Knowledge graph generation apparatus, knowledge graph generation method, and knowledge graph generation program

Abstract

To provide a knowledge graph generation apparatus for generating a knowledge graph in which reliability is given to edges.SOLUTION: A knowledge graph generation apparatus 10 comprises: a text data acquisition unit 121 that acquires text data from an information source; a cleansing processing unit 122 that cleanses the text data; a triple extraction unit 131 that applies an information extraction model to the cleansed text data to extract triples; and a weighting unit 132 that assigns weights to the triples based on the reliability of the triples.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an apparatus for automatically generating a knowledge graph from text, which constructs a knowledge graph to which weights based on the reliability of an information source and the reliability of an information extraction model are added. [Background technology]

[0002] Conventionally, image information and text information are extracted from videos, and weighted knowledge graphs are generated for each (Patent Document 1). Conventionally, a weighted knowledge graph has been generated, with weights being the relevance scores between multiple words extracted from multiple documents (Patent Document 2). Previously, a weighted knowledge graph embedding method has been proposed (Non-Patent Document 1). Previously, we proposed a framework that uses probabilistic classification methods to automatically infer accurate and ambiguous knowledge from web text containing uncertain information (Non-Patent Document 2). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Special Publication No. 2022-502376 [Patent Document 2] Special Publication No. 2023-513886 [Non-patent literature]

[0004] [Non-Patent Document 1] Kun, Kong Wei, et al. "WeExt: A Framework of Extending Deterministic Knowledge Graph Embedding Models for Embedding Weighted Knowledge Graphs." IEEE Access (2023). [Non-patent document 2] Wu, Wentao, et al. "Probase: A probabilistic taxonomy for text understanding." Proceedings of the 2012 ACM SIGMOD international conference on management of data. 2012. Summary of the Invention [Problem to be solved by the invention]

[0005] Various information is made public through web pages, social media posts, etc. Meanwhile, in the security domain, various threat information is published on platforms such as social media, and it is essential to collect and utilize this information early. However, since web pages and social media posts are written in natural language, it is difficult to apply them to mechanical processing.

[0006] To efficiently utilize unstructured data written in natural language, it is necessary to convert it into structured data. However, manually extracting the necessary data from the text and converting it into structured data is costly. Therefore, it is desirable to use natural language processing to automatically extract information.

[0007] One of the structured data formats that can be used for information extraction and transformation is a structured format called a knowledge graph. A knowledge graph has a flexible graph structure with entities as nodes and relationships as edges.

[0008] By using information extraction technology to appropriately extract subject / predicate / object pairs (a combination of these three elements is called a "triple") from a sentence, it becomes possible to automatically extract a knowledge graph. Efforts are being made to automatically generate knowledge graphs using this knowledge acquisition technology based on natural language processing.

[0009] In the automatic generation of knowledge graphs, false positives in natural language processing models or errors in the input information sources themselves can lead to the inclusion of incorrect or unreliable knowledge. When this happens, the nodes adjacent to the node containing the incorrect knowledge are also affected by the error, which can reduce the accuracy of inference tasks using the knowledge graph and lead to a deterioration in the visibility of the knowledge graph itself.

[0010] On the other hand, since there is a limit to the accuracy of model extraction, it is necessary to allow the inclusion of incorrect knowledge due to false positives in the automatically generated knowledge graph.Furthermore, if the input sources that form the basis of the knowledge graph are restricted to reliable sources in order to avoid the inclusion of incorrect knowledge, the diversity and immediacy of the knowledge contained in the knowledge graph may decrease.

[0011] Based on the above background, we propose a method for automatically generating knowledge graphs that allows incorrect knowledge to be mixed into the knowledge graph, and when adding knowledge to the knowledge graph, weights the edges of the knowledge graph based on how likely it is that the knowledge is correct. This makes it possible to reduce the impact of incorrect knowledge on the entire knowledge graph while maintaining the diversity and immediacy of the information within the knowledge graph.

[0012] Possible use cases include removing edges with reliability below a threshold, or reducing the impact of low-reliability nodes on the embedding by considering edge weights when deriving a graph embedding. Also, weighting a graph makes it possible to apply weighted graph embedding techniques.

