Computer system and information retrieval method

Abstract

To accurately retrieve a document including information on a structure having a predetermined feature.SOLUTION: A computer system manages a database for storing structure data in a graph format with elements constituting a document as nodes. The document includes information about the structure as knowledge. The knowledge includes a superordinate concept which is knowledge in which a partial structure of a structure is abstractly expressed, and a subordinate concept which is knowledge in which the abstractly expressed structure is embodied. The computer system receives a search query including, as a search key, data corresponding to a superordinate concept or a subordinate concept for search, inputs the search key and one of the superordinate concept and the subordinate concept included in a document to the model, and identifies a document including the superordinate concept or the subordinate concept similar to the search key.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a technique for storing data in which the relationships between elements of a document are structured, and a technique for searching the stored data. [Background technology]

[0002] In materials development at companies, efforts are sometimes made to predict the results of experiments to create new materials before they are conducted by analyzing experimental data extracted from public documents such as patent documents. When extracting experimental data, for example, the raw materials used in the synthesis of a certain material, the experimental procedures, and the experimental conditions are extracted from the documents, organized into searchable structured data, and then stored in a database.

[0003] Attempts to structure searchable data are often done manually and are costly. Therefore, there is a need for a system that can efficiently extract words related to a given case from documents and further associate the extracted words with each other.

[0004] Patent Document 1 describes a device that uses a knowledge graph to appropriately associate hypernyms and hyponyms of a word. A knowledge graph is constructed by extracting named entities from documents such as patents and papers, identifying the relationships between the named entities, and graphing them. A word in a document is designated as a hypernym, and its hyponym is searched for in the knowledge graph, thereby associating the hypernym with the hyponym. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] International Publication No. 2022 / 208822 Summary of the Invention [Problem to be solved by the invention]

[0006] The technology in Patent Document 1 uses graph data that represents the relationships between words, making it difficult to search for information about a single, cohesive structure such as a compound. This is particularly difficult when the structure's partial configuration is expressed abstractly.

[0007] An object of the present invention is to provide a technique for accurately searching for documents containing information about structures having predetermined characteristics. [Means for solving the problem]

[0008] A representative example of the invention disclosed in the present application is as follows: That is, a computer system including a processor, a storage device connected to the processor, and a network interface connected to the processor manages a database storing structure data in a graph format in which elements constituting documents are nodes, the documents including information about structures as knowledge, the knowledge including a higher-level concept that is knowledge that abstractly expresses a portion of the structure of the structure, and a lower-level concept that is knowledge that embodies the abstractly expressed structure, maintains a machine learning model that accepts the lower-level concept and the higher-level concept as input and outputs a score that represents the similarity between the input lower-level concept and the input higher-level concept, accepts a search query that includes data corresponding to the higher-level concept or the lower-level concept for search as a search key, inputs the search key and either the higher-level concept or the lower-level concept included in each document stored in the database to the machine learning model, and identifies documents that include the higher-level concept or the lower-level concept that is similar to the search key based on the score obtained from the machine learning model. [Effects of the Invention]

[0009] According to the present invention, it is possible to accurately search for documents containing information about structures having predetermined characteristics. Problems, configurations, and effects other than those described above will become apparent from the following description of the embodiments. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 illustrates an example of a system configuration according to a first embodiment. [Figure 2] FIG. 1 is a diagram illustrating an example of an information processing device that realizes a device constituting a system according to a first embodiment. [Figure 3] FIG. 2 is a diagram illustrating the flow of processing executed by the data management / search device according to the first embodiment. [Figure 4] FIG. 4 is a diagram illustrating an example of a data structure of document management information according to the first embodiment. [Figure 5] FIG. 2 is a diagram showing an example of a data structure of structure data according to the first embodiment. [Figure 6] 10 is a flowchart illustrating an example of a structured data generation process S1 executed by the data management / search device according to the first embodiment. [Figure 7] FIG. 2 is a diagram showing an example of a data structure of superordinate concept data according to the first embodiment. [Figure 8] FIG. 1 is a diagram illustrating an example of conversion from a Markush structure to a molecular graph. [Figure 9] FIG. 2 is a diagram showing an example of a data structure of subordinate concept data according to the first embodiment. [Figure 10] FIG. 10 is a diagram illustrating an example of a data structure of superordinate concept-subordinate concept pair data according to the first embodiment. [Figure 11] 10 is a flowchart illustrating an example of a similar knowledge determination model generation process S2 executed by the data management / search device according to the first embodiment. [Figure 12] FIG. 10 is an image diagram of a similar knowledge determination model generation process S2 executed by the data management / search device according to the first embodiment. [Figure 13A] FIG. 10 is a diagram illustrating an example of a data structure of a search query according to the first embodiment. [Figure 13B] FIG. 10 is a diagram illustrating an example of a data structure of a search query according to the first embodiment. [Figure 14] FIG. 10 is a diagram illustrating an example of a data structure of search target-search query pair data according to the first embodiment. [Figure 15] 10 is a flowchart illustrating an example of a knowledge search process S3 executed by the data management / search device according to the first embodiment. [Figure 16] 10 is a flowchart illustrating an example of a similar knowledge determination process executed by the data management / search device according to the first embodiment. [Figure 17] FIG. 2 is an image diagram of a similar knowledge determination process executed by the data management / search device according to the first embodiment. [Figure 18A] FIG. 10 is a diagram showing an example of a search screen displayed on a user terminal 2 according to the first embodiment. [Figure 18B] FIG. 10 is a diagram showing an example of a search screen displayed on a user terminal 2 according to the first embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, embodiments of the present invention will be described with reference to the drawings. The following description and drawings are examples for explaining the present invention, and some omissions and simplifications have been made as appropriate for clarity of explanation. The present invention can be implemented in various other forms. Unless otherwise specified, each component may be singular or plural.

