (related to the particle "ga").

This list is created by the indexing unit 211 in S503 of FIG. 5, for example. According to this list, the target word related to the predicate is the subject, but a list with a similar structure is created for target words other than the subject.

The indexing unit 211 refers to the arg_subj15 column of the document DB 200, counts the frequency (number) of each subject related to the predicate, and creates a ranking list (for example, a list in descending order of subject frequency) representing the counted result. Then, for each subject, the indexing unit 211 refers to all entries having the subject as arg_subj15, and for each predicate-argument structure containing the subject and the predicate (predicate11 and/or predicate_hyouki12) related to the subject, Aggregate the frequency (number) of predicate-argument structures. FIG. 20 shows the counting result (ranking list in descending order of frequency of the predicate-argument structure) when the subject is "work burden". The indexing unit 211 registers each predicate-argument structure and its frequency in the semantic structure index 201 as a semantic structure 1300 and a frequency count 1301, and registers the ID of the document having that semantic structure in the semantic structure index 201 as a document ID 1302. (see Figure 4).

In S600 of FIG. 6, the user support unit 212 may refer to the semantic structure index 201, and in S601 of FIG. 6, the user support unit 212 may display the UI 2100 illustrated in FIG. However, at this stage, the search result of the predicate-argument structure in the UI 2100 may be blank. The UI 2100 has selection tools 2101 and 2102, a search execution button 2103, and a tagging button 2104 (elements 2101 to 2104 are examples of GUI components).

Using the selection tool 2101, a user-desired predicate term can be selected from predicate terms related to predicates (for example, subjects) in the document DB 200 (the selection tool 2102 will be described later). When the search execution button 2103 is pressed, the user support unit 212 searches the semantic structure index 201 for a predicate argument structure (semantic structure) including the predicate argument selected using the selection tool 2101, and searches for the found predicate argument. A structured document is searched from the document DB 200 . In S602 of FIG. 6, the user support unit 212 displays the found predicate-argument structures and a list of documents (search results) on the same UI 2100. FIG.

After that, when the tagging button 2104 is pressed, the user support unit 212 displays the UI 2200 illustrated in FIG. 22 for assigning classification tags in S603 of FIG. 6, for example.

According to the search results illustrated in FIG. 21, it can be seen that there is a tendency that the burden of IT-related work is disproportionately placed on some employees with high IT skills. For this reason, it is conceivable that the user may add a classification tag of "disproportionate workload to employees" to documents having this combination.

<Regarding “(2) It is possible to absorb notation variations”>

According to the search results illustrated in FIG. 21, there are scattered opinions about "employees' unbalanced workload". We would like to associate the same classification tag "employee's work burden biased" with multiple opinion sentences in one operation, but there are various spelling variations for the predicate "biased" related to "work burden".

Therefore, the user support unit 212 identifies a similar predicate-argument structure by applying rule-based or machine-learning-based processing to the found predicate-argument structure or an opinion sentence having the predicate-argument structure.

For example, rule-based processing is predicate normalization. Specifically, for example, the user support unit 212 removes the auxiliary verb from the predicate "too biased", "biased", etc. obtain. The user support unit 212 replaces the predicate in the predicate-argument structure with the original form and identifies two or more similar predicate-argument structures. At this time, the user support unit 212 removes negative forms such as "not biased" (specifically, sentences corresponding to entries with negation13 turned on).

Also, for example, machine learning-based processing is to calculate similarity of predicates in distributed representations of languages such as Word2Vec, FastText, or BERT. The user support unit 212 may identify two or more predicate-argument structures that are close to each other (the similarity difference is equal to or less than a certain value) (similarity may be estimated for the entire sentence).

