Document retrieval support system, document retrieval support method, and document retrieval support program

Abstract

To provide a document retrieval support system, a document retrieval support method, and a document retrieval support program, capable of retrieving a document while preventing an increase in cost.SOLUTION: Training data in which tag information is assigned to some document files extracted from a plurality of document files to be retrieved are acquired by a training data acquisition unit 13. A tag estimation model for estimating the tag information to be assigned to the document file is constructed by a construction unit 14 by applying the acquired training data to a Transformer machine learning model on which learning has been carried out in advance using a corpus. The tag information is assigned to each of the plurality of document files to be retrieved by an assign unit using the constructed tag estimation model.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a document retrieval support system, a document retrieval support method, and a document retrieval support program that support document retrieval. [Background technology]

[0002] In recent years, electronic documents have been used to obtain various information on the Internet, etc. Users can search for desired tag information and select and use documents that have the tag information.

[0003] For example, in the document search learning system described in Patent Document 1, content read by an electronic document reader is displayed on a content display unit. When a user selects a character string in the content by dragging it with a mouse or the like and clicks the search button, the selected character string is transferred to a search engine. This causes the Internet to be searched for documents related to the selected character string, and a list of search results is displayed on a web browser display unit. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2018-195222 Summary of the Invention [Problem to be solved by the invention]

[0005] To enable document search, tag information must be added to the documents to be searched in advance. However, because the number of documents to be searched is enormous, the burden on the workers who add tag information is heavy. Furthermore, the names (labels) of tag information are not always standardized and often vary from worker to worker. As a result, adding, maintaining, and arranging tag information incurs significant costs. Furthermore, it takes a considerable amount of time for newly created documents to become searchable.

[0006] It is possible to use AI (artificial intelligence) to estimate and automatically assign tag information to documents. However, if there is not enough training data, it is not possible to estimate tag information with high accuracy. Therefore, a large amount of training data is required to assign tag information to documents using AI. Therefore, creating a large amount of training data incurs a large cost.

[0007] An object of the present invention is to provide a document retrieval support system, a document retrieval support method, and a document retrieval support program that enable document retrieval while suppressing increases in costs. [Means for solving the problem]

[0008] (1) A document retrieval support system according to a first aspect of the present invention includes a training data acquisition unit that acquires training data to which tag information is assigned to some document files extracted from a plurality of document files to be searched; a construction unit that constructs a tag estimation model for estimating tag information to be assigned to the document files by applying the training data acquired by the training data acquisition unit to a Transformer-based machine learning model that has been trained in advance using a corpus; and an assignment unit that assigns tag information to each of the plurality of document files to be searched using the tag estimation model constructed by the construction unit. a search unit that receives an input of a character string after tag information has been assigned to each of the plurality of document files to be searched by the assigning unit, searches for a document file that matches the character string from the plurality of document files to be searched, and outputs a result screen showing the search results; Equipped with The result screen output by the search unit further displays, as tag information, tag information assigned to the document file searched by the search unit and tag information belonging to the same layer as the tag information, and the search unit searches for a document file assigned with tag information selected by a user from the tag information displayed on the result screen, and outputs a result screen showing the search results. do.

[0009] In this document retrieval support system, a Transformer-based machine learning model is trained in advance using a corpus, so a tag information estimation model that can estimate tag information with high accuracy is constructed even when the amount of training data is small. In this case, there is no need to acquire a large amount of training data. Therefore, by using a tag information estimation model constructed using only some, but not all, of the multiple document files to be searched as training data, tag information can be assigned to each of the multiple document files to be searched.

[0010] With this configuration, users do not need to assign tag information to a huge number of document files. Furthermore, there is no variation in the expression of the labels in the tag information. Therefore, there is almost no cost involved in assigning, maintaining, or arranging tag information. This allows documents to be searched while suppressing increases in costs. Furthermore, since tag information is immediately assigned to newly created document files, newly created document files can also be searched immediately. The document search support system further includes a search unit that, after tag information is assigned to each of a plurality of document files to be searched by the assigning unit, accepts input of a character string, searches the plurality of document files to be searched for document files that match the character string, and outputs a results screen showing the search results. In this case, a user can easily use document files that match the character string by inputting a desired character string. The result screen output by the search unit further displays accessible tag information related to the searched document files, and the search unit searches for document files assigned with tag information selected by the user from the tag information displayed on the result screen and outputs a result screen showing the search results. In this case, the user can easily use the document file assigned with the tag information by selecting the desired tag information displayed on the result screen. (7) The accessible tag information includes tag information assigned to the document file searched by the search unit and candidate tag information belonging to the same hierarchical level as the tag information. In this case, the accessible tag information can be easily displayed on the result screen.

[0011] (2) The training data may indicate the relationship between the content of the document file, which is an explanatory variable, and the tag information assigned to the document file, which is a target variable, and the tag estimation model may be constructed by a Transformer-based machine learning model learning the relationship between the content of the document file and the tag information assigned to the document file based on the training data. In this case, the tag estimation model can be easily constructed.

[0012] (3) The Transformer-based machine learning model may be BERT (Bidirectional Encoder Representations from Transformers), which allows the machine learning model to easily learn using a corpus.

[0013] (4) The document search support system may further include a creation unit that creates a tag information list showing candidate tag information to be assigned to the multiple document files to be searched, and an extraction unit that extracts some document files from the multiple document files to be searched, and the training data acquisition unit may generate training data by assigning any tag information in the tag information list created by the creation unit to the some document files extracted by the extraction unit. In this case, training data can be easily generated from the some document files extracted from the multiple document files to be searched.

[0017] (5 The document search support system further includes a receiving unit that receives an instruction to edit tag information assigned to one of a plurality of document files to be searched, and an editing unit that edits the tag information of the document file instructed by the receiving unit and edits the tag information of other document files to be searched, and the assigning unit may update the tag information assigned to the document file to the tag information edited by the editing unit. With this configuration, even if inappropriate tag information is assigned to a document file, the tag information assigned to the document file can be appropriately corrected.