[0012] In the following explanation, the same or similar components may be assigned the same reference numerals, and redundant explanations may be omitted. Furthermore, in the following explanation, the letter "S" before a reference numeral indicates a processing step. Furthermore, in the following explanation, various types of information may be described using expressions such as "table" and "information," but the information may also be expressed using other data structures.

[0013] Other challenges include the following: In chemical literature, some of the compositions of raw materials used in material synthesis include information specific to chemical systems, such as Markush structures, IUPAC names, and SMILES format. This information is structurized manually. For example, to structure information shown in an image, it is necessary to refer to and organize the image information and text related to the image.

[0014] The prior art such as Patent Document 1 is a technology that targets natural language, and therefore it is difficult to structure documents that include information other than natural language into a searchable data format. The present invention aims to provide a technology that also structures documents that include information other than text into a searchable data format. [Example]

[0015] FIG. 1 is a diagram illustrating an example of a system configuration according to a first embodiment.

[0016] This system 1 extracts elements such as words and images from documents such as chemical literature, determines whether there is a semantic relationship between the elements, and based on the determination results, generates graph-like structure data 120 that connects words that have a semantic relationship, and stores the data in a structure data DB 118. The system 1 also accepts input of a search query 21 from a user and searches for documents containing desired information.

[0017] As shown in Fig. 1, system 1 includes a data management / retrieval device 100, a user terminal 2, and a data management device 4. All of these are configured using an information processing device 10 (see Fig. 2). Data management / retrieval device 100 and data management device 4 are connected to each other via a communication medium 5 so that they can communicate with each other bidirectionally. Communication medium 5 is a communication infrastructure that provides wired or wireless communication functions, and is, for example, various communication networks such as a LAN (Local Area Network), a WAN (Wide Area Network), the Internet, a dedicated line, various public communication networks, various communication devices that comply with the USB (Universal Serial Bus) standard, etc.

[0018] The data management / search device 100 extracts elements from a document, determines whether there is a relationship between the elements, and generates structured data 120 by connecting the elements based on the determination result, and stores the structured data 120 in the structured data DB 118.

[0019] The documents handled by the system 1 are, for example, documents such as patent documents or papers describing polymer synthesis processes, and are managed as electronic data.

[0020] The document may be any document that describes knowledge of a specific domain. The document includes a superordinate concept, which is knowledge expressed abstractly, and a subordinate concept, which is knowledge that specifically expresses the abstract structure. For example, in the case of a chemistry document that includes knowledge of chemical compounds, the superordinate concept is information about the compound (structure) in which part of the structure is expressed abstractly, and the subordinate concept is the specific structure itself and information about the compound (structure) in which the structure is specifically expressed.

[0021] The structural data 120 can be used, for example, as information that provides hints for designing polymeric materials, or as training data for data-driven material development.

[0022] The data management / search device 100 includes, as its main functions, a storage unit 110, a structured data generation unit 130, a knowledge integration unit 140, a similar knowledge determination model generation unit 150, and a knowledge search unit 160.

[0023] The storage unit 110 stores document management information 111, a structured data schema 112, a similar knowledge determination model 113, superordinate concept data 114, subordinate concept data 115, superordinate concept-subordinate concept pair data 116, search target-search query pair data 117, and a structured data DB 118. These will be described in detail later.

[0024] The user terminal 2 provides a user interface for performing operations on various types of information managed by the data management / search device 100. The user terminal 2 provides a user interface for referring to or editing, for example, the document management information 111, the structured data 120, and the structured data schema 112. The user terminal 2 also transmits a search query 21 received from the user to the data management / search device 100. The search query 21 includes a search key for searching the structured data DB 118 for documents containing desired knowledge. Data such as characters or images corresponding to either a superordinate concept or a subordinate concept is used as the search key.

[0025] The data management device 4 includes a data management unit 41. The data management unit 41 manages a document data DB 42 that stores document data, which is an electronic document. The data management unit 41 also communicates with the data management / search device 100 and transmits document data to the data management / search device 100. The data management device 4 acquires the document data stored in the document data DB 42, for example, via a communication medium 5. The document data can also be registered by the user via a user interface provided by the user terminal 2.

[0026] FIG. 2 is a diagram illustrating an example of an information processing device that realizes the devices constituting the system according to the first embodiment.

