Similar case extraction support system, similar case extraction support method, and program

Abstract

To support easy extraction of appropriate similar cases.SOLUTION: A similar case extraction support system 1 comprises: storage means for storing sentences representing the contents of items having correlations and content sentence vectors where the sentences have been converted to vectors; vector generation means for converting a sentence into a vector using a transformer-based model; calculation means for calculating a degree of similarity between an input sentence vector representing a vector of the input sentence converted by the vector generation means and a content sentence vector; order determination means for determining an order of the sentences representing the contents in accordance with the degree of similarity; and display means for displaying results including the sentences representing the contents, the degrees of similarity, and the orders.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a similar case extraction support system, a similar case extraction support method, and a program. [Background technology]

[0002] There is known a technology that uses case data to extract information related to a sentence entered by a user from past cases, thereby making it useful for manufacturing process work, product design, etc. For example, in Patent Document 1, keywords are extracted from a plurality of sentences that represent failure modes related to the input information from an existing FMEA (Failure Mode and Effect Analysis) sheet, according to input information indicating information on parts and functions entered by a user into a new FMEA sheet. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2018-45548 Summary of the Invention [Problem to be solved by the invention]

[0004] Because the notation and expressions used in FMEA formats vary depending on the writer, it is desirable to utilize existing case data after understanding the context. However, the technology in Patent Document 1 does not take context into consideration, so it is not always possible to extract appropriate similar cases from existing cases, depending on the input text, for example. Therefore, there is a need for support that makes it easy to extract appropriate similar cases.

[0005] The present disclosure has been made in consideration of the above-described problems, and aims to provide a similar case extraction support system, a similar case extraction support method, and a program that support the easy extraction of appropriate similar cases. [Means for solving the problem]

[0006] In order to achieve the above-mentioned object, the similar case extraction support system according to the present disclosure includes a storage means for storing sentences representing the content of correlated items and content sentence vectors obtained by converting the sentences into vectors; a vector generation means for converting the sentences into vectors using a transformer-based model; a calculation means for determining the similarity between the content sentence vector and an input sentence vector representing the vector of the input sentence converted by the vector generation means; a ranking determination means for determining the ranking of the sentences representing the content according to the similarity; and a display means for displaying results including the sentences representing the content, the similarity, and the ranking. [Effects of the Invention]

[0007] According to the present disclosure, a sentence is converted into a vector, the similarity between an input sentence vector representing the vector of the input sentence and a content sentence vector is calculated, the ranking of the sentences representing the content is calculated according to the calculated similarity, and the results including the sentences representing the content, the similarity, and the ranking are displayed, thereby providing support for easily extracting appropriate similar cases. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram showing the configuration of a similar case extraction support system according to a first embodiment. [Figure 2] FIG. 1 is a block diagram showing a functional configuration of a user terminal device according to a first embodiment. [Figure 3] FIG. 1 is a block diagram showing a functional configuration of a similar case extraction support device according to a first embodiment. [Figure 4] FIG. 1 is a diagram showing an example of a hardware configuration of a similar case extraction support device according to a first embodiment. [Figure 5] Flowchart showing vector generation processing according to the first embodiment [Figure 6] FIG. 1 is a diagram showing an example of a technical document storage unit according to the first embodiment. [Figure 7] Flowchart showing similar case extraction support processing 1 according to the first embodiment [Figure 8]1 is a flowchart showing a similarity score calculation process according to the first embodiment; [Figure 9] FIG. 10 is a diagram illustrating an example of a calculation information storage unit according to the first embodiment. [Figure 10] FIG. 10 is a diagram illustrating an example of a calculation information storage unit according to the first embodiment. [Figure 11] A diagram to explain an example of how to search for similar cases in both work and risk [Figure 12] FIG. 10 is a block diagram showing the functional configuration of a similar case extraction support device according to a second embodiment. [Figure 13] Flowchart showing similar case extraction support processing 2 according to the second embodiment [Figure 14] An example of the correspondence between assumed risks and clusters [Figure 15] A diagram for explaining correction values ​​for each cluster [Figure 16] An example of clusters, scores, and ranks of assumed risks [Figure 17] A diagram for explaining correction values ​​for each cluster [Figure 18] A diagram to explain score correction and re-ranking [Figure 19] FIG. 10 is a block diagram showing the functional configuration of a similar case extraction support device according to a third embodiment. [Figure 20] Flowchart showing model creation processing according to the third embodiment [Figure 21] Flowchart showing inference processing according to the third embodiment [Figure 22] Flowchart showing similar case extraction support processing 3 according to the third embodiment [Figure 23] A diagram showing an example of the correspondence between work risks, scores, ranks, categories, etc. [Figure 24] (a) A diagram for explaining the re-identification of the order of combinations of work content, work risks, etc. belonging to the work-related category, (b) A diagram for explaining the re-identification of the order of combinations of work content, work risks, etc. belonging to the purchased item-related category. DETAILED DESCRIPTION OF THE INVENTION

[0009] (Embodiment 1) A similar case extraction support system, a similar case extraction support method, and a program according to embodiments of the present disclosure will be described in detail below with reference to the drawings. Note that identical or corresponding parts in the drawings are designated by the same reference numerals.

[0010] A similar case extraction support system according to an embodiment of the present disclosure is a system that supports the task of extracting similar cases by searching a database (hereinafter referred to as DB) that stores multiple cases, each containing multiple pairs of a first item and a second item that are correlated, and extracting cases similar to an input case. In the following description, pairs of work content and risks associated with the work are used as an example of multiple cases. More specifically, the similar case extraction support system according to this embodiment includes a database (DB) that pre-stores sentences that describe the work content and the risks associated with the work, searches the DB using an input sentence that describes the work content as a key, and extracts sentences that describe the work content and the risks associated with the work that are similar to the input sentence. When searching the DB, the similar case extraction support system vectorizes the input sentence and the sentences stored in the DB, and extracts sentences that describe the work content and the risks associated with the work using the similarity of the vectors. This will be described in detail below.