[0018] ( 6 )The editorial department, Estimated probability of estimated tag information The tag information of the document files to be searched may be selectively edited based on the threshold value. In this case, the tag information of the plurality of document files can be selectively edited collectively based on the threshold value.

[0019] ( 7 The result screen output by the search unit may further display a GUI (Graphical User Interface) that is operated by the user, and the reception unit may receive an instruction to edit tag information assigned to one of the multiple document files that are the search targets by operating the GUI. In this case, the user can easily instruct editing of tag information assigned to a desired document file.

[0020] ( 8 The editing unit may update the tag estimation model constructed by the construction unit based on the editing results. In this case, it becomes easier to assign appropriate tag information to document files.

[0021] ( 9A document retrieval support method according to a second aspect of the present invention includes: acquiring training data to which tag information is assigned to some document files extracted from a plurality of document files to be searched; constructing a tag estimation model for estimating tag information to be assigned to the document files by applying the acquired training data to a Transformer-based machine learning model that has been trained in advance using a corpus; and assigning tag information to each of the plurality of document files to be searched using the constructed tag estimation model. After tag information is assigned to each of the plurality of document files to be searched, an input of a character string is accepted, and a document file matching the character string is searched from the plurality of document files to be searched is searched, and a result screen is output on which the search results, the tag information assigned to the searched document files, and candidate tag information belonging to the same layer as the tag information are displayed; and a document file assigned with tag information selected by a user from the tag information displayed on the result screen is searched, and a result screen is output on which the search results are displayed. Includes.

[0022] This document search support method eliminates the need to acquire large amounts of training data. Furthermore, users do not need to assign tag information to a huge number of document files. Furthermore, there is no fluctuation in expression in the labels of tag information. Therefore, there is almost no cost involved in assigning, maintaining, or arranging tag information. This makes it possible to search for documents while suppressing increases in costs. Furthermore, since tag information is immediately assigned to newly created document files, newly created document files can also be searched immediately.

[0023] ( 10 A document retrieval support program according to a third aspect of the present invention is a document retrieval support program executable by a processing device, the document retrieval support program comprising: a process of acquiring training data to which tag information is assigned to some document files extracted from a plurality of document files to be searched; a process of constructing a tag estimation model for estimating tag information to be assigned to the document files by applying the acquired training data to a Transformer-based machine learning model that has been trained in advance using a corpus; and a process of assigning tag information to each of the plurality of document files to be searched using the constructed tag estimation model. a process of receiving an input of a character string after tag information has been assigned to each of a plurality of document files to be searched, searching for document files that match the character string from the plurality of document files to be searched, and outputting a result screen that displays the search results, tag information assigned to the searched document files, and candidate tag information that belongs to the same hierarchical level as the tag information; and a process of searching for document files that have been assigned tag information selected by a user from the tag information displayed on the result screen, and outputting a result screen that shows the search results. The processing device executes the above.

[0024] This document search support program eliminates the need to acquire large amounts of training data. Furthermore, users do not need to assign tag information to a huge number of document files. Furthermore, there is no fluctuation in expression in the labels of tag information. Therefore, there is almost no cost involved in assigning, maintaining, or arranging tag information. This makes it possible to search for documents while suppressing increases in costs. Furthermore, since tag information is immediately assigned to newly created document files, newly created document files can also be searched immediately. [Effects of the Invention]

[0025] According to the present invention, it is possible to search for documents while suppressing increases in costs. [Brief explanation of the drawings]

[0026] [Figure 1] 1 is a diagram showing a configuration of a document search support system according to an embodiment of the present invention; [Figure 2] FIG. 2 is a diagram illustrating the configuration of a tag estimation device and a search device in FIG. [Figure 3] 3 is a diagram showing a tag information list created by the creation unit of FIG. 2. FIG. [Figure 4] 3 is a diagram showing an example of a method for specifying a document file to be extracted by the extraction unit in FIG. 2. FIG. [Figure 5] 3 is a diagram showing a table including document files extracted by the extraction unit of FIG. 2. FIG. [Figure 6] FIG. 10 is a diagram illustrating generated teacher data. [Figure 7] FIG. 10 is a diagram illustrating an example of creating training data having a multi-label structure. [Figure 8] FIG. 10 is a diagram illustrating an example of creating training data having a multi-label structure. [Figure 9] FIG. 10 is a diagram illustrating an example of a search screen. [Figure 10] FIG. 10 is a diagram showing an example of a result screen. [Figure 11] FIG. 10 is a diagram illustrating an example of an editing screen. [Figure 12] FIG. 10 is a diagram showing a first display example of a file column on a result screen. [Figure 13] FIG. 10 is a diagram showing a second display example of the file column on the result screen. [Figure 14] FIG. 10 is a diagram showing a third display example of the file column on the result screen. [Figure 15] FIG. 10 is a diagram showing a third display example of the file column on the result screen. [Figure 16] FIG. 10 is a diagram showing a fourth display example of the file column on the result screen. [Figure 17] FIG. 10 is a diagram showing a fifth display example of the file column on the result screen. [Figure 18] FIG. 10 is a diagram illustrating an example of an editing screen when training data has a multi-label structure. [Figure 19] 3 is a flowchart showing a tag estimation process performed by the tag estimation device of FIG. 2. [Figure 20] 3 is a flowchart showing a search process performed by the search device of FIG. 2. [Figure 21] 3 is a flowchart showing a search process performed by the search device of FIG. 2. DETAILED DESCRIPTION OF THE INVENTION

[0027] (1) Structure of the document search support system A document retrieval support system, a document retrieval support method, and a document retrieval support program according to embodiments of the present invention will be described below with reference to the accompanying drawings. Fig. 1 is a diagram showing the configuration of a document retrieval support system according to an embodiment of the present invention. As shown in Fig. 1, the document retrieval support system 100 is realized by an information processing device such as a personal computer, and includes a RAM (random access memory) 110, a ROM (read only memory) 120, a CPU (central processing unit) 130, a storage unit 140, an operation unit 150, a display unit 160, and a database storage device 170.