[0027] The information processing device 10 includes a processor 11, a main memory device 12, an auxiliary memory device 13, an input device 14, an output device 15, and a communication device 16. Examples of the information processing device 10 include a personal computer, a server device, a smartphone, and a tablet.

[0028] All or part of the information processing device 10 may be realized using virtual information processing resources provided using virtualization technology, process space separation technology, and the like, such as a virtual server provided by a cloud system. Furthermore, all or part of the functions provided by the information processing device 10 may be realized by a service provided by a cloud system via an API (Application Programming Interface), for example. Furthermore, all or part of the functions provided by the information processing device 10 may be realized using, for example, Software as a Service (SaaS), Platform as a Service (PaaS), Infrastructure as a Service (IaaS), or the like.

[0029] The data management / search device 100 and the user terminal 2 may be realized by an information processing device 10 having the same hardware configuration. Furthermore, the data management / search device 100 may be realized by using, for example, a plurality of information processing devices 10 connected to each other so as to be able to communicate with each other.

[0030] The processor 11 shown in the figure is configured using, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), an AI (Artificial Intelligence) chip, etc.

[0031] The main memory device 12 is a device that stores programs and data, and is, for example, a read-only memory (ROM), a random access memory (RAM), or a non-volatile memory (NVRAM).

[0032] The auxiliary storage device 13 is, for example, an optical storage device such as an SSD (Solid State Drive), a hard disk drive, a CD (Compact Disc) or a DVD (Digital Versatile Disc), a storage system, a reading / writing device for recording media such as an IC card, an SD card or an optical recording medium, or a storage area of ​​a cloud server. Programs and data can be read into the auxiliary storage device 13 via a recording medium reading device and a communication device 16. The programs and data stored in the auxiliary storage device 13 are read into the main storage device 12 as needed.

[0033] The input device 14 is an interface that accepts input from the outside, and is, for example, a keyboard, a mouse, a touch panel, a card reader, a pen-input tablet, a voice input device, or the like.

[0034] The output device 15 is an interface that outputs various information such as the progress of processing and the results of processing. The output device 15 is, for example, a display device such as a liquid crystal monitor, LCD (Liquid Crystal Display), or graphic card that visualizes the various information, an audio output device such as a speaker that converts the various information into audio, or a printer that converts the various information into text. Note that, for example, the information processing device 10 may be configured to input and output information to and from other devices via the communication device 16.

[0035] The input device 14 and the output device 15 constitute a user interface that realizes interactive processing with the user (receiving information, presenting information, etc.).

[0036] The communication device 16 is a device that realizes communication with other devices. The communication device 16 is a wired or wireless communication interface that realizes communication with other devices via the communication medium 5, and is, for example, a NIC (Network Interface Card), a wireless communication module, a USB module, or the like.

[0037] The information processing device 10 may be equipped with, for example, an operating system, a file system, a DBMS (DataBase Management System) such as a relational database or NoSQL, and a KVS (Key-Value Store).

[0038] The functions of the data management / search device 100, the user terminal 2, and the data management device 4 are realized by the respective processors 11 reading and executing programs stored in the respective main memories 12, or by the hardware (FPGA, ASIC, AI chip, etc.) that constitute these devices themselves.

[0039] The various functions provided by the data management / search device 100 are realized using, for example, various known data mining methods such as text data mining, various known natural language processing methods (morphological analysis, syntactic analysis, semantic analysis, context analysis, feature extraction, word embedding, named entity extraction, text classification, sequence labeling), and various known machine learning methods (deep learning (DNN: Deep Neural Network), RNN: Recurrent Neural Network, etc.). The data management / search device 100 stores the various pieces of information described above, for example, as tables in a database or files managed by a file system.

[0040] FIG. 3 is a diagram illustrating the flow of processing executed by the data management / search device 100 according to the first embodiment.

[0041] The processes executed by the data management / search device 100 are mainly three processes: structured data generation process S1, similar knowledge determination model generation process S2, and knowledge search process S3.

[0042] In the structured data generation process S1, the data management / search device 100 generates knowledge contained in the document, for example, information on a polymer material synthesis procedure, as structured data 120 in a graph format, and stores it in the structured data DB 118.

[0043] In the similar knowledge determination model generation process S2, the data management / search device 100 generates a similar knowledge determination model 113 for searching structured data 120 (documents) including a superordinate or subordinate concept similar to the search key included in the search query 21.

[0044] In the knowledge search process S3, the data management / search device 100 uses the similar knowledge determination model 113 to search for structured data 120 that includes a superordinate or subordinate concept similar to the search key included in the search query 21.

[0045] The following explanation focuses on three processes: structural data generation process S1, similar knowledge determination model generation process S2, and knowledge search process S3. Note that in the examples, explanations will be given using Patent Application No. 2013-544157, which is a document containing knowledge about polymer material synthesis, as an example.

[0046] The structure data generation process S1 will now be described. First, the data structure of the information and data handled in the structure data generation process S1 will be described.

[0047] FIG. 4 is a diagram illustrating an example of the data structure of the document management information 111 according to the first embodiment.