[0011] As shown in FIG. 1, a similar case extraction support system 1 according to this embodiment includes a similar case extraction support device 100 that supports extraction of similar cases, and a user terminal device 200 used by a user.

[0012] The similar case extraction support device 100 and the user terminal device 200 are connected to each other so that they can communicate with each other via a network NW. The similar case extraction support device 100 and the user terminal device 200 are each configured with a computer. The network NW may be wired, wireless, a WAN (Wide Area Network), a LAN (Local Area Network), an intranet, or an extranet.

[0013] The configuration of the user terminal device 200 will be described with reference to FIG. The user terminal device 200 is operated by the user to input a sentence (hereinafter, "input sentence") indicating an operation for which similar cases are to be searched, and displays the operation extracted as a similar case and its risk.

[0014] The user terminal device 200 comprises a registration reception unit 201 that receives input text indicating work content, a search input identification information reception unit 202 that identifies, from the accumulated input text, text to be searched for similar cases as search input information, a search input information sending unit 203 that sends the search input information to the similar case extraction support device 100, a search result information acquisition unit 204 that acquires search result information sent from the similar case extraction support device 100, a search result information output unit 205 that outputs the search result information, a display unit 206 that displays various information including the search result information, and a second memory unit 230 that stores information.

[0015] The second storage unit 230 includes a search input information storage unit 231 that stores input sentences for which similar cases are to be searched, and a search result information storage unit 232 that stores search result information. The search input information storage unit 231 cumulatively stores input sentences accepted by the registration acceptance unit 201. Furthermore, the search result information storage unit 232 stores the search result information that the search result information acquisition unit 204 receives and acquires from the similar case extraction support device 100.

[0016] The registration receiving unit 201 receives the registration of an input sentence, and stores the received input sentence in the search input information storage unit 231. The input sentence is, for example, a sentence indicating the details of work in a manufacturing process.

[0017] The search input specification information receiving unit 202 receives information specifying search input information from among the input sentences stored in the search input information storage unit 231. The search input information sending unit 203 sends the search input information to the similar case extraction support device 100.

[0018] The search result information acquisition unit 204 acquires the search result information output by the similar case extraction support device 100. The search result information output unit 205 outputs the search result information acquired by the search result information acquisition unit 204 to the display unit 206.

[0019] The display unit 206 displays various information including the search result information output by the search result information output unit 205. The display unit 206 includes, for example, a liquid crystal display. The display unit 206 is an example of a display means according to the present disclosure.

[0020] Next, the configuration of the similar case extraction support device 100 will be described with reference to FIG.

[0021] The similar case extraction support device 100 includes a search input information acquisition unit 101 that acquires search input information, a vector generation unit 102 that converts sentences into vectors, a similarity calculation unit 103 that calculates the similarity between vectors, a search unit 104 that searches for work content, a rank sorting unit 105 that ranks and rearranges the search results, a transmission unit 106 that outputs the results ranked and rearranged by the rank sorting unit 105, and a first memory unit 120 that stores information, programs, etc.

[0022] The first storage unit 120 includes a search input information storage unit 121 that stores received search input information, a vector information storage unit 122 that stores vector information, a calculation information storage unit 123 that stores similarities calculated by the similarity calculation process, a search result information storage unit 124 that stores result information indicating the processing results, a technical document storage unit 125 that stores technical documents, and a trained model storage unit 126 that stores trained models. The first storage unit 120 also stores programs executed by the search input information acquisition unit 101, the vector generation unit 102, the similarity calculation unit 103, the search unit 104, the rank sorting unit 105, the transmission unit 106, etc. The first storage unit 120 is an example of a storage means according to the present disclosure.

[0023] The search input information storage unit 121 stores search input information sent from the user terminal device 200. The vector information storage unit 122 stores vector information generated by the vector generation unit 102. The calculation information storage unit 123 stores similarities calculated by the similarity calculation unit 103. The search result information storage unit 124 stores information ranked and sorted by the rank sorting unit 105. The technical document storage unit 125 corresponds to a case DB and stores technical documents to be searched. In this embodiment, the technical document has a configuration in which work content, work risks, countermeasures, and key information are associated with each other, as exemplified in FIG. 6. Details of the technical document will be described later. Note that the work content and work risks are examples of the first and second matters related to the present disclosure. The trained model storage unit 126 stores a trained model of BERT (Bidirectional Encoder Representations from Transformers), which is a transformer-based model.

[0024] The search input information acquisition unit 101 acquires search input information transmitted from the user terminal device 200 via the network NW. The search input information acquisition unit 101 stores the acquired search input information in the search input information storage unit 121.

[0025] The vector generation unit 102 converts the search input information stored in the search input information storage unit 121 and sentences representing work content and work risks in technical documents stored in the technical document storage unit 125 into vectors by applying them to a trained model stored in the trained model storage unit 126. The BERT trained model converts one sentence into a 768-dimensional vector. However, the number of dimensions varies depending on the model. The vector generation unit 102 stores the vectors corresponding to the generated search input information in the vector information storage unit 122. Furthermore, the vector generation unit 102 associates the vectors corresponding to the work content and the vectors corresponding to the work risks in the technical documents with key information and stores them in the vector information storage unit 122. The vector generation unit 102 is an example of a vector generation means according to the present disclosure. The vectors corresponding to the work content and the vectors corresponding to the work risks are examples of a first vector and a second vector according to the present disclosure.

[0026] The similarity calculation unit 103 calculates the cosine similarity between the vector corresponding to the search input information generated by the vector generation unit 102 and each vector corresponding to a sentence representing the work content in the technical document. The similarity calculation unit 103 stores the calculated cosine similarity in the calculated information storage unit 123. The similarity calculation unit 103 also calculates the similarity between the work content vector determined to have the highest similarity and the associated work risk vector, and other work content vectors and work risk vectors in the technical document. Note that cosine similarity is a measure of the similarity between two vectors. Specifically, cosine similarity is the cosine value of the angle formed by the two vectors. The similarity calculation unit 103 is an example of a calculation means according to the present disclosure. The vector corresponding to the sentence representing the work content in the technical document is an example of a content sentence vector according to the present disclosure.