[0028] The RAM 110, ROM 120, CPU 130, storage unit 140, operation unit 150, display unit 160, and database storage device 170 are connected to a bus 180. The RAM 110, ROM 120, and CPU 130 constitute a tag estimation device 10 and a search device 20. In this embodiment, the tag estimation device 10 and the search device 20 are configured as a common document retrieval support system 100, but the embodiment is not limited to this. The tag estimation device 10 and the search device 20 may be configured as separate document retrieval support systems 100 as long as they are connectable to each other.

[0029] The RAM 110 is made of, for example, a volatile memory and is used as a work area for the CPU 130. The ROM 120 is made of, for example, a non-volatile memory and stores a document search support program. The document search support program includes a tag estimation program and a search program. The CPU 130 executes the tag estimation program stored in the ROM 120 on the RAM 110 to perform tag estimation processing as the tag estimation device 10. The CPU 130 also executes the search program stored in the ROM 120 on the RAM 110 to perform search processing as the search device 20. The tag estimation processing and search processing will be described in detail later.

[0030] Document search assistance The program may be stored in the storage unit 140 instead of the ROM 120. Document search assistance The program may be provided in a form stored in a computer-readable storage medium and installed in the ROM 120 or the storage unit 140. Alternatively, if the document retrieval support system 100 is connected to a network such as the Internet, the program may be distributed from a server (including a cloud server) on the network. Document search assistance The program may be installed in the ROM 120 or the storage unit 140 .

[0031] The storage unit 140 includes a storage medium such as a hard disk, a semiconductor memory, an optical disk, a magnetic disk, or a memory card, and stores a Transformer-based machine learning model for estimating tag information to be added to an arbitrary document file. The Transformer-based machine learning model is trained in advance using a large-scale corpus. In this example, the Transformer-based machine learning model is BERT (Bidirectional Encoder Representations from Transformers).

[0032] The operation unit 150 includes, for example, a pointing device such as a mouse, a keyboard, or a touch panel. By operating the operation unit 150, the user can make predetermined selections or give instructions to the tag estimation device 10 or the search device 20. The display unit 160 includes, for example, a liquid crystal display, and displays a predetermined GUI (Graphical User Interface) or the results of the search process. The database storage device 170 includes a large-capacity storage device such as a server, and stores in advance document files indicating a large number of documents to be searched.

[0033] (2) Configuration of tag estimation device Fig. 2 is a diagram showing the configuration of the tag estimation device 10 and the search device 20 of Fig. 1. As shown in Fig. 2, the tag estimation device 10 includes, as functional units, a creation unit 11, an extraction unit 12, a teacher data acquisition unit 13, a construction unit 14, an estimation unit 15, and an editing unit 16. The functional units of the tag estimation device 10 are realized by the CPU 130 of Fig. 1 executing a tag estimation program stored in the ROM 120 or the like. Some or all of the functional units of the tag estimation device 10 may be realized by hardware such as electronic circuits.

[0034] In response to an operation of the operation unit 150, the creation unit 11 creates a list of candidates for tag information to be assigned to the document file to be searched (referred to as a tag information list). In the tableIn the tag information table, candidates for tag information are arranged in a hierarchical structure. The user can use the operation unit 150 to perform an operation for creating a tag information list.

[0035] In response to an operation of the operation unit 150, the extraction unit 12 extracts some of the document files to be used for generating teacher data from the many document files stored in the database storage device 170. The user can use the operation unit 150 to specify some of the document files to be used for generating teacher data.

[0036] The teacher data acquisition unit 13 acquires the teacher data created by the creation unit 11 for each document file extracted by the extraction unit 12. Tag information list By assigning any of the tag information in the above, each document file is generated as training data to be used in constructing a tag estimation model, which will be described later. The training data indicates the relationship between the content of the document file, which is an explanatory variable, and the tag information assigned to the document file, which is a target variable.

[0037] The construction unit 14 acquires the BERT stored in the storage unit 140. The construction unit 14 also causes the BERT to learn the relationship between the contents of the document file and the tag information assigned to the document file, using a plurality of training data generated by the training data acquisition unit 13. In this way, the BERT is constructed as a tag estimation model.

[0038] The estimation unit 15 estimates tag information of a document file specified by the search device 20 using the tag estimation model constructed by the construction unit 14, and transmits the estimation result to the search device 20. Estimation part 15 evaluates the estimation probability (reliability) of the estimated tag information. As mentioned above, BERT is trained in advance using a large corpus, so it is possible to estimate tag information with high accuracy even when the number of training data is relatively small. In this example, the number of training data per tag information is approximately 10.

[0039] When the editing unit 16 receives a request from the search device 20 to edit tag information assigned to one of the document files, it creates an edited file F1 by editing the tag information assigned to the document file. The editing unit 16 also edits tag information for other document files (pre-edited files F2) to be searched, in the same manner as for the edited file F1. The editing unit 16 can selectively and collectively edit tag information for the pre-edited file F2 based on a threshold value for estimation probability. The editing unit 16 transmits the edited tag information to the search device 20 and updates the tag estimation model constructed by the construction unit 14 based on the editing results.

[0040] Maintenance of tag information may result in the discovery of tag information that is no longer in use or has a changed name. Furthermore, the estimation probability of tag information estimated by the estimation unit 15 is not always 100%, and incorrect tag information may be estimated. Furthermore, there may be cases where new tag information that has never existed before is desired to be assigned. Even in such cases, the editing unit 16 can edit the tag information to assign appropriate tag information to a document. Furthermore, since there is no need to retrain the tag estimation model to estimate new tag information, downtime can be prevented.