[0048] The document management information 111 stores entries each consisting of an ID 1111, a Patent_number 1112, and a content 1113. One entry corresponds to one piece of document data. Note that the fields included in an entry are merely an example and are not limited to these.

[0049] ID 1111 is a field for storing the ID of the document data. Patent_number 1112 is a field for storing the patent application number, which is identification information assigned to a patent document (document). Content 1113 is a field for storing the content of the document. The content 1113 stores the entity of the document data.

[0050] Although this embodiment is based on the assumption that the system is for managing patent documents, the data structure of the document management information 111 for managing documents other than patent documents can be changed as appropriate.

[0051] FIG. 5 is a diagram showing an example of the data structure of the structure data 120 according to the first embodiment.

[0052] The structure data 120 is a directed graph made up of a plurality of nodes and a plurality of links.

[0053] A node has variables such as Node_ID, Class, Name, Text, File, and Conditions. Node_ID is the ID of the node. Class is the content of the node. Name is the entity in the document corresponding to the node. Text indicates the sentence corresponding to the node. File indicates the file related to the node. Conditions indicates the conditions related to the node.

[0054] For example, a node whose Class is "Process" represents an experimental operation, and "Conditions" represents the experimental conditions.

[0055] Links connecting nodes are given attributes such as Has_component, Input, Output, and Next. Has_component is an attribute that indicates the containment relationship between nodes. Input and Output respectively indicate the input and output for a Process node. Next indicates the next experimental operation.

[0056] In the structured data generation process S1, knowledge contained in patent documents, particularly experimental procedures / conditions, is structured as a directed graph of the above-mentioned schema. The schema of the structured data 120 is managed by the structured data schema 112. The structured data schema 112 can be edited by a user using a user terminal 2 or the like.

[0057] The schema of the structure data 120 shown in FIG. 5 is an example, and the schema can be defined appropriately depending on the knowledge to be structured.

[0058] The structured data generation process S1 will now be described in detail with reference to Fig. 6. Fig. 6 is a flowchart illustrating an example of the structured data generation process S1 executed by the data management / search device 100 according to the first embodiment.

[0059] The structure data generation process S1 is started when an execution instruction is received from a user via the user terminal 2, when a scheduled timing arrives, or when new document data is added to the document management information 111. Here, the process executed for one piece of document data will be described.

[0060] The structured data generating unit 130 extracts elements such as words and images from document data using known named entity recognition (NER) technology and image recognition technology (step S11).

[0061] The structure data generation unit 130 determines the relationships between elements (step S12). For example, the structure data generation unit 130 determines the relationships between words using a known relation extraction technique (RE: Relation Extraction). The structure data generation unit 130 also determines the relationships between images and words by converting the images into text using a known image recognition technique.

[0062] The structure data generator 130 generates structure data 120 representing the relationships between the elements based on the determination result of the relationships between the elements (step S13).

[0063] The knowledge integration unit 140 updates the structure data 120 by integrating the generated structure data 120 with the structure data 120 stored in the structure data DB 118 (step S14). One possible integration method is to combine the generated structure data 120 with the structure data 120 stored in the structure data DB 118 using nodes containing the same word as a key, but the integration method is not limited to one and may be determined arbitrarily. Note that integration of the structure data 120 does not necessarily have to be performed.

[0064] The similar knowledge determination model generation process S2 will be described below. First, the data structure of the information and data handled in the similar knowledge determination model generation process S2 will be described.

[0065] Fig. 7 is a diagram showing an example of the data structure of the superordinate concept data 114 in Example 1. Fig. 8 is a diagram showing an example of conversion from a Markush structure to a molecular graph.

[0066] The superordinate concept data 114 is data for managing superordinate concepts, which are abstracted domain knowledge. In the case of a system that handles patent documents, the contents of claims are treated as superordinate concepts. FIG. 7 shows the superordinate concept data 114 generated from the contents of claims of Japanese Patent Application No. 2013-544157. The superordinate concept data 114 includes an ID 1141, a Patent_number 1142, a Claim 1143, a Material 1144, and a Definition 1145.

[0067] ID 1141 is a field that stores the ID of the generic concept data 114. Patent_number 1142 is a field that stores identification information of a patent document that includes a generic concept corresponding to the generic concept data 114. Claim 1143 is a field that stores the number of the claim that defines the generic concept.

[0068] Material 1144 is a field for storing a Markush structure representing a compound as described in the claims. Material 1144 may store the Markush structure as an image, or may store a molecular graph converted from the Markush structure as shown in FIG. 8. For example, a method of converting the Markush structure into text using SMILES notation is conceivable. FIG. 7 shows text converted from a Markush structure using SMILES notation.

[0069] It should be noted that the SMILES notation cannot represent the repetition of molecular structures in Markush structures, and therefore information on this point is missing.

[0070] Definition 1145 is a field for storing text defining the abstract molecular structure of the Markush structure described in the claims. In the example of Fig. 7, definitions of the abstracted "M" and "Q" in the molecular structure are described.

[0071] The superordinate concept data 114 can be generated by merging information of nodes in the hierarchy below the node whose Class in the structure data 120 is "Claim."