[0027] The search unit 104 searches the similarities stored in the calculated information storage unit 123, identifies those with a high similarity that meet the conditions, and extracts sentences that represent the work content corresponding to the identified similarity.

[0028] The rank sorting unit 105 sorts the combinations of work content and risk according to the similarity and ranks them in descending order of similarity. Details will be described later. The rank sorting unit 105 is an example of a rank specifying means according to the present disclosure.

[0029] The transmission unit 106 stores the search result information indicating the search results rearranged and ranked by the rank sorting unit 105 in the search result information storage unit 124 and transmits it to the user terminal device 200 .

[0030] Next, an example of the hardware configuration of the similar case extraction support device 100 having the above configuration will be described with reference to Fig. 4. As shown in Fig. 4, the similar case extraction support device 100 is realized by a computer such as a personal computer or a microcontroller. The hardware configuration of the user terminal device 200 is the same as that of the similar case extraction support device 100.

[0031] The similar case extraction support device 100 includes a processor 1001 that executes an operating program, a memory 1002 that serves as the main storage area of ​​the processor 1001, an interface 1003 that realizes a communication function, and a secondary storage device 1004 that stores the operating program for executing processing. The processor 1001, memory 1002, interface 1003, and secondary storage device 1004 are connected to one another via a bus 1000.

[0032] The processor 1001 is, for example, a CPU (Central Processing Unit). The processor 1001 loads an operating program stored in a secondary storage device 1004 into a memory 1002 and executes the program, thereby realizing the functions of the similar case extraction support device 100, specifically, the functions of a search input information acquisition unit 101, a vector generation unit 102, a similarity calculation unit 103, a search unit 104, a rank sorting unit 105, and a transmission unit 106.

[0033] The memory 1002 is a main storage device configured, for example, by a RAM (Random Access Memory). The memory 1002 stores the operating program that the processor 1001 reads from the secondary storage device 1004. The memory 1002 also functions as a work memory when the processor 1001 executes the operating program.

[0034] The interface 1003 is an I / O (Input / Output) interface such as a serial port, a USB (Universal Serial Bus) port, a network interface, etc. The interface 1003 realizes the communication function of the similar case extraction support device 100.

[0035] The secondary storage device 1004 is, for example, a flash memory, a hard disk drive (HDD), or a solid state drive (SSD). The secondary storage device 1004 stores the operating program executed by the processor 1001, search input information, vector information, calculation information, search result information, technical documents, and trained models.

[0036] Next, an operation of the similar case extraction support system 1 having the above configuration to detect and output cases similar to a case specified by a user will be described.

[0037] In the following explanation, an example will be given in which a user inputs a sentence indicating the work content in the manufacturing of electrical and electronic equipment, and work content similar to the work content indicated by the input sentence, as well as corresponding risks and countermeasures, are extracted from technical documents pre-stored in the technical document storage unit 125 and presented to the user.

[0038] As a premise, a person in charge stores the technical document to be searched in the technical document storage unit 125 of the similar case extraction support device 100. As illustrated in FIG. 6, the technical document is associated with key information and includes sentences indicating the work content, sentences indicating the work risks, sentences indicating the countermeasures, and key information. The key information corresponds to identification information that associates the work content, work risks, and countermeasures. For example, the key information "TD_0001" represents the technical document's identification number "TD" and serial number "0001." The sentences indicating the work content and the sentences indicating the work risks are examples of a first sentence representing a first item and a second sentence representing a second item according to the present disclosure.

[0039] Next, the person in charge starts the vector generation unit 102 of the similar case extraction support device 100, and starts a vector generation process for vectorizing the technical documents stored in the technical document storage unit 125.

[0040] As shown in FIG. 5, when the vector generation process starts, the vector generation unit 102 generates a vector of the technical document information by applying the technical document information stored in the technical document storage unit 125 to the BERT trained model stored in the trained model storage unit 126 of the first storage unit 120 (step S11). More specifically, the vector generation unit 102 vectorizes each of the documents of the work content, work risks, and countermeasures illustrated in FIG. 6. The work content vector is an example of a work content vector according to the present disclosure. The work risk vector is an example of a risk vector according to the present disclosure.

[0041] The vector generating unit 102 stores the generated vector as technical document vector information in the vector information storage unit 122 together with the key information (step S12).

[0042] Next, a similar case extraction support process 1 for extracting similar cases from the generated technical document vector information will be described. For example, the user inputs a sentence describing the work content into the user terminal device 200 in order to learn about the risks associated with the work. For example, with the improvement of the GUI (Graphical User Interface), the user inputs the sentence describing the work content into an entry field displayed on the display unit 206 of the user terminal device 200. As an example, assume that the user inputs the sentence "connector installation." After finishing inputting the sentence, the user selects an icon representing "register," "confirm," "OK," or the like.

[0043] In response to the user's selection, the registration acceptance unit 201 generates search input information by adding user identification information, date and time information, work process identification information, etc. to the input text, and stores the search input information in the search input information storage unit 231. For example, in the case of work content for work process P4 input by user A at 14:10 on March 1, 2024, the input text is stored in a folder with a folder name representing "A202403011410P4." The text "connector installation" is an example of an input text according to the present disclosure.

[0044] Next, the user specifies an input sentence to be searched from among the input sentences stored in the search input information storage unit 231. For example, the user specifies the folder name "A202403011410P4" as described above. The search input specification information receiving unit 202 receives this specification operation, and the search input information sending unit 203 sends the search input information saved in the folder with the received folder ID to the similar case extraction support device 100.

[0045] The similar case extraction support device 100 receives search input information and starts a similar case extraction support process 1 illustrated in FIG. First, the search input information acquisition unit 101 of the similar case extraction support device 100 acquires search input information from the user terminal device 200 and stores it in the search input information storage unit 121 (step S21).

[0046] The vector generation unit 102 generates a vector of the search input by inputting the search input stored in the search input storage unit 121 into the trained model of BERT stored in the trained model storage unit 126 (step S22). The vector generation unit 102 stores the generated vector as an input sentence vector in the vector information storage unit 122 together with the sentence information and the identification number of the search input.