[0041] (3) Configuration of the search device 2, search device 20 includes, as functional units, an index acquisition unit 21, a registration unit 22, a file acquisition unit 23, a search unit 24, and a reception unit 25. The functional units of search device 20 are realized by CPU 130 in Fig. 1 executing a search program stored in ROM 120 or the like. Some or all of the functional units of search device 20 may be realized by hardware such as electronic circuits.

[0042] The index acquisition unit 21 acquires index information for each document file by crawling the database storage device 170. The index information includes the path, name, creator, creation date, content, etc. of the document file stored in the database storage device 170. Every time a new document file is added to the database storage device 170, the index acquisition unit 21 further acquires index information for the added document file. The registration unit 22 registers the index information acquired by the index acquisition unit 21 in the search unit 24.

[0043] As will be described later, the registration unit 22 receives tag information estimated for each document file by the estimation unit 15 of the tag estimation device 10. The registration unit 22 further registers the tag information received for each document file from the estimation unit 15 in the search unit 24 as part of the index information for that document file. This assigns tag information to the document file, making it possible to search for that document file. Therefore, the registration unit 22 functions as an assigning unit that assigns tag information to document files.

[0044] Furthermore, as will be described later, the registration unit 22 receives tag information for each document file edited by the editing unit 16 of the tag estimation device 10. The registration unit 22 updates the tag information registered in the search unit 24 to the edited tag information received from the editing unit 16.

[0045] The file acquisition unit 23 acquires the document file to be searched based on the index information excluding the tag information registered in the search unit 24, and specifies the acquired document file to the tag estimation device 10. In this case, the tag information of the specified document file is estimated by the estimation unit 15 of the tag estimation device 10. The tag information estimated by the estimation unit 15 is transmitted to the registration unit 22 as described above.

[0046] After the tag information of each document file is registered by the registration unit 22, the search unit 24 displays a predetermined search screen on the display unit 160. The user can input a desired character string into the search screen by operating the operation unit 150. The search unit 24 searches for document files to which tag information matching the character string input into the search screen has been assigned, and displays a result screen showing the search results on the display unit 160.

[0047] The accepting unit 25 accepts an instruction to edit tag information from the result screen displayed on the display unit 160. The user can issue an instruction to edit the tag information assigned to a desired document file on the result screen by operating the operation unit 150. Here, in the initial setting, the threshold value of the estimated probability used to edit the tag information is a default value, but the user can also change the setting of the threshold value of the estimated probability by operating the operation unit 150.

[0048] When the receiving unit 25 receives an instruction to edit tag information, it requests the tag estimation device 10 to edit the tag information assigned to the document file. In this case, the editing unit 16 of the tag estimation device 10 edits the tag information assigned to the document file, and also edits tag information assigned to other document files in the same manner. The tag information edited by the editing unit 16 is transmitted to the registration unit 22 as described above.

[0049] (4) Example of operation of tag estimation device Fig. 3 is a diagram showing a tag information list created by the creation unit 11 in Fig. 2. As shown in Fig. 3, the creation unit 11 creates a tag information list 101 using a table displayed on the display unit 160 based on the operation of the operation unit 150 by the user.

[0050] In the example of tag information list 101 in FIG. 3, the tag information includes items such as "industry," "theme," "method," and "file type." Labels for the item "industry" include "food," "information," and "automobile." Labels for the item "theme" include "fermented food," "service," and "autonomous driving." Labels for the item "method" include "yeast," "SaaS," "PaaS," "MaaS," and "high-precision GPS." Labels for the item "file type" include "research report" and "design document."

[0051] 4 is a diagram showing an example of a method for specifying a document file to be extracted by extraction unit 12 of FIG. 2. As shown in FIG. 4, a document file is extracted by specifying a path to a destination to save it on display unit 160. In the example of FIG. 4, the user operates operation unit 150 to specify "Folder CCC" in "Folder BBB" in "Folder AAA" in "C Drive". The user also operates operation unit 150 to specify "Folder WWW" and "Folder ZZZ" in "Folder YYY" in "Folder XXX" in "C Drive".

[0052] In this case, the extraction unit 12 extracts the document files "aaa.txt" and "bbb.pptx" saved in the "folder CCC." The extraction unit 12 also extracts the document file "xxx.pdf" saved in the "folder WWW." The extraction unit 12 also extracts the document files "yyy.docx" and "zzz.pptx" saved in the "folder ZZZ."

[0053] The paths specified in FIG. 4 and the extracted document files are displayed in table form on the display unit 160 in association with the tag information list 101 created in FIG. 3. FIG. 5 is a diagram showing a table including document files extracted by the extraction unit 12 of FIG. 2. As shown in FIG. 5, the user operates the operation unit 150 to select labels for each item corresponding to each document file in the table 102 from the labels included in the tag information list 101. In the example of FIG. 5, the labels for each item are selected from a pull-down menu corresponding to the tag information list 101. Some of the labels may be selected automatically.

[0054] The labels of the selected tag information are assigned to each document file by the training data acquisition unit 13 in FIG. 2. As a result, each document file is generated as training data. FIG. 6 is a diagram showing the generated training data. In the example of FIG. 6, five training data are generated from each of the five extracted document files. The training data "xxx.pdf" is assigned the labels "food," "fermented food," "fermented food," and "research report."

[0055] Similarly, the training data "yyy.docx" is assigned the labels "information," "services," "SaaS," and "design document." The training data "zzz.pptx" is assigned the labels "information," "services," "PaaS," and "design document." The training data "aaa.txt" is assigned the labels "automobile," "services," "MaaS," and "investigation report." The training data "bbb.pptx" is assigned the labels "automobile," "autonomous driving," "high-precision GPS," and "design document."