[0072] The data structure of the superordinate concept data 114 is merely an example and is not limiting, and any data structure that can express knowledge as a superordinate concept may be used.

[0073] FIG. 9 is a diagram showing an example of the data structure of the subordinate concept data 115 according to the first embodiment.

[0074] The lower concept data 115 is data for managing lower concepts, which are embodied domain knowledge. In the case of a system that handles patent documents, the contents of the examples are treated as lower concepts. FIG. 9 shows the lower concept data 115 generated from the contents of the examples of Japanese Patent Application No. 2013-544157. The lower concept data 115 includes ID 1151, Patent_number 1152, Embodiment 1153, Paragraph_number 1154, Instance_candidate_from_text 1155, and Text 1156.

[0075] ID 1151 is a field that stores the ID of the subordinate concept data 115. Patent_number 1152 is a field that stores identification information of a patent document that includes a subordinate concept that corresponds to the subordinate concept data 115. Embodiment 1153 is a field that stores the number of the embodiment in which the subordinate concept is described. Paragraph_number 1154 is a field that stores the paragraph in which the subordinate concept is specifically described.

[0076] Instance_candidate_from_text 1155 is a field that stores candidates for specific information of abstracted information in a claim. For example, in the claim of Japanese Patent Application No. 2013-544157, compounds corresponding to "M" and "Q" shown in the Markush structure are abstracted. In the example of Figure 9, "PSSNa" is stored in Instance_candidate_from_text 1155 as a candidate for a specific compound corresponding to "M" and "Q."

[0077] Text 1156 is a field that stores the text written in the paragraph corresponding to Paragraph_number 1154.

[0078] The lower concept data 115 can be generated by merging information of nodes in the hierarchy below the node whose Class of the structure data 120 is "Embodiment".

[0079] The data structure of the lower concept data 115 is merely an example and is not limited to this, and any data structure that can express knowledge that is a lower concept may be used.

[0080] FIG. 10 is a diagram illustrating an example of the data structure of the superordinate concept-subordinate concept pair data 116 according to the first embodiment.

[0081] The superordinate concept-subordinate concept pair data 116 is training data used for learning the similar knowledge determination model 113. Fig. 10 shows the superordinate concept-subordinate concept pair data 116 generated from Japanese Patent Application No. 2013-544157. The superordinate concept-subordinate concept pair data 116 includes an ID 1161, a Patent_number 1162, a Claim 1163, a Material 1164, a Definition 1165, an Instance_candidate_from_text 1166, a Text 1167, and a Label 1168.

[0082] In this embodiment, the data structure is as described above in order to perform contrastive learning using paired data that associates the Markush structure and its definition statement described in the claims with the substance information of the embodiment.

[0083] ID 1161 is a field that stores the ID of the superordinate concept-subordinate concept pair data 116. Patent_number 1162, Claim 1163, Material 1164, and Definition 1165 are the same fields as those described in FIG. 7. Instance_candidate_from_text 1166 and Text 1167 are the same fields as those described in FIG. 9.

[0084] Label 1168 is a field that stores a label indicating whether the content of Text 1167 is related to the content of Definition 1165. In other words, information indicating whether the subordinate concept that constitutes the pair corresponds to information that embodies the superordinate concept is stored. If the subordinate concept that constitutes the pair corresponds to information that embodies the superordinate concept, "True" is stored, and if not, "False" is stored.

[0085] The data structure of the superordinate concept-subordinate concept pair data 116 is merely an example and is not limited to this.

[0086] The similar knowledge determination model generation process S2 will be described in detail. Fig. 11 is a flowchart illustrating an example of the similar knowledge determination model generation process S2 executed by the data management / search device 100 of the first embodiment. Fig. 12 is an image diagram of the similar knowledge determination model generation process S2 executed by the data management / search device 100 of the first embodiment.

[0087] The similar knowledge determination model generation process S2 is started when an execution instruction is received from the user via the user terminal 2 or when the structure data generation process S1 is completed.

[0088] In the similar knowledge determination model generation process S2, a similar knowledge determination model 113 is generated by comparative learning using superordinate concept data 114 and subordinate concept data 115. The superordinate concept data 114 and subordinate concept data 115 can be defined appropriately depending on the information to be treated as knowledge. For example, when extracting information from chemical patent documents, it is often the case that superordinate concepts are described in claims and subordinate concepts are described in examples, so a similar knowledge determination model 113 is generated by comparative learning of claims and examples.

[0089] The similar knowledge determination model 113 is, for example, a neural network that receives a superordinate concept and a subordinate concept as input, and outputs a similarity indicating whether the subordinate concept is information that embodies the superordinate concept.

[0090] The similar knowledge determination model generation unit 150 generates superordinate concept data 114 from the structure data 120 (step S21), and also generates subordinate concept data 115 from the structure data 120 (step S22). The superordinate concept data 114 and subordinate concept data 115 may be generated, for example, based on a preset schema. Alternatively, a user may refer to the structure data 120 and manually set the superordinate concept data 114 and subordinate concept data 115.