[0047] The similarity calculation unit 103 executes a similarity score calculation process (step S23). Specific processing details of the similarity score calculation process are shown in Fig. 8. In this embodiment, the similarity score corresponds to, for example, a cosine similarity that takes into consideration both the work content and the work risk exemplified in Fig. 6. Details will be described later.

[0048] The similarity calculation unit 103 calculates the cosine similarity between the vector of the input sentence and the vector of each work content of the technical document, which are stored in the vector information storage unit 122 (step S23a). The cosine similarity is calculated by dividing the inner product of the two vectors by the product of the magnitudes of the two vectors. For example, the vector a → and vector b → Cosine similarity cos(a → ,b → ) is expressed by the following equation 1.

[0049] [Number 1] cos(a → ,b → )=(a → ×b → ) / (||a → ||×||b → ||)

[0050] 9, the similarity calculation unit 103 associates the calculated cosine similarity with information indicating the work content and key information, and stores them in the calculated information storage unit 123. Fig. 9 shows an example of the cosine similarity between a vector corresponding to the work content in the technical document and a vector corresponding to the input text, in the case where the input text is "connector installation."

[0051] The rank sorting unit 105 associates the technical documents with the calculated cosine similarities as shown in Fig. 9, and stores them in the calculated information storage unit 123. Next, the rank sorting unit 105 sorts the documents by cosine similarities as shown in Fig. 10 (step S23b). Specifically, the rank sorting unit 105 sorts the cosine similarities in descending order and assigns ranks indicating the order.

[0052] In this way, as shown in FIG. 11, the task content that is most similar to the input sentence is identified based on the vector similarity. Next, the similarity calculation unit 103 performs a process of extracting similar cases based on the similarities between a plurality of tasks and the similarities between task risks.

[0053] As shown in FIG. 11, the similarity calculation unit 103 calculates the cosine similarity between the vector corresponding to the work content of rank 1, "inserting a connector into a circuit board," and the vectors of other work contents.

[0054] The search unit 104 identifies and extracts task contents whose cosine similarity with the top-ranked task content Q1 is higher than a preset reference value (step S23c). The reference value is a predetermined numerical value, such as a similarity of 0.75 or 0.8. Here, it is assumed that two task contents, task content Q1.1 and task content Q1.2, have been identified, as shown in FIG. 11. The search unit 104 identifies and extracts the pair of task content Q1.1 and its task risk A1.1, and the pair of task content Q1.2 and its task risk A1.2, as candidates for similar cases. The search unit 104 stores the extracted pair of task content Q1.1 and its task risk A1.1, and the pair of task content Q1.2 and its task risk A1.2 in the calculation information storage unit 123. In the example of FIG. 11, there are two work contents, Q1.1 and Q1.2, that are highly similar to the work content Q1, but there may be zero or one, or three or more.

[0055] The search unit 104 also identifies the task risk A1 corresponding to the task content Q1 ranked first as the top risk (step S23d).

[0056] Next, the similarity calculation unit 103 extracts task risks similar to the top-ranked task risk from among the task risks included in the technical document. Specifically, the cosine similarity between the vector of the top-ranked task risk and the vectors of the other task risks is calculated, and task risks with cosine similarities equal to or greater than a reference value are identified and extracted (step S23e). In the example shown in FIG. 10, the cosine similarity between the vector of the task risk "Partially Inserted Connector" ranked 1 and the vectors of the other task risks stored in the technical document storage unit 125 is calculated, and task risks with cosine similarities equal to or greater than a reference value are identified. Here, as shown in FIG. 11, it is assumed that two task risks, task risks A1a and A1b, have been identified. The search unit 104 identifies and extracts the pair of task risk A1a and corresponding task content Q1a and the pair of task risk A1b and corresponding task content Q1b as candidates for similar cases. The search unit 104 stores the extracted pair of task risk A1a and corresponding task content Q1a, and the pair of task risk A1b and corresponding task content Q1b in the calculation information storage unit 123. Note that in the example of Fig. 11, there are two task risks, A1a and A1b, that are highly similar to task risk A1, but there may be zero or one, or three or more.

[0057] Next, the similarity calculation unit 103 calculates a similarity score SC for each extracted pair (step S23f). The similarity calculation unit 103 calculates the similarity score SC, for example, using the following equation 2. In equation 2, DQ represents the similarity of the work content, and DA represents the similarity of the work risk. Furthermore, W Q is the weight of the work content, W A represents the weight for the work risk, and W Q and W A The sum of these is assigned to 1. The weighting for the work content is W Q and the weight for work risk W A is determined in advance and stored in the calculation information storage unit 123.

[0058] [Number 2] Similarity score SC=DQ×W Q +DA×W A

[0059] The similarity calculation unit 103 calculates a similarity score SC for pairs of work content and work risk extracted in steps S23c to S23e and stored in the calculation information storage unit 123. In the example of FIG. 11, the similarity score SC is calculated for each of the pair of work content Q1 and work risk A1, the pair of work content Q1.1 and work risk A1.1, the pair of work content Q1.2 and work risk A1.2, the pair of work content Q1a and work risk A1a, and the pair of work content Q1b and work risk A1b. The similarity calculation unit 103 stores the similarity score SC calculated for each pair in the calculation information storage unit 123. This completes the similarity score calculation process (step S23), and the process proceeds to step S24 shown in FIG. 7.

[0060] The rank sorting unit 105 ranks the pair of work content and work risk with the largest similarity score SC obtained in step S23 as number 1, and then ranks the other pairs of work content and work risk in order according to their similarity scores (step S24).The rank sorting unit 105 rearranges the pairs of work content and work risk and the measures corresponding to those pairs in order from number 1, and stores them in the search result information storage unit 124 as search result information.

[0061] The transmission unit 106 outputs the search result information stored in the search result information storage unit 124 to the user terminal device 200 (step S25). The search result information acquisition unit 204 of the user terminal device 200 receives the search result information and stores it in the search result information storage unit 232. The search result information output unit 205 outputs the search result information stored in the search result information storage unit 232 to the display unit 206. The display unit 206 displays the search result information output by the search result information output unit 205 (step S26). Then, the similar case extraction support process 1 ends.