[0056] It is also possible to select multiple labels (multi-labels) for one item. Figures 7 and 8 are diagrams for explaining an example of creating training data with a multi-label structure. On the left side of Figure 7, the structure of "industry" classified under "industry" is shown in a tree format. Similarly, on the right side of Figure 7, the structure of "accident details" classified under "work-related accidents" is shown in a tree format.

[0057] As shown in FIG. 7, the "Industry" of the label "Other" in the item "Industry" is further classified into "Automotive Maintenance," "Drying," or "Machine Repair." Furthermore, the "Accident Details" of the label "Fall" in the item "Accident Details" is further classified into "Slip," "Stumble," or "Misstep." When creating training data having the multi-label structure of FIG. 7, table 103 of FIG. 8 is displayed on display unit 160 instead of table 102 of FIG. 5.

[0058] In table 103 of FIG. Based on Tag Information Table 101 Based on the created creation unit 11, one or more labels that can be assigned to each item of the document file extracted by the extraction unit 12 are displayed corresponding to the document file. The user operates the operation unit 150 to select one or more labels for each item corresponding to each document file on the table 103. In the example of FIG. 8, a "◯" is added to the box in the table 103 corresponding to the selected label. With this configuration, it is possible to assign multiple labels to each item corresponding to each document file. This makes it possible to create training data with a multi-label structure.

[0059] (5) Example of search device operation After the tag information of each document file is registered as part of the index information in the search unit 24 of FIG. 2, a search screen is displayed on the display unit 160 by the search unit 24. FIG. 9 is a diagram showing an example of the search screen. As shown in FIG. 9, the search screen 30 includes an input field 31. The user can input a desired character string into the input field 31 by operating the operation unit 150 of FIG. 2. The character string may be a word or a sentence.

[0060] When a character string is entered in the input field 31, the search unit 24 searches for document files to which tag information matching the character string has been assigned, and a results screen showing the search results is displayed on the display unit 160. In the example of FIG. 9, the word "work-related accident" has been entered in the input field 31. FIG. 10 is a diagram showing an example of the results screen. As shown in FIG. 10, the results screen 40 includes an input field 41, a file field 42, a tag search field 43, and buttons 44 and 45.

[0061] The input field 41 has the same function as the input field 31 on the search screen 30. The user can also input a desired character string into the input field 41. Therefore, the search unit 24 may display the result screen 40 on the display unit 160 instead of the search screen 30 after the tag information of each document file is registered as part of the index information.

[0062] The file field 42 displays the beginning of the document based on the searched document file, the path of the document file, index information (including tag information), etc. Also displayed in the file field 42 is an edit icon 46. The edit icon 46 is a GUI that is operated to accept editing of tag information. When multiple document files are searched, multiple file fields 42 are arranged in a vertical direction, corresponding to the multiple document files, respectively.

[0063] The tag search field 43 displays accessible tag information related to the searched document file in a tree structure. The accessible tag information includes tag information assigned to the searched document file and tag information belonging to the same hierarchy as the tag information. A check box is displayed to the left of each tag information, and the number of matching tag information items is displayed in parentheses to the right of each tag information item. Tag information items with a matching number of zero items are displayed grayed out.

[0064] When the user operates operation unit 150 to check a check box corresponding to desired tag information and operates button 44, only the tag information corresponding to the check box is displayed in tag search field 43. When the user operates operation unit 150 to operate button 45, all accessible tag information is displayed in tag search field 43.

[0065] Furthermore, when the user operates the operation unit 150 to select desired tag information, the search unit 24 searches for document files to which the tag information has been added, and a result screen showing the search results is displayed on the display unit 160. Therefore, the user can search for document files by selecting only tag information.

[0066] Furthermore, the user operates the operation unit 150 to operate the edit icon 46 in the desired file field 42, thereby selecting the document file corresponding to that file field 42. An edit screen for editing tag information assigned to the selected document file is displayed as a pop-up screen. FIG. 11 is a diagram showing an example of the edit screen. As shown in FIG. 11, the edit screen 50 includes a tag edit field 51 and a button 52.

[0067] The tag editing field 51 displays tag information assigned to the selected document file for each item hierarchy. Also, check boxes 53 are displayed in the tag editing field 51. If multiple pieces of tag information are assigned to the selected document file, multiple tag editing fields 51 are arranged vertically, one for each piece of tag information.

[0068] When the user operates the operation unit 150 to uncheck the check box 53 in the desired tag editing field 51 and then operates the button 52, the tag information corresponding to the check box 53 is deleted from the selected document file. On the other hand, when the user operates the operation unit 150 to check the check box 53 in the desired tag editing field 51 and then operates the button 52, the tag information corresponding to the check box 53 is added to the selected document file.

[0069] Furthermore, the user operates the operation unit 150 to change the desired label in the tag information in the desired tag editing field 51, and then operates the button 52, whereby the label of the tag information attached to the selected document file is updated to the changed label. In the example of Fig. 11, as the label for the item "details of the accident" in the tag information, "reaction to the action / unreasonable action" is selected from the pull-down menu corresponding to the tag information list 101 in Fig. 3. As a result, the label "caught / entangled" for the item "details of the accident" is changed to the label "reaction to the action / unreasonable action."

[0070] When the above operation is performed, editing of tag information is requested to the editing unit 16 of the tag estimation device 10 in Fig. 2. The tag information assigned to the selected document file is edited by the editing unit 16 to create an edited file F1. In addition, tag information of another pre-edited file F2 is also edited in the same way as the edited file F1, based on the threshold value of the estimation probability.

[0071] For example, suppose the estimated probabilities of the label "caught / entangled" assigned to the first to fourth document files, which are the pre-editing file F2, are 85%, 50%, 70%, and 55%, respectively. If the threshold for the estimated probability is set to 60%, the label for the item "Details of the accident" for the first and third document files, which have been assigned labels with estimated probabilities equal to or greater than the threshold, will be changed to "Reaction to action / Unreasonable action." On the other hand, the label for the item "Details of the accident" for the second and fourth document files, which have been assigned labels with estimated probabilities less than the threshold, will be changed to "No choice."