[0091] In this example, the information described in the claims is considered to be a generic concept, and the information described in the examples is considered to be a subordinate concept. This is based on a standard specific to patent documents, in which abstract information in claims is often specifically described in the examples.

[0092] The generation of the superordinate concept data 114 and the subordinate concept data 115 may be executed in the structure data generation process S1.

[0093] The similar knowledge determination model generation unit 150 determines the relationship between the superordinate concept data 114 and the subordinate concept data 115 (step S23).

[0094] In this embodiment, the similar knowledge determination model generation unit 150 determines whether or not the information in the Instance_candidate_from_text 1155 of the lower concept data 115 is specific information (instance) of the abstracted information in the Definition 1145 of the higher concept data 114 for all combinations of the higher concept data 114 and the lower concept data 115.

[0095] Based on the result of the determination, the similar knowledge determination model generation unit 150 generates the superordinate concept-subordinate concept pair data 116 (step S24). The determination result of step S23 is set in the label 1168 of the superordinate concept-subordinate concept pair data 116.

[0096] The similar knowledge determination model generation unit 150 performs contrastive learning of the similar knowledge determination model 113 using the superordinate concept-subordinate concept pair data 116 as training data (step S25).

[0097] Specifically, the similar knowledge determination model generation unit 150 refers to the Label 1168 of the superordinate concept-subordinate concept pair data 116, and converts it to "1" if it is True, and converts it to "0" if it is False. The similar knowledge determination model generation unit 150 performs learning using the Material 1164, the Definition 1165, and the Instance_candidate_from_text 1166 as explanatory variables, and the Label 1168 as a target variable.

[0098] The similar knowledge determination model generation unit 150 stores the generated similar knowledge determination model 113 in the storage unit 110 (step S26).

[0099] The knowledge search process S3 will now be described. First, the data structure of the information and data handled in the knowledge search process S3 will be described.

[0100] 13A and 13B are diagrams showing an example of the data structure of the search query 21 according to the first embodiment.

[0101] The search query 21 includes an ID 211 , a Type 212 , and a Query 213 .

[0102] ID211 is a field that stores the ID of the search query 21. Type212 is a field that stores the data format of the search key of the search query 21. Either "Text" or "Image" is stored in Type212. Note that the data format may be other than these. Query213 is a field that stores the search key of the search query 21. Query213 stores a word or an image. Note that the content of the search query 21 may be other than these.

[0103] FIG. 14 is a diagram illustrating an example of the data structure of the search target_search query pair data 117 according to the first embodiment.

[0104] The search target_search query pair data 117 is data input to the similar knowledge determination model 113. Fig. 13 shows the search target_search query pair data 117, which searches for the superordinate concept data 114 generated from the contents of the embodiment of Japanese Patent Application No. 2013-544157. The search target_search query pair data 117 includes ID 1171, Target_ID 1172, Patent_number 1173, Claim 1174, Material 1175, Definition 1176, Query ID 1177, Type 1178, Query 1179, and Score 1180.

[0105] The ID 1171 is a field that stores the ID of the search target_search query pair data 117. The Target_ID 1172 is a field that stores the ID of the search target. In this embodiment, the Target_ID 1172 stores the ID of the superordinate concept data 114. The query ID 1177 is a field that stores the ID of the search query 21.

[0106] The Patent_number 1173, Claim 1174, Material 1175, and Definition 1176 are the same fields as those described in FIG.

[0107] Type 1178 and Query 1179 are the same as the fields described in Figures 12A and 12B. If the content of the search query 21 is an image such as a Markush structure, text using SMILES notation is stored in Query 1179 as described in Figure 8.

[0108] Score 1180 is a field for storing a score calculated based on the output of the similar knowledge determination model 113 .

[0109] The knowledge search process S3 will now be described in detail with reference to Fig. 15, which is a flowchart illustrating an example of the knowledge search process S3 executed by the data management / search device 100 according to the first embodiment.

[0110] The knowledge search process S3 is started when a search query 21 is received via the user terminal 2 or the like.

[0111] The similar knowledge determination unit 161 of the knowledge search unit 160 executes a similar knowledge determination process using the similar knowledge determination model 113 (step S31). The similar knowledge determination process will be described in detail later.

[0112] The search result output unit 162 of the knowledge search unit 160 identifies documents containing similar knowledge based on the scores calculated in the similar knowledge determination process, and outputs the documents as search results to the user terminal 2 (step S32).

[0113] Fig. 16 is a flowchart illustrating an example of the similar knowledge determination process executed by the data management / search device 100 according to the first embodiment. Fig. 17 is an image diagram of the similar knowledge determination process executed by the data management / search device 100 according to the first embodiment.

[0114] The similar knowledge determination unit 161 acquires the search query 21 (step S311), and acquires data to be searched from the structure data DB 118 (step S312).

[0115] The data to be searched may be all of the structure data 120 in the structure data DB 118, or may be data extracted from the structure data 120 based on predetermined narrowing conditions.

[0116] In this embodiment, the similar knowledge determination unit 161 extracts superordinate concept data 114 or subordinate concept data 115 from the structure data 120 in accordance with the search query 21. In the case of the search query 21 shown in Figures 13A and 13B, the superordinate concept data 114 is extracted because the subordinate concept is included as a search key. If the superordinate concept is included as a search key, the subordinate concept data 115 is extracted.