[0013] In Patent Document 1, an image knowledge graph and a text knowledge graph are created from video data, and a weighted graph is generated from them. In contrast, in the present invention, a plurality of text data are input to the knowledge graph generation device. In Patent Document 2, the relevance between words is used as the weight of a node, whereas in the present invention, the reliability of the knowledge graph is used as the weight.

[0014] The present invention aims to assign the probability of correct knowledge to edges in an apparatus that automatically generates a knowledge graph from text. [Means for solving the problem]

[0015] When automatically creating a knowledge graph from natural language text, the present invention assigns the probability of correct knowledge to edges, thereby constructing a weighted knowledge graph in which less reliable knowledge has edges with small weights and more reliable knowledge has edges with large weights.

[0016] Here, the reliability that serves as the weight of the edge is calculated from the "reliability assigned in advance to the information source itself" and the "quantitative value of the reliability score in the information extraction task." The latter "quantitative value of reliability score in information extraction task" is as follows: (A) A reliability score is defined in advance for the model itself (for example, the reliability of the model itself, such as the F1 score of the information extraction model itself). (a) A case where reliability is assigned to each triple that is the output of the information extraction task (e.g., the output value of the softmax function in the final layer). We assume two patterns:

[0017] Then, for example, the knowledge graph is generated by performing the following steps. (Step S1) Obtain text data from a specified information source. This acquisition method can be broadly classified into two methods: acquisition by web crawling and acquisition by API.

[0018] (Step S2) Information that is not used for information extraction, such as unnecessary HTML tags, is deleted from the acquired text data. This process of deleting information that is not used for information extraction is called "cleansing."

[0019] (Step S3) Using the information extraction model corresponding to the information source, named entity information and relationships are extracted. Proper nouns such as place names, personal names, and organization names, plus quantities, dates, times, etc., are called "named entities."

[0020] (Step S4) The reliability of the information source i is calculated as W i The reliability of information extraction model j corresponding to information source i is denoted by V j Then, for the extracted triple (S, V, O), the weight W i *V j is added and stored in the graph database. Note that the symbol "*" represents multiplication. The above operations are performed for all target information sources.

[0021] The present inventors have discovered that in a knowledge graph generation device including a text data acquisition unit that acquires text data from an information source, a text data cleansing processing unit that cleanses the text data, and a triple extraction unit that applies an information extraction model to the cleansed text data to extract triples, weights based on the reliability of the triples can be assigned to the triples, and have completed the present invention.

[0022] (1) A knowledge graph generation device according to the present invention includes a text data acquisition unit that acquires text data from an information source, a text data cleansing processing unit that cleanses the text data, a triple extraction unit that applies an information extraction model to the cleansed text data to extract triples, and a weighting unit that assigns weights to the triples based on the reliability of the triples.

[0023] (2) In the knowledge graph generation device, the reliability of the triple may be the reliability of the information source.

[0024] (3) In the knowledge graph generation device, the reliability of the triple may be the reliability of the information extraction model.

[0025] (4) In the knowledge graph generation device, the reliability of the triple may be a product of the reliability of the information source and the reliability of the information extraction model.

[0026] (5) In the knowledge graph generation device, the reliability of the triple may be an output value of a softmax function in a final layer of a neural network of the information extraction model.

[0027] (6) In the knowledge graph generation device, the reliability of the triple may be a product of the reliability of the information source and an output value of a softmax function in a final layer of a neural network of the information extraction model.

[0028] (7) A knowledge graph generation method according to the present invention includes a text data acquisition step of acquiring text data from an information source, a text data cleansing processing step of cleansing the text data, a triple extraction step of applying an information extraction model to the cleansed text data to extract triples, and a weighting step of assigning weights to the triples based on the reliability of the triples.