[0072] (6) Other display examples FIG. 12 is a diagram showing a first display example of the file column 42 on the result screen 40. As shown in the upper part of FIG. 12, in the first display example, tag information is displayed in each file column 42 with each type of tag information grouped on a single line. In the example of FIG. 12, the labels for the item "Industry" on the left of FIG. 7 are grouped on a single line, and the labels for the item "Accident Details" on the right of FIG. 7 are grouped on another single line. As shown in the lower part of FIG. 12, the display of the header for the tag information may be omitted. In this case, the number of characters displayed in the file column 42 is reduced, thereby improving the readability of the file column 42.

[0073] FIG. 13 is a diagram showing a second display example of the file column 42 on the result screen 40. As shown in the upper part of FIG. 13, in the second display example, tag information is displayed in each file column 42 with each type of tag information grouped on a single line, similar to the first display example. Here, labels in a parent-child relationship are displayed in the same color. In this case, the user can easily recognize the parent-child relationship of the labels in the tag information. As shown in the lower part of FIG. 13, the display of the header of the tag information may be omitted.

[0074] 14 and 15 are diagrams showing a third display example of the file column 42 on the result screen 40. As shown in FIG. 14, in the third display example, tag information is displayed in each file column 42 with each parent-child relationship grouped on a single line. In this case, the user can easily recognize the parent-child relationship of the labels in the tag information. As shown in the upper part of FIG. 15, the display of the header of the tag information may be omitted. Furthermore, as shown in the lower part of FIG. 15, tag information of the same type may be displayed in the same color.

[0075] Fig. 16 is a diagram showing a fourth display example of the file column 42 on the result screen 40. As shown in Fig. 16, in the fourth display example, buttons 47 are displayed in each file column 42 so as to correspond to the type of tag information. Each time the user operates the operation unit 150 to operate a desired button 47, the label of the tag information corresponding to that button 47 is alternately expanded and cleared.

[0076] Fig. 17 is a diagram showing a fifth display example of the file column 42 on the result screen 40. In the fifth display example shown in Fig. 17, tag information is displayed in each file column 42 with each type grouped on one line and subdivided into the lowest level items.

[0077] Fig. 18 is a diagram showing an example of an editing screen when the training data has a multi-label structure. When the user operates the operation unit 150 to operate the edit icon 46 in the desired file field 42 in Fig. 10, an editing screen 60 in Fig. 18 is displayed as a pop-up screen. As shown in Fig. 18, the editing screen 60 includes multiple tabs 61, a tag editing field 62, and a button 63.

[0078] The multiple tabs 61 correspond to multiple items of tag information, respectively. When the user operates the operation unit 150 to operate a desired tab 61, the labels of the items in the tag information corresponding to that tab 61 are displayed in a tree format in the tag editing field 62. Each time the user operates a desired label using the operation unit 150, the label is alternately selected and deselected. The selected label is displayed with a background color so that it can be identified. It is also possible to select multiple labels.

[0079] The user can perform operations to instruct editing, such as adding, deleting, or changing, for each selected label, in the same way as the operations on the editing screen 50 of Fig. 11. When the user operates the button 63 while instructing editing of each selected label, the label of the selected tag information attached to the selected document file is edited.

[0080] (7) Tag estimation process Fig. 19 is a flowchart showing tag estimation processing by the tag estimation device 10 of Fig. 2. The tag estimation processing of Fig. 19 is performed by the CPU 130 of Fig. 1 executing on the RAM 110 a tag estimation program in a document retrieval support program stored in the ROM 120 or the like. The tag estimation processing will be described below using the flowchart of Fig. 19 together with the tag estimation device 10 of Fig. 2.

[0081] First, the creation unit 11 creates the tag information list 101 based on the user's operation of the operation unit 150 (step S1). Furthermore, the extraction unit 12 extracts some of the document files stored in the database storage device 170 based on the user's operation of the operation unit 150 (step S2). Either step S1 or step S2 may be executed first.

[0082] Next, the teacher data acquisition unit 13 acquires the data created in step S1. Tag Information List 101 The teacher data is generated by assigning labels of the tag information to the document files extracted in step S2 (step S3). Next, the construction unit 14 acquires a BERT that has been trained in advance using a large-scale corpus, and constructs a tag estimation model by training the BERT using the teacher data generated in step S3 (step S4).

[0083] Thereafter, the estimation unit 15 determines whether or not a document file to be searched has been specified (step S5). When step S23 or step S37 in the search process described below is executed, the estimation unit 15 determines that a document file has been specified. When a document file has not been specified, the estimation unit 15 proceeds to step S8.

[0084] When a document file is specified, the estimation unit 15 estimates tag information of the document file using the tag estimation model constructed in step S4 (step S6). The estimation unit 15 also transmits the tag information estimated in step S6 to the search device 20 (step S7), and the process proceeds to step S8.

[0085] In step S8, the editing unit 16 determines whether editing of tag information has been requested for any document file (step S8). When step S31 in the search process described below is executed, the editing unit 16 determines that editing of tag information has been requested. When editing of tag information has not been requested, the editing unit 16 returns to step S5.

[0086] When editing of tag information is requested, the editing unit 16 creates an edited file F1 by editing the tag information assigned to the document file in accordance with the request (step S9).The editing unit 16 also edits the tag information of another unedited file F2 based on the threshold value of the estimated probability (step S10).

[0087] The editing unit 16 also edits the tag information edited in step S10. Search device 20 (Step S11). Furthermore, the editing unit 16 updates the tag estimation model constructed in step S4 based on the editing result in step S10 (Step S12), and returns to step S5. Either step S11 or step S12 may be executed first.