[0117] The similar knowledge determination unit 161 generates the search target_search query pair data 117 (Step S313). Specifically, the similar knowledge determination unit 161 generates the search target_search query pair data 117 by combining the search query 21 with the data to be searched.

[0118] The similar knowledge determination unit 161 selects one piece of search target-search query pair data 117 (step S314).

[0119] The similar knowledge determination unit 161 calculates a score representing the similarity between the search target and the search query 21 that constitute the search target-search query pair data 117 using the similar knowledge determination model 113 (step S315).

[0120] 14 , the similar knowledge determination unit 161 inputs Material 1175, Definition 1176, and Query 1179 to the similar knowledge determination model 113. The similar knowledge determination unit 161 sets the output of the similar knowledge determination model 113 to Score 1180 of the search target_search query pair data 117.

[0121] The similar knowledge determining unit 161 determines whether or not the process has been completed for all the search target-search query pair data 117 (step S316).

[0122] If the process has not been completed for all the search target_search query pair data 117, the similar knowledge determination unit 161 returns to step S314 and executes the same process.

[0123] When the process is completed for all the search target_search query pair data 117, the similar knowledge determination unit 161 outputs the search target_search query pair data 117 to the search result output unit 162 (step S317), and ends the similar knowledge determination process.

[0124] In this way, by using the similar knowledge determination model 113, it becomes possible to search for data that requires multimodal analysis between images and text, such as Markush structure.

[0125] The search result output unit 162 may output the search target_search query pair data 117 as the search result as it is, or may output data that has been processed in some way as the search result. For example, the search result output unit 162 sorts the search target_search query pair data 117 in descending order of Score 1180, and outputs the result.

[0126] Next, the screen (user interface) displayed on the user terminal 2 will be described.

[0127] 18A and 18B are diagrams showing examples of a search screen displayed on the user terminal 2 of the first embodiment.

[0128] The search screen 2000 includes an input area 2001 , a search query display area 2002 , a search result display area 2003 , and a knowledge display area 2004 .

[0129] The input area 2001 is an area for inputting various pieces of information to be included in the search query 21. The input area 2001 includes a search button 2011, a text input field 2012, and a file input field 2013.

[0130] The search query display area 2002 is an area where the content of the search query 21 is displayed.

[0131] The search result display area 2003 is an area that displays the search results for the search query 21. The search result display area 2003 includes an output button 2031 and a search result list 2032.

[0132] The knowledge display area 2004 is an area for displaying knowledge contained in the searched document. The knowledge display area 2004 includes a plurality of tabs 2041 and a knowledge display column 2042.

[0133] The user inputs data to be used as a search key in the text input field 2012 or file input field 2013 of the input area 2001 of the search screen 2000. An image file is input in the file input field 2013. When the user presses the search button 2011, the user terminal 2 generates a search query 21 based on the input in the input area 2001 and transmits the search query 21 to the data management / search device 100. This executes the knowledge search process S3. Details of the generated search query 21 are displayed in the search query display area 2002, so the user can check the search query 21 to be transmitted by referring to this area.

[0134] When the user terminal 2 receives the search target_search query pair data 117 from the data management / search device 100, it displays it as a search result list 2032 in the search result display area 2003. In FIGS. 18A and 18B, the contents of some fields are omitted for simplicity of explanation. By pressing an output button 2031, the user can obtain the search result list 2032 as a file in a predetermined data format. For example, a CSV file may be used.

[0135] Furthermore, the user can refer to the knowledge contained in the selected document in the knowledge display area 2004 by operating the checkbox 2033 in the search result list 2032 .

[0136] In the knowledge display area 2004, a plurality of tabs 2041 are displayed for selecting the knowledge to be displayed. When the "synthesis procedure graph" tab 2041 is selected as shown in FIG. 18A, a graph showing the synthesis procedure is displayed in the knowledge display field 2042. When the "compound structure" tab is selected as shown in FIG. 18B, the Markush structure showing the entire compound, its basic skeleton, variables showing abstracted information, variable definition text, etc. are displayed in the knowledge display field 2042. This information can be extracted from the structure data 120. The number of types of tabs can be increased or decreased depending on the use case.

[0137] As described above, the data management / search device 100 can generate and store structured data 120 by structuring documents such as patent documents based on the structured data schema 112. The data management / search device 100 can accurately search for information that requires multimodal analysis, such as compounds with Markush structures, by executing a knowledge search process S3 using the similar knowledge determination model 113.

[0138] Although one embodiment of the present invention has been described above, it goes without saying that the present invention is not limited to the above embodiment and can be modified in various ways without departing from the spirit of the present invention. For example, the above embodiment has been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to an embodiment having all of the described configurations. Furthermore, it is possible to add, delete, or replace part of the configuration of the above embodiment with other configurations.