[0029] (8) A knowledge graph generation program according to the present invention causes a computer to execute a text data acquisition step of acquiring text data from an information source, a text data cleansing processing step of cleansing the text data, a triple extraction step of applying an information extraction model to the cleansed text data to extract triples, and a weighting step of assigning weights to the triples based on the reliability of the triples. [Effects of the Invention]

[0030] According to the present invention, in a knowledge graph that is automatically generated using information extraction technology, the reliability of triples is used as a weight, thereby reducing the influence of low-accuracy knowledge that gets mixed into the knowledge graph. [Brief explanation of the drawings]

[0031] [Figure 1] FIG. 2 is a diagram illustrating functions of the knowledge graph generation device according to the first embodiment of the present invention. [Figure 2] FIG. 2 is a diagram illustrating a hardware configuration of a server according to the first embodiment of the present invention. [Figure 3] FIG. 2 is a diagram showing a processing flow of the first embodiment of the present invention. [Figure 4] FIG. 2 is a diagram illustrating an example of triple extraction according to the first embodiment of the present invention. [Figure 5] FIG. 4 is a diagram illustrating an example of weighting according to the first embodiment of the present invention. [Figure 6] FIG. 10 is a diagram illustrating an example of weighting according to the second embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0032] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. [First embodiment] The knowledge graph generation device according to the first embodiment of the present invention is characterized in that it uses the reliability of the information source and the reliability of the information extraction model. The knowledge graph generation device of the first embodiment includes: (1) Information collection server (2) Model management and inference server (3) Knowledge graph storage server It has three servers:

[0033] FIG. 1 is a diagram showing the functions of three servers that constitute a knowledge graph generation device 10 of the first embodiment. The information collection server 120 performs a process of acquiring text data from an information source, and also a process of deleting unnecessary information from the acquired text data. The information collection server 120 includes a text data acquisition unit 121, a cleansing processing unit 122, a collection target domain list 123, a collection target API list 124, an API key list 125, a collection database 126, a post-cleansing database 127, and a text data cleansing processing rule list 128.

[0034] The text data acquisition unit 121 acquires text data from an information source. The cleansing processing unit 122 deletes unnecessary information from the acquired text data. The collection target domain list 123 lists domains from which text data is to be acquired. The collection target API list 124 lists APIs from which text data is to be acquired. An API serves as an interface that enables functions and data provided by applications of other programs and services to be utilized according to certain rules. To use an API, a user first registers as a member with the service that publishes the API they want to use and acquires an API key, which is an authentication code for using the API. The API can be used by embedding the acquired API key in a program that uses the API. Other programs and services use the API key to identify users and verify requests. The API key list 125 is a list of API keys. The collection database 126 is a database of text collected by the text data acquisition unit 121. The post-cleansing database 127 is a database that stores text data after being cleansed by the cleansing processing unit 122. The text data cleansing processing rule list 128 is a list that defines rules for how to perform text cleansing for each information source.

[0035] The model management and inference server 130 uses an information extraction model corresponding to the information source to The model management and inference server 130 performs a process to extract triples and assigns weights to the triples based on their reliability. The model management and inference server 130 includes a triple extraction unit 131 , a weighting unit 132 , an information extraction model list 133 , an inference log 134 , an information source reliability list 135 , and an information extraction model reliability list 136 .

[0036] The triple extraction unit 131 extracts triples from the cleansed text data. The weighting unit 132 assigns weights to the triples based on their reliability. The information extraction model list 133 lists information extraction models for extracting triples in association with information sources. The inference log 134 is a log of inferences performed by the model management and inference server 130. The information source reliability list 135 lists the reliability of information sources in association with the information sources. The information extraction model reliability list 136 lists the reliability of information extraction models in association with the information extraction models.

[0037] The knowledge graph storage server 140 includes a knowledge graph database 141. The knowledge graph storage server 140 stores the weighted triples in the knowledge graph database 141.

[0038] All three of the above servers have the hardware configuration shown in Figure 2. The server 100 is composed of a CPU (Central Processing Unit) 101, a ROM (Read Only Memory) 102, a RAM (Random access memory) 103, a storage unit 104, an input unit 105, an output unit 106, a communication unit 107, and a bus 108. The CPU 101, the ROM 102, the RAM 103, the storage unit 104, the input unit 105, the output unit 106, and the communication unit 107 are interconnected via the bus 108.