[0088] (8) Search processing Figures 20 and 21 are flowcharts showing the search processing by the search device 20 of Figure 2. The search processing of Figures 20 and 21 is performed by the CPU 130 of Figure 1 executing on the RAM 110 a search program in a document search support program stored in the ROM 120 or the like. The search processing will be explained below using the search device 20 of Figure 2 and the flowcharts of Figures 20 and 21.

[0089] First, the index acquisition unit 21 acquires index information excluding tag information of each document file by crawling the database storage device 170 (step S21). Next, the registration unit 22 registers the index information acquired in step S21 in the search unit 24 (step S22).

[0090] Next, the file acquisition unit 23 specifies the document file to be searched based on the index information registered in step S22 (step S23). In this case, steps S6 and S7 of the tag estimation process are sequentially executed, whereby tag information of the specified document file is estimated and the estimated tag information is transmitted to the search device 20.

[0091] The registration unit 22 receives the tag information transmitted from the tag estimation device 10 (step S24). The registration unit 22 also registers the tag information received in step S24 in the search unit 24 as part of the index information of the document file specified in step S23 (step S25). Thereafter, the search unit 24 displays the search screen 30 of FIG. 9 on the display unit 160 (step S26).

[0092] Next, the search unit 24 determines whether or not a character string has been input in the input field 31 of the search screen 30 (step S27). Note that in step S26, the search unit 24 may cause the display unit 160 to display the result screen 40 of Fig. 10 instead of the search screen 30. In this case, in step S27, the search unit 24 determines whether or not a character string has been input in the input field 41 of the result screen 40.

[0093] If no character string is input, the search unit 24 proceeds to step S30. If a character string is input, the search unit 24 searches for document files to which tag information matching the input character string has been assigned (step S28). Next, the search unit 24 displays a result screen 40 indicating the search results of step S28 on the display unit 160 (step S29), and proceeds to step S30. In step S30, the acceptance unit 25 determines whether an instruction to edit tag information for any document file has been accepted from the result screen 40 displayed in step S29 (step S30).

[0094] When an instruction to edit tag information is received, the receiving unit 25 requests the tag estimation device 10 to edit the tag information assigned to the document file (step S31). In this case, steps S9 to S11 of the tag estimation process are sequentially executed, whereby the tag information assigned to the specified document file is edited in accordance with the request, and the edited tag information is transmitted to the search device 20.

[0095] The registration unit 22 receives the edited tag information transmitted from the tag estimation device 10 (step S32). Furthermore, the registration unit 22 updates the tag information registered in step S25 to the edited tag information received in step S32 (step S33), and returns to step S27.

[0096] If an instruction to edit tag information is not received in step S30, the receiving unit 25 proceeds to step S34. In step S34, the index acquisition unit 21 crawls the database storage device 170 to determine whether a new document file has been added to the database storage device 170 (step S34). If a new document file has not been added, the index acquisition unit 21 returns to step S27.

[0097] When a new document file is added, the index acquisition unit 21 acquires index information excluding tag information of the new document file (step S35). S35 The index information acquired in step S36 is registered in the search unit 24. Next, the file acquisition unit 23 specifies a new document file as a search target based on the index information registered in step S36 (step S37), and the process returns to step S27.

[0098] (9) Effects In the document retrieval support system 100 according to this embodiment, training data in which tag information is assigned to some document files extracted from a plurality of document files to be searched is acquired by the training data acquisition unit 13. The training data acquired by the training data acquisition unit 13 is applied to a Transformer-based machine learning model that has been trained in advance using a corpus, and a tag estimation model for estimating tag information to be assigned to the document files is constructed by the construction unit 14. Using the tag estimation model constructed by the construction unit 14, the registration unit 22 assigns tag information to each of the plurality of document files to be searched.

[0099] In this document retrieval support system 100, BERT, a Transformer-based machine learning model, is trained in advance using a corpus, so that a tag information estimation model that can estimate tag information with high accuracy is constructed even when the amount of training data is small. In this case, there is no need to acquire a large amount of training data. Therefore, by using a tag information estimation model constructed using some, but not all, of the multiple document files to be searched as training data, tag information can be assigned to each of the multiple document files to be searched.

[0100] With this configuration, users do not need to assign tag information to a huge number of document files. Furthermore, there is no variation in the expression of the labels in the tag information. Therefore, there is almost no cost involved in assigning, maintaining, or arranging tag information. This allows documents to be searched while suppressing increases in costs. Furthermore, since tag information is immediately assigned to newly created document files, newly created document files can also be searched immediately.

[0101] The user can easily use document files that match a desired character string by entering the desired character string in the input field 31 of the search screen 30 or the input field 41 of the result screen 40. In addition, the user can easily use document files that have the tag information added by selecting the desired tag information displayed in the tag search field 43 of the result screen 40.

[0102] Furthermore, the user can easily instruct editing of tag information assigned to a desired document file by operating the edit icon 46 in the file field 42 on the result screen 40. This allows the tag information assigned to a document file to be appropriately corrected even if inappropriate tag information has been assigned to the document file. Furthermore, when tag information is edited, the tag estimation model is updated based on the editing results, making it easy to assign appropriate tag information to a document file.

[0103] (10) Other embodiments In the above embodiment, the document retrieval support system 100 includes the memory unit 140, the operation unit 150, the display unit 160, and the database storage device 170, but the embodiment is not limited to this. The document retrieval support system 100 does not have to include some or all of the memory unit 140, the operation unit 150, the display unit 160, and the database storage device 170, as long as the document retrieval support system 100 is connectable to the memory unit 140, the operation unit 150, the display unit 160, and the database storage device 170.

[0104] Furthermore, the teacher data acquisition unit 13 generates teacher data by assigning any tag information in the tag information list 101 created by the creation unit 11 to some of the document files extracted by the extraction unit 12, but the embodiment is not limited to this. The teacher data acquisition unit 13 may also acquire teacher data in which tag information is assigned to some of the document files extracted from multiple document files to be searched from an external storage device, etc. In this case, the tag estimation device 10 does not need to include the creation unit 11 and the extraction unit 12.