Two or more predicate-argument structures (similar predicate-argument structures) specified as described above are arranged on the UI 2200 illustrated in FIG. As a result, similar predicate-argument structures are identified and arranged with high accuracy even if there is variation in the notation of predicates in the found predicate-argument structures. Therefore, the user can easily select an opinion sentence to which a desired classification tag is attached. The user enters the classification tag “employee's work burden imbalance” in the tag input field 2201 , selects the opinion sentence to which the classification tag is assigned (enters a check mark), and presses the tagging execution button 2202 . The user support unit 212 associates the input classification tag "employee's workload bias" with all the selected opinion sentences, and assigns the associated classification tags to the opinion sentences (documents) to which the classification tags are assigned. Include in the classification tag 1204 (see FIG. 3) in the entry.

<Regarding “(3) Document retrieval starting from predicates instead of or in addition to predicate terms”>

An example of document retrieval using a predicate as a starting point in place of or in addition to a predicate term such as the subject "business burden" will be described with reference to FIGS. 23 and 24. FIG.

As illustrated in FIG. 23, a selection tool 2102 of the UI 2100 can be used to select a user-desired predicate (for example, "insufficient") from the predicates in the document DB 200. FIG. When the search execution button 2103 is pressed, the user support unit 212 searches the semantic structure index 201 for a predicate-argument structure (semantic structure) including the predicate selected using the selection tool 2102, and searches for the found predicate-argument structure. The document DB 200 is searched for a document having In S602 of FIG. 6, the user support unit 212 displays the list (search result) of the found predicate-argument structures and documents on the same UI 2100. FIG.

After that, when the tagging button 2104 is pressed, the user support unit 212 displays the UI 2400 illustrated in FIG. 24 for assigning classification tags in S603 of FIG. 6, for example. Specifically, even if there is variation in the notation of a predicate term (for example, a subject) in the found predicate term structure, similar predicate term structures are identified with high accuracy and arranged on the UI 2400 . For example, the user enters the classification tag “shortage of training time” in the tag input field 2401 , selects the opinion sentence to which the classification tag is assigned (enters a check mark), and presses the tagging execution button 2402 . The user support unit 212 associates the input classification tag “insufficient training time” with all the selected opinion sentences, and sets the associated classification tag to the entry of the opinion sentence (document) to which the classification tag is assigned. in the classification tag 1204 (see FIG. 3).

The above is the description of the second embodiment. In the second embodiment, the semantic structure is a predicate-argument structure containing predicates and predicate-terms. In the second embodiment, the relationship with the first embodiment may be, for example, as follows.
- The semantic structure condition specified in (A) is a condition for one or both of the predicate and the predicate term that constitute the predicate-argument structure (for example, "want: consideration" in the first embodiment ( For example, see FIG. 8) is an example of a condition for both the predicate and the predicate term).
・In (B), a document that meets the semantic structure condition specified in (A) is a document that has a predicate-argument structure that includes a predicate or predicate-argument corresponding to the condition specified in (A).
The user support unit 212 performs rule-based or machine-learning-based processing on predicate terms or predicate terms in the predicate term structure that includes the predicate or predicate term corresponding to the condition specified in (A), and performs similar predicate Identify term structure. In (C), the displayed documents are those with similar predicate-argument structures. "Similar predicate-argument structures" may be predicate-argument structures whose similarity between predicate-argument structures is less than a certain value.

Although some embodiments of the present invention have been described above, these are examples for explanation of the present invention, and are not intended to limit the scope of the present invention only to these embodiments. The present invention can also be implemented in various other forms.

For example, in the above description, the user support unit 212 performs "(A) accepts selection of the semantic structure on the index 201 from the user." may be specified automatically without receiving input from the user. For example, the user support unit 212 accumulates a user's designation history (history of designation of semantic structure conditions), identifies semantic structure conditions that are considered important to the user from the history, forms rules, and creates semantic structures according to the rules. Kozo conditions can be automatically specified. Specifically, for example, when a history is accumulated that the user has specified mainly negative semantic structures, the user support unit 212 selects negative semantic structures (e.g., [ Decrease: Sales], [Retire: Employee], [Increase: Dissatisfaction]) may be automatically specified.