[0062] As described above, the similar case extraction support system 1 according to the first embodiment converts text into vectors using BERT, calculates the similarity between an input text and a text describing work content or work risk, calculates a ranking of combinations of work and risk based on the calculated similarity, and displays a result including the combination, similarity, and ranking. Therefore, it is possible to extract risks associated with the work content described by the input text, thereby supporting the easy extraction of appropriate similar cases. The extracted appropriate cases, specifically the extracted risks, can be used to improve correlated work, design the product involved, and so on.

[0063] (Modification of the first embodiment) In the first embodiment, similar work contents and work risks are found only for the work content Q1 and work risk A1 with the highest similarity, as shown in Fig. 11. However, this is not limiting, and similar work contents and work risks may be found for each of the work contents and corresponding work risks with the highest similarity up to the nth (n is a natural number) ranking. Furthermore, only the top m (m is a natural number) pairs of work content and work risk may be transmitted to the user terminal device 200 based on the calculated similarity score SC.

[0064] In the first embodiment, the various types of information in the first storage unit 120 are provided in the similar case extraction support device 100, but may be provided outside the similar case extraction support device 100. For example, the various types of information in the first storage unit 120 may be provided in a cloud-based server outside the similar case extraction support device 100.

[0065] In the first embodiment, the similar case extraction support system 1 includes the similar case extraction support device 100 and the user terminal device 200, but it does not have to include the user terminal device 200. In other words, the similar case extraction support device 100 may perform all of the processing in the user terminal device 200, including processing such as receiving search input information and displaying search result information.

[0066] Additionally, the risk priority number (hereinafter referred to as RPN), which is used as a standard for risk assessment in FMEA, can be taken into consideration. RPN is a standard that quantifies the degree to which risks hidden in manufacturing processes, products, etc. should be addressed with priority, and is calculated using the following formula 3. [Number 3] RPN = Severity x Occurrence x Detection

[0067] The numerical values ​​of the severity, occurrence, and detection in Equation 3 and the RPN value obtained by multiplying them are associated with the task risk and stored in the technical document storage unit 125. The similarity calculation unit 103 may correct the obtained similarity score SC using the RPN value of the corresponding task risk. For example, the similarity calculation unit 103 may correct the similarity score SC by multiplying or adding the RPN value divided by the maximum RPN value, expressed as (RPN / (maximum RPN value)), to the obtained similarity score SC. This allows the magnitude of the RPN to be reflected in the similarity score.

[0068] In the first embodiment, "connector installation" representing the work content was input as the search input information, but a sentence representing the work risk may also be input. For example, if a product is scratched, searching for a sentence such as "will be scratched" as a risk will extract a list of work content that may cause the risk. Using this as a clue, the process that caused the scratch can be analyzed.

[0069] Furthermore, in the first embodiment, the input is the work content and the output is the expected risk, but this is not limited to this as long as there is a corresponding correlation between the input and output. For example, examples include changes and their risks in DRBFM, change management tables, etc., defects, target processes and areas of concern in quality audits, processes and quality assurance methods in QC process charts, target areas and areas to be improved in improvement activities, occurrences and root causes in defect case studies, work and disaster risks in safety KYT, etc.

[0070] In the first embodiment, an example was described in which a result is output in response to a single sentence input by the user. The input does not have to be a single sentence, and can be multiple sentences prepared in advance. For example, a work procedure manual, a list of change elements, a QC process chart, a list of parts elements, etc. Multiple sentences can be input, and a search and ranking calculation can be performed for each sentence, with the results being displayed on the screen or output as a text file. The processing remains the same for each sentence; after sentence 1 is executed, sentence 2 is executed, and the execution results of all sentences are combined vertically and output.

[0071] In the first embodiment, the sentence input by the user is stored as search input information in the search input information storage unit 231, but it does not have to be stored. For example, after the user inputs a sentence to the user terminal device 200, the input sentence may be immediately sent from the user terminal device 200 to the similar case extraction support device 100 without being stored in the search input information storage unit 231.

[0072] (Embodiment 2) 12 to 18, a similar case extraction support system 1 according to a second embodiment will be described. The second embodiment differs from the first embodiment in that the similar case extraction support device 100 further includes a clustering unit 107 and a rank re-sorting unit 108 in addition to the configuration of the first embodiment, and in that the similar case extraction support device 100 executes a similar case extraction support process 2. The following description of the second embodiment will focus on the differences from the first embodiment.

[0073] As shown in FIG. 12, the similar case extraction support device 100 according to the second embodiment further includes, in addition to the respective functional components of the similar case extraction support device 100 according to the first embodiment, a clustering unit 107 that groups, in other words, clusters, anticipated risks, and a ranking re-sorting unit 108 that re-ranks and rearranges the risks.

[0074] The clustering unit 107 clusters the task risks ranked in step S24 of the similar case extraction support process 1 in the first embodiment as assumed risks. The clustering unit 107 clusters the assumed risks based on the similarity of meaning, as will be described in detail later.

[0075] The rank re-sorting unit 108 re-ranks and re-sorts the data using the clustering results. Details will be described later. The rank re-sorting unit 108 is an example of a rank re-determining means according to the present disclosure.

[0076] In the second embodiment, similar case extraction support processing 2 shown in Fig. 13 is performed. Steps S21 to S24 and steps S25 and S26 are the same as those in the similar case extraction support processing 1 in the first embodiment. As shown in Fig. 13, in the similar case extraction support processing 2 in the second embodiment, after step S24, steps S31 and S32 are performed, and then step S25 is performed.

[0077] After step S24, the clustering unit 107 performs clustering (step S31), a process different from that of the first embodiment. For example, sentences regarding the expected work risks, such as "can cause scratches," "will be scratched," "scratch occurs," "can cause scratches," "scratched," and "scratch occurs," are each vectorized in advance in the technical document vector generation process of the first embodiment. The clustering unit 107 reduces the dimensions of these vectors using techniques such as UMAP, t-SNE, and PCA. Dimension reduction is a method of reducing the dimensions while retaining the features of a vector. For example, the clustering unit 107 compresses a 768-dimensional vector to 7 dimensions.