[0139] Furthermore, the above-described configurations, functional units, processing units, processing means, etc. may be partially or entirely implemented in hardware, for example, by designing them as integrated circuits. The above-described configurations, functions, etc. may also be implemented in software, with a processor interpreting and executing a program that implements each function. Information such as the programs, tables, and files that implement each function can be stored in a memory, a hard disk, a recording device such as an SSD (Solid State Drive), an IC card, an SD card, a DVD, or other recording media.

[0140] Furthermore, the layout of the various functional units, processing units, and databases of each information processing device described above is merely an example, and the layout of the various functional units, processing units, and databases can be changed to an optimal layout in terms of the performance, processing efficiency, communication efficiency, etc. of the hardware and software that these devices are equipped with.

[0141] Furthermore, the configuration (schema, etc.) of the database that stores the various types of data described above can be flexibly changed from the viewpoint of efficient use of resources, improved processing efficiency, improved access efficiency, improved search efficiency, and the like. [Explanation of symbols]

[0142] 1 System 2. User terminal 4 Data management device 5 Communication medium 10. Information processing equipment 11 processors 12 Main storage 13 Auxiliary storage device 14 Input Devices 15 Output Devices 16. Communications equipment 21 search queries 41 Data Management Department 42 Document Data DB 100 Data management / retrieval device 110 Storage section 111 Document Management Information 112 Structural Data Schema 113 Similarity Knowledge Judgment Model 114 Superordinate Concept Data 115 Sub-concept data 116 Superordinate-subordinate pair data 117 Search target_search query pair data 118 Structural Data DB 120 Structural Data 130 Structure data generation unit 140 Knowledge Integration Department 150 Similarity knowledge determination model generation unit 160 Knowledge Search Department 161 Similar Knowledge Judgment Department 162 Search result output section 2000 Search Screen

Claims

1. A computer system comprising a processor, a storage device connected to the processor, and a network interface connected to the processor, Manages a database that stores structured data in a graph format, with the elements that make up a document as nodes, The document contains information about the structure as knowledge; The knowledge includes a superordinate concept, which is knowledge that abstractly expresses a part of the structure of the structure, and a subordinate concept, which is knowledge that embodies the abstractly expressed structure, the computer system holds a machine learning model that receives the lower-level concept and the higher-level concept as input and outputs a score that represents a similarity between the input lower-level concept and the input higher-level concept; accepting a search query including data corresponding to the superordinate concept or the subordinate concept for search as a search key; For each document stored in the database, the search key and either the higher-level concept or the lower-level concept contained in the document are input to the machine learning model; A computer system characterized by identifying the documents that include the higher-level concept or the lower-level concept that is similar to the search key based on the score obtained from the machine learning model.

2. 2. The computer system of claim 1, The processor: If the search key included in the search query is the lower concept, the higher concept included in the document is input to the machine learning model; A computer system characterized in that, when the search key included in the search query is the higher-level concept, the lower-level concept included in the document is input to the machine learning model.

3. 3. The computer system according to claim 2, The processor: a first process for extracting the element from the document; a second process for analyzing relationships between the elements; a third process of generating a graph connecting the elements based on the results of the analysis and storing the generated graph in the database as the structure data.

4. 4. The computer system according to claim 3, The processor: extracting the superordinate concept and the subordinate concept from the structural data; generating pairs of the hypernym and the hypernym; determining whether the subordinate concept constituting the pair corresponds to the subordinate concept of the superordinate concept constituting the pair, and generating training data by assigning the result of the determination to the pair; A computer system that generates the machine learning model by performing contrastive learning using the learning data.

5. 4. The computer system according to claim 3, the document is a patent document, The generic concept is information described in the claims of the patent document, A computer system characterized in that the subordinate concept is information described in an example of the patent document.

6. 6. The computer system according to claim 5, The patent document contains information about chemical substances as the knowledge, The patent document includes a claim that describes a Markush structure representing a compound.

7. 7. The computer system of claim 6, In the first process, if the Markush structure is extracted as the element, the processor converts the Markush structure into text.

8. 4. The computer system according to claim 3, the processor provides an interface for setting a schema of the structure data; In the first process, the processor extracts the elements from the document based on the schema.

9. 2. The computer system of claim 1, The computer system, wherein the processor provides an interface for presenting the identified document.

10. An information retrieval method executed by a computer system, comprising: The computer system a processor, a storage device connected to the processor, and a network interface connected to the processor; Manages a database that stores structured data in a graph format, with the elements that make up a document as nodes, The document contains information about the structure as knowledge; The knowledge includes a superordinate concept, which is knowledge that abstractly expresses a part of the structure of the structure, and a subordinate concept, which is knowledge that embodies the abstractly expressed structure, the computer system holds a machine learning model that receives the lower-level concept and the higher-level concept as input and outputs a score that represents a similarity between the input lower-level concept and the input higher-level concept; The information retrieval method includes: a step of receiving a search query including, as a search key, data corresponding to the superordinate concept or the subordinate concept for search by the processor; a step in which the processor inputs the search key and either the higher-level concept or the lower-level concept contained in the document to the machine learning model for each document stored in the database; and identifying the document that includes the higher-level concept or the lower-level concept that is similar to the search key based on the score obtained from the machine learning model.

Images