13... document analysis support system

Claims (8)

an indexing unit that extracts a semantic structure from each of a plurality of documents and creates an index of the extracted semantic structure;
a user support unit that performs the following (A) to (E); death,
(B) if among the plurality of documents, there is a document that matches the semantic structure condition specified in (A), at least a part or a summary of the document is displayed to the user;
(C) accepting from the user a selection of one or more documents from the displayed documents and designation of a classification tag, which is a tag for the one or more documents;
(D) associate each of the one or more documents selected in (C) with a classification tag specified in (C);
(E) searching for documents similar to at least one document associated with a specified classification tag from among the plurality of documents not associated with at least a classification tag, and if there is such a document, displaying at least a portion or summary of the similar document to the user, and performing (C);
document analysis support system. For each document to be displayed,
a character string including at least semantic elements constituting a semantic structure extracted from the document is to be displayed;
If a classification tag is associated with the document, the classification tag is also subject to display.
The document analysis support system according to claim 1. The user support unit
Each time (D), determine whether or not the conditions for stopping the similar document search are satisfied,
If the result of the judgment is false, perform (E),
The stopping condition is any of the following:
- The ratio of documents with no associated classification tag to the plurality of documents is less than a certain ratio.
・The number of executions of (E) has reached a predetermined number,
The document analysis support system according to claim 1. The user support unit identifies at least part of or a summary of the classification tags to be associated with the document using the learned language model and displays it to the user;
the user assistance unit displaying predicted classification tags to a user;
input to the trained language model includes at least a portion or summary of a document;
output from the trained language model includes at least a portion or summary of classification tags;
The document analysis support system according to claim 1. the semantic structure is a predicate-argument structure containing a predicate and a predicate-argument;
The semantic structure condition specified in (A) is a condition for one or both of the predicate and the predicate term that make up the predicate-argument structure,
In (B), the document that conforms to the semantic structure condition specified in (A) is a document that has a predicate-argument structure that includes a predicate or predicate-term corresponding to the condition specified in (A).
A document analysis support system according to any one of claims 1 to 4. The user support unit performs rule-based or machine-learning-based processing on predicate terms or predicate terms in a predicate term structure that includes a predicate or predicate term corresponding to the condition specified in (A), and performs similar predicate terms identify the structure,
In (C), the displayed documents are documents with similar predicate-argument structures,
The document analysis support system according to claim 5. (1) a computer extracts a semantic structure from each of a plurality of documents and creates an index of the extracted semantic structure;
(2) the computer receives from the user a designation of a semantic structure condition based on the viewpoint that the user wants to know based on the index, or automatically designates the semantic structure condition;
(3) if the computer matches the semantic structure condition specified in (2) among the plurality of documents, the computer displays the document to the user;
(4) the computer receives from the user a selection of one or more documents from the displayed documents and designation of a classification tag, which is a tag for the one or more documents;
(5) the computer associates each of the one or more documents selected in (4) with the classification tag specified in (4);
(6) The computer searches for documents similar to at least one document associated with a designated classification tag from among the plurality of documents not associated with at least a classification tag, and If so, display at least a portion or summary of the similar document to the user and perform (4);
Document analysis support method. (1) extracting a semantic structure from each of a plurality of documents and creating an index of the extracted semantic structure;
(2) receiving from the user a specification of a semantic structure condition based on a viewpoint that the user wants to know based on the index, or automatically specifying the semantic structure condition;
(3) if there is a document among the plurality of documents that matches the semantic structure condition specified in (2), the document is displayed to the user;
(4) accepting from the user a selection of one or more documents from the displayed documents and designation of a classification tag, which is a tag for the one or more documents;
(5) associate the classification tag specified in (4) with each of the one or more documents selected in (4);
(6) searching documents similar to at least one document associated with a specified classification tag from among the plurality of documents not associated with at least a classification tag; displaying at least a portion or summary of the similar document to the user, and performing (4);
A computer program that makes a computer do something.

Images