[0105] (11) Correspondence between each component of the claims and each part of the embodiment Below, examples of correspondence between each element of the claims and each element of the embodiments will be described, but the present invention is not limited to the following examples. Various other elements having the configuration or function described in the claims can also be used as each element of the claims.

[0106] In the above embodiment, the teacher data acquisition unit 13 is an example of a teacher data acquisition unit, the construction unit 14 is an example of a construction unit, the registration unit 22 is an example of an assignment unit, and the document search support system 100 is an example of a document search support system. The creation unit 11 is an example of a creation unit, the extraction unit 12 is an example of an extraction unit, the result screen 40 is an example of a result screen, the search unit 24 is an example of a search unit, the reception unit 25 is an example of a reception unit, and the editing unit 16 is an example of an editing unit. [Explanation of symbols]

[0107] 10...tag estimation device, 11...creation unit, 12...extraction unit, 13...teaching data acquisition unit, 14...construction unit, 15...estimation unit, 16...editing unit, 20...search device, 21...index acquisition unit, 22...registration unit, 23...file acquisition unit, 24...search unit, 25...reception unit, 30...search screen, 31, 41...input field, 40...result screen, 42...file field, 43...tag search field, 44, 45, 47, 52, 63...button, 46...editing Icons, 50, 60...Editing screen, 51, 62...Tag editing field, 53...Check box, 61...Tab, 100...Document search support system, 101...Tag information list, 102, 103...Table, 110...RAM, 120...ROM, 130...CPU, 140...Memory unit, 150...Operation unit, 160...Display unit, 170...Database storage device, 180...Bus, F1...Edited file, F2...File before editing

Claims

1. a teacher data acquisition unit that acquires teacher data to which tag information is added to some document files extracted from a plurality of document files to be searched; a construction unit that constructs a tag estimation model for estimating tag information to be assigned to a document file by applying the training data acquired by the training data acquisition unit to a Transformer-based machine learning model that has been trained in advance using a corpus; an assigning unit that assigns tag information to each of a plurality of document files to be searched using the tag estimation model constructed by the constructing unit; a search unit that, after tag information is assigned to each of the plurality of document files to be searched by the assigning unit, accepts an input of a character string, searches the plurality of document files to be searched for for a document file that matches the character string, and outputs a result screen showing the search results; the result screen output by the search unit further displays, as tag information, tag information assigned to the document file searched by the search unit and tag information belonging to the same layer as the tag information; A document search support system in which the search unit searches for document files to which tag information selected by the user has been assigned from the tag information displayed on the result screen, and outputs the result screen showing the search results.

2. The training data indicates a relationship between the content of a document file, which is an explanatory variable, and tag information assigned to the document file, which is a target variable; 2. The document search support system according to claim 1, wherein the tag estimation model is constructed by the Transformer-based machine learning model learning the relationship between the contents of document files and tag information assigned to the document files based on the training data.

3. 3. The document retrieval support system according to claim 1, wherein the Transformer-based machine learning model is BERT (Bidirectional Encoder Representations from Transformers).

4. a generating unit for generating a tag information list showing a list of candidates for tag information to be assigned to a plurality of document files to be searched; an extracting unit that extracts some document files from the plurality of document files to be searched; A document search support system described in any one of claims 1 to 3, wherein the teacher data acquisition unit generates the teacher data by assigning any tag information in the tag information list created by the creation unit to a portion of the document files extracted by the extraction unit.

5. a receiving unit that receives an instruction to edit tag information assigned to any of a plurality of document files to be searched; an editing unit that edits tag information of the document file designated by the receiving unit and edits tag information of other document files to be searched, 5. The document search support system according to claim 1, wherein the tagging unit updates tag information added to a document file to tag information edited by the editing unit.

6. 6. The document retrieval support system according to claim 5, wherein said editing section selectively edits the tag information of the document file to be retrieved based on a threshold value of the estimated probability of said estimated tag information.

7. the result screen output by the search unit further displays a GUI (Graphical User Interface) that is operated by a user; 7. The document search support system according to claim 5, wherein the accepting unit accepts an instruction to edit tag information assigned to any one of a plurality of document files to be searched by operating the GUI.

8. 8. The document retrieval support system according to claim 5, wherein the editing unit updates the tag estimation model constructed by the construction unit based on an editing result.

9. Obtaining teacher data to which tag information is added to some document files extracted from a plurality of document files to be searched; constructing a tag estimation model for estimating tag information to be assigned to a document file by applying the acquired training data to a Transformer-based machine learning model that has been trained in advance using a corpus; assigning tag information to each of a plurality of document files to be searched using the constructed tag estimation model; After tag information is assigned to each of the plurality of document files to be searched, input of a character string is accepted, and document files matching the character string are searched from the plurality of document files to be searched, and a result screen is output that displays the search results, the tag information assigned to the searched document files, and candidate tag information belonging to the same layer as the tag information; A document search support method including searching for document files to which tag information selected by a user is assigned from the tag information displayed on the results screen, and outputting the results screen showing the search results.

10. A document retrieval support program executable by a processing device, A process of acquiring training data to which tag information is added to some document files extracted from multiple document files to be searched; A process of constructing a tag estimation model for estimating tag information to be assigned to a document file by applying the acquired training data to a Transformer-based machine learning model that has been trained in advance using a corpus; a process of assigning tag information to each of a plurality of document files to be searched using the constructed tag estimation model; a process of receiving an input of a character string after tag information has been assigned to each of the plurality of document files to be searched, searching for document files that match the character string from the plurality of document files to be searched, and outputting a result screen that displays the search results, the tag information assigned to the searched document files, and candidate tag information that belongs to the same layer as the tag information; a process of searching for document files to which tag information selected by a user is assigned from among the tag information displayed on the result screen, and outputting the result screen showing the search results; A document search support program executed by the processing device.

Images