[0078] Next, the clustering unit 107 clusters this seven-dimensional vector using unsupervised machine learning. Specific methods include, for example, a method of specifying the number of clusters in k-means and a method of not specifying the number of classes in DBSCAN. These methods are commonly used for numerical values, images, etc., and are similar when dealing with text, since it is vectorized. This allows text with similar meanings to belong to the same cluster. For example, the above-mentioned "can scratch," "wound," and "scratch occurs" belong to cluster A, while "wrong direction" and "wrong orientation" belong to cluster B. As illustrated in Figure 14, each assumed risk belongs to cluster A, B, etc. The attribute clusters of assumed risks clustered by the clustering unit 107 are stored in the calculation information storage unit 123.

[0079] Next, based on the clustering result, the similarity calculation unit 103 corrects the similarity score. Based on the corrected similarity score, the rank re-sorting unit 108 re-determines the rank (step S32). Here, how the correction value for correcting the similarity score is determined will be described. As illustrated in FIG. 15, the number of data belonging to cluster A is 5, the number of data belonging to cluster B is 2, the number of data belonging to cluster C is 2, and the number of data belonging to cluster D is 1. In this case, when ranked by the number of data, cluster A is ranked first, clusters B and C are ranked second, and cluster D is ranked fourth. In this embodiment, the correction value includes the reciprocal of the rank of the number of clusters. As illustrated in FIG. 15, cluster A is ranked first, so the correction value is 1 / 1. Furthermore, clusters B and C are ranked second, so the correction value is 1 / 2. Furthermore, cluster D is ranked fourth, so the correction value is 1 / 4.

[0080] For the example of FIG. 16, a flow of applying a correction value to the similarity score and re-calculating the ranking will be described. The similarity calculation unit 103 calculates the number of assumed risks belonging to cluster A as three, namely, "abrasion," "scratch," and "scratch." The similarity calculation unit 103 also calculates the number of assumed risks belonging to cluster B as one, namely, "misdirection," the number of assumed risks belonging to cluster C as one, namely, "drop," and the number of assumed risks belonging to cluster D as one, namely, "forgotten to wear." The similarity calculation unit 103 ranks clusters A to D in descending order of the number of data items, as shown in FIG. 17, with cluster A ranked first and clusters B, C, and D ranked second. As described above, the correction value is the reciprocal of the rank of the number of data items, so the similarity calculation unit 103 calculates the correction value for cluster A as 1 and the correction values ​​for clusters B, C, and D as 0.5.

[0081] The similarity calculation unit 103 applies the correction values ​​of clusters A to D to the example of Fig. 16, and corrects the similarity scores as shown in Fig. 18. The rank re-sorting unit 108 rearranges the similarity scores corrected by the similarity calculation unit 103 in descending order, and determines the rank of the assumed risks again (step S32).

[0082] As described above, the similar case extraction support system 1 according to the second embodiment clusters assumed risks, corrects the similarity scores according to the number of data items belonging to a cluster, and re-determines the ranking of assumed risks. This makes it possible to reflect important risks that have occurred in the past or for which many cases have been extracted in the ranking, thereby supporting the extraction of similar cases.

[0083] (Modification of the second embodiment) In the second embodiment, the similarity calculation unit 103 calculates the correction value of the similarity score SC using the reciprocal of the rank as illustrated in FIG. 15 , but the present disclosure is not limited to this. For example, the similarity calculation unit 103 may correct the similarity score SC using a value obtained by multiplying a correction coefficient predetermined for each cluster by the number of data items of the assumed risk belonging to the same cluster as the correction value. The formula is expressed as follows: Correction value = Correction coefficient × Number of data items belonging to the same cluster. The similarity calculation unit 103 adds the correction value calculated by this formula to the calculated similarity score SC, thereby calculating the similarity score SC reflecting the correction value.

[0084] (Embodiment 3) 19 to 24, a similar case extraction support system 1 according to a third embodiment will be described. The third embodiment differs from the first embodiment in that the similar case extraction support device 100 further includes a rank re-sorting unit 108, a pre-processing unit 109, a learning unit 110, and an inference unit 111 in addition to the configuration of the first embodiment, and in that the similar case extraction support device 100 executes learning processing, inference processing, and similar case extraction support processing 3. The following description of the third embodiment will focus on the differences from the first and second embodiments.

[0085] As shown in FIG. 19, in addition to the configuration of embodiment 1, the similar case extraction support device 100 further includes a rank re-sorting unit 108 that re-ranks and rearranges the data, a pre-processing unit 109 that standardizes the notation, variation, etc. of the technical document information, a learning unit 110 that creates a trained model for inferring categories, and an inference unit 111 that infers categories using the trained model created by the learning unit 110.

[0086] The ranking re-sorting unit 108 re-ranks and re-sorts the items in accordance with the category designated by the user, as will be described in detail later.

[0087] The preprocessing unit 109 standardizes variations in notation in technical document information. For example, the preprocessing unit 109 converts half-width notation of "connector" into full-width notation of "connector."

[0088] The learning unit 110 executes the learning process in the model creation process shown in FIG. As a prerequisite for the model creation process, the preprocessing unit 109 reads the target technical document information and preprocesses the text. The preprocessing unit 109 standardizes, for example, half-width and full-width characters, uppercase and lowercase letters, and other variations. The preprocessing unit 109 saves the preprocessed technical document information in a folder representing preprocessed identification information in the technical document storage unit 125. The user extracts a portion of the text from the preprocessed technical document information and labels it according to the user's intention. The user labels some of the work risks in the technical document information with causes of occurrence, such as "work-related," "equipment-related," "design-related," and "purchased item-related," as a classification of categories indicating the causes of defects. The work content, countermeasures, key information, etc. associated with the work risk are saved in the technical document storage unit 125 along with the category labeled as the work risk. Note that "work-related," "equipment-related," "design-related," "purchased item-related," etc. are examples of category information related to the present disclosure.

[0089] 20 starts when an operation unit (not shown) in the similar case extraction support device 100 receives an operation by a user to instruct execution. The vector generation unit 102 generates vectors by using BERT to convert sentences indicating task risks labeled by the user and sentences indicating task details, countermeasures, etc. corresponding to the task risks, which are stored in the technical document storage unit 125 (step S41). The vector generation unit 102 stores the generated vectors in the vector information storage unit 122.

[0090] The learning unit 110 uses BERT to learn combinations of vectors corresponding to sentences representing work risks and vectors corresponding to sentences representing categories corresponding to the work risks (step S42). The learning unit 110 stores the trained model created by the training in the trained model storage unit 126 (step S43). Then, the model creation process ends. The learning unit 110 is an example of a training means according to the present disclosure.

[0091] The inference unit 111 infers a category in the inference process shown in Fig. 21. The inference process shown in Fig. 21 is started when an operation unit (not shown) in the similar case extraction support device 100 receives an operation by a user instructing execution of the process after the model creation process is completed. The inference unit 111 is an example of an inference means according to the present disclosure.

[0092] The vector generation unit 102 performs vector conversion using BERT on sentences representing the remaining task risks not labeled by the user, which are stored in the technical document storage unit 125, and sentences representing the task details, countermeasures, etc. corresponding to the task risks, to generate vectors (step S51). The vector generation unit 102 stores the generated vectors in the vector information storage unit 122.

[0093] The inference unit 111 infers a category by inputting vectors corresponding to the sentences representing the remaining task risks that were not labeled by the user into the trained model stored in the trained model storage unit 126 (step S52). The inferred category is stored in the vector information storage unit 122 together with the vectors of the sentences representing the corresponding task risks (step S53). Then, the inference process ends.

[0094] In the third embodiment, similar case extraction support processing 3 shown in Fig. 22 is performed. Steps S21 to S24 and steps S25 and S26 are the same as those in the similar case extraction support processing 1 in the first embodiment. As shown in Fig. 22, in the similar case extraction support processing 3 in the third embodiment, after step S24, steps S61 and S62 are performed, and then step S25 is performed.

[0095] After step S24, the rank resorting unit 108 determines whether or not a category has been designated by the user (step S61). The rank resorting unit 108 determines whether or not a category designation by the user has been accepted by an operation unit (not shown) of the similar case extraction support device 100. If it is determined that a category has been designated by the user (step S61: Yes), the rank resorting unit 108 calculates the rank again (step S62). If it is determined that a category has not been designated by the user (step S61: No), the processing of step S61 is repeated.

[0096] The processing of steps S61 and S62 will be described in detail with reference to Figures 23 and 24. Assume that in the processing of step S24, a ranking is determined as shown in Figure 23. In step S61, for example, when the categories "task-induced" and "purchased item-induced" are specified, the rank re-sorting unit 108 sorts and ranks the combinations of task content, task risks, and countermeasures belonging to the "task-induced" category in descending order of similarity score, as shown in Figure 24(a). Furthermore, the rank re-sorting unit 108 sorts and ranks the combinations of task content, task risks, and countermeasures belonging to the "purchased item-induced" category in descending order of similarity score, as shown in Figure 24(b).

[0097] As described above, the similar case extraction support system 1 according to the third embodiment re-determines the ranking for each category specified by the user. This makes it possible to support the extraction of combinations of task risks and task details that have a higher correlation with the search input information.

[0098] (Modification of the third embodiment) In the third embodiment, the user specifies a category, but the specified category may be reflected in the similarity score SC. For example, if the user specifies "task-related" from among multiple categories, the similarity calculation unit 103 may re-calculate the similarity score SC by setting a correction coefficient of 1 for the specified category "task-related" and setting correction coefficients of 0.5 for the other categories. This allows adjustment to result in a category that is prioritized being ranked higher.

[0099] Although the preferred embodiments have been described in detail above, the present invention is not limited to the above-described embodiments, and various modifications and substitutions can be made to the above-described embodiments without departing from the scope of the claims.

[0100] Various aspects of the present disclosure are summarized below as appendices.

[0101] (Appendix 1) a storage means for storing sentences representing the contents of correlated items and content sentence vectors obtained by converting the sentences into vectors; a vector generator that converts sentences into vectors using a transformer-based model; a calculation means for calculating a similarity between an input sentence vector representing a vector of the input sentence converted by the vector generation means and the content sentence vector; a ranking specifying means for determining a ranking of the sentences expressing the content according to the similarity; a display means for displaying the results including the sentence representing the content, the similarity, and the ranking; Equipped with A support system for extracting similar cases. (Appendix 2) The above matters include the work involved in the production and manufacturing of products and the risks involved in such work. The similar case extraction support system described in Appendix 1. (Appendix 3) the calculation means calculates a similarity between a work content vector representing a content of the work or a risk vector representing a content of the risk and the input sentence vector; The similar case extraction support system described in Appendix 2. (Appendix 4) the rank specifying means determines the rank of the combination of the work and the risk having a correlation, which has the maximum similarity, as number 1; the calculation means calculates a similarity between the work in the combination determined to be first by the rank determination means and other work stored in the storage means, and a similarity between the risk in the combination determined to be first by the rank determination means and other risks stored in the storage means, further comprising a ranking re-identification means for re-determining the ranking of the combination of the work and the risk in accordance with the similarity determined by the calculation means; The similar case extraction support system described in Appendix 3. (Appendix 5) A clustering unit that classifies the risks into clusters, The calculation means corrects the similarity in accordance with the number of the risks belonging to the cluster. 5. A similar case extraction support system according to any one of appendices 2 to 4. (Appendix 6) a learning means for generating a trained model by learning combinations of some of the risks stored in the storage means and category information representing each factor of the risks; and an inference means for inferring the category information using each remaining risk stored in the storage means and the trained model, the rank re-identification means re-determines the rank of the combination of the work and the risk for each category information specified by a user among the category information. 6. A similar case extraction support system according to claim 4 or 5. (Appendix 7) The items include a plurality of pairs of a first item and a second item that are correlated with each other; the storage means stores a pair of a first sentence representing a first item and a second sentence representing a second item, a first vector of the first sentence, and a second vector of the second sentence; the input sentence is a sentence that specifies a first matter, the calculation means calculates a similarity between the vector of the input sentence and the vector of the first sentence of each of the pairs; the ranking specification means specifies the top n (n is a natural number) pairs of first sentences and second sentences in descending order of the similarity calculated by the calculation means, the calculation means calculates a similarity between a vector of a first sentence in a pair ranked in the top n similarity ranking and a vector of a first sentence in another pair, and a similarity between a vector of a second sentence in a pair ranked in the top n similarity ranking and a vector of a second sentence in another pair, a ranking re-identification means for re-determining the ranking of the pair of the first item and the second item according to the similarity calculated by the calculation means; The display means outputs the top m (m is a natural number) pairs identified by the rank re-identification means. 7. A similar case extraction support system according to any one of appendices 1 to 6. (Appendix 8) storing content sentence vectors in which sentences expressing the contents of a plurality of items are converted into vectors; Convert the input sentence into a vector, Calculating the similarity between the input sentence vector representing the converted vector and the content sentence vector; displaying a sentence representing the content according to the similarity; Similar case extraction support method. (Appendix 9) On the computer, storing content sentence vectors in which sentences expressing the contents of a plurality of items are converted into vectors; Convert the input sentence into a vector, Calculating the similarity between the input sentence vector representing the converted vector and the content sentence vector; displaying a sentence representing the content according to the similarity; A program that executes a process. [Explanation of symbols]

[0102] 1 Similar case extraction support system, 100 Similar case extraction support device, 101 Search input information acquisition unit, 102 Vector generation unit, 103 Similarity calculation unit, 104 Search unit, 105 Rank sorting unit, 106 Transmission unit, 107 Clustering unit, 108 Rank re-sorting unit, 109 Preprocessing unit, 110 Learning unit, 111 Inference unit, 120 First memory unit, 121 Search input information storage unit, 122 Vector information storage unit, 123 Calculation information storage unit, 124, 232 Search result information storage unit, 125 Technical document storage unit, 126 Trained model storage unit, 200 User terminal device, 201 Registration reception unit, 202 Search input specific information reception unit, 203 Search input information sending unit, 204 Search result information acquisition unit, 205 Search result information output unit, 206 Display unit, 230 Second memory unit, 231 Search input information storage unit, 232 search result information storage unit, 1000 bus, 1001 processor, 1002 memory, 1003 interface, 1004 secondary storage device.

Claims

1. a storage means for storing sentences representing the contents of correlated items and content sentence vectors obtained by converting the sentences into vectors; a vector generator that converts sentences into vectors using a transformer-based model; a calculation means for calculating a similarity between an input sentence vector representing a vector of the input sentence converted by the vector generation means and the content sentence vector; a ranking specifying means for determining a ranking of the sentences expressing the content according to the similarity; a display means for displaying the results including the sentence representing the content, the similarity, and the ranking; Equipped with A support system for extracting similar cases.

2. The above matters include the work involved in the production and manufacturing of products and the risks involved in such work. The similar case extraction support system according to claim 1 .

3. the calculation means calculates a similarity between a work content vector representing a content of the work or a risk vector representing a content of the risk and the input sentence vector; The similar case extraction support system according to claim 2 .

4. the rank specifying means determines the rank of the combination of the work and the risk having a correlation, which has the maximum similarity, as first; the calculation means calculates a similarity between the work in the combination determined to be first by the ranking determination means and other work stored in the storage means, and a similarity between the risk in the combination determined to be first by the ranking determination means and other risks stored in the storage means, further comprising a ranking re-identification means for re-determining the ranking of the combination of the work and the risk in accordance with the similarity determined by the calculation means; The similar case extraction support system according to claim 3 .

5. A clustering unit that classifies the risks into clusters, The calculation means corrects the similarity in accordance with the number of the risks belonging to the cluster. The similar case extraction support system according to claim 2 .

6. a learning means for generating a trained model by learning combinations of some of the risks stored in the storage means and category information representing each factor of the risks; and an inference means for inferring the category information using each remaining risk stored in the storage means and the trained model, the rank re-identification means re-determines the rank of the combination of the work and the risk for each category information specified by a user among the category information. The similar case extraction support system according to claim 4.

7. The items include a plurality of pairs of a first item and a second item that are correlated with each other; the storage means stores a pair of a first sentence representing a first item and a second sentence representing a second item, a first vector of the first sentence, and a second vector of the second sentence; the input sentence is a sentence that specifies a first item, the calculation means calculates a similarity between the vector of the input sentence and the vector of the first sentence of each of the pairs; the ranking specification means specifies pairs of first sentences and second sentences in the top n (n is a natural number) ranking in descending order of the similarity calculated by the calculation means; the calculation means calculates a similarity between a vector of a first sentence in a pair ranked in the top n similarity ranking and a vector of a first sentence in another pair, and a similarity between a vector of a second sentence in a pair ranked in the top n similarity ranking and a vector of a second sentence in another pair, a ranking re-identification means for re-determining the ranking of the pair of the first item and the second item according to the similarity calculated by the calculation means; the display means outputs the top m (m is a natural number) pairs identified by the rank re-identification means. The similar case extraction support system according to claim 1 .

8. storing content sentence vectors in which sentences expressing the contents of a plurality of items are converted into vectors; Convert the input sentence into a vector, Calculating the similarity between the input sentence vector representing the converted vector and the content sentence vector; displaying a sentence representing the content according to the similarity; Similar case extraction support method.

9. On the computer, storing content sentence vectors in which sentences expressing the contents of a plurality of items are converted into vectors; Convert the input sentence into a vector, Calculating the similarity between the input sentence vector representing the converted vector and the content sentence vector; displaying a sentence representing the content according to the similarity; A program that executes a process.

Images