[0039] The CPU 101 operates according to a control program stored in the ROM 102 . The ROM 102 is made up of a read-only non-volatile semiconductor memory, and stores a control program, which is a computer program for causing the server 100 to execute processes, and the like. The RAM 103 is made up of a volatile semiconductor memory that can be read and written, and provides a work area when the CPU 101 executes a program.

[0040] The CPU 101 uses the RAM 103 as a work area and operates according to the control program stored in the ROM 102, whereby the CPU 101, ROM 102, and RAM 103 functionally constitute various processing units. The CPU 101, ROM 102, and RAM 103 of the information collection server 120 functionally constitute a text data acquisition unit 121 and a cleansing processing unit 122. The CPU 101, ROM 102, and RAM 103 of the model management and inference server 130 functionally constitute a triple extraction unit 131 and a weighting unit 132.

[0041] The storage unit 104 is configured, for example, from a hard disk unit. The input unit 105 writes data into the storage unit 104 . The output unit 106 reads out the data stored in the storage unit 104 . The communication unit 107 exchanges data with other servers.

[0042] FIG. 3 is a process flow diagram of the first embodiment of the present invention. The information source D1 is a website or a post by a user on a social networking site. In step S102, the text data acquisition unit 121 acquires text data from the information source D1. More specifically, the text data acquisition unit 121 crawls the sites in the collection target domain list 123 and stores the results in the collection database 126. The text data acquisition unit 121 also calls an API in the collection target API list 124 and stores the response in the collection database 126. The processing result of step S102 is the text data D2 stored in the collection database 126.

[0043] In step S104, the cleansing processing unit 122 performs cleansing processing on the text data D2. Since the format of the text data D2 differs for each target site and API, the cleansing processing on the text data is performed by referring to a rule list (stored in the text data cleansing processing rule list 128) on how to perform cleansing processing for each information source D1. The results are stored in the post-cleansing database 127. As a concrete example of cleansing processing, for example, in the case of a website, the necessary information is <h2>< / h2> Since only the text of the tag part is present, the tag part is extracted. The processing result of step S104 is the cleansed text data D3 stored in the post-cleansing database 127.

[0044] In step S106, the triple extraction unit 131 extracts triples from the cleansed text data D3. The triple extraction unit 131 extracts triples from the cleansed text data D3 using an information extraction model that is stored in the information extraction model list 133 and that corresponds to the information source D1. As is well known, triple extraction is performed by performing two steps: named entity extraction and relation extraction.

[0045] IREX (Information Retrieval and Extraction Exercise), a workshop on information extraction and information retrieval, defines the following eight types of named entities: (1)Organization name (ORGANIZATION:ORG) (2) Person name (PERSON:PER) (3) Place name (LOCATION: LOC) (4) Date Expression (DATE:DAT) (5) Time expression (TIME:TIM) (6) Amount expression (MONEY:MNY) (7) Percentage Expression (PERCENT:PNT) (8) Unique object name (ARTIFACT:ART) These named entities become the subjects or objects of triples and become nodes in the knowledge graph. Named entities that become nodes in the knowledge graph are sometimes called entities.

[0046] The relation extraction process is a process for estimating the relation between the extracted named entities. It is important to note that the meanings of the subject, predicate, and object of triples in a knowledge graph may differ from the meanings of subjects, predicates, and objects in ordinary grammar. FIG. 4 shows an example of triple extraction in the first embodiment of the present invention. For example, suppose that the cleansed text data D3 obtained from information source A is the sentence "The author of Sherlock Holmes is Conan Doyle." Here, suppose triple extraction is performed using information extraction model a. Using a general information extraction model, after organizing the information into "Sherlock Holmes has an author named Conan Doyle," "Sherlock Holmes" is treated as the subject, "author" as the predicate, and "Conan Doyle" as the object. As a result, the triple t1 = (Sherlock Holmes, author, Conan Doyle) is obtained.

[0047] Also, suppose that the cleansed text data D3 obtained from information source B is the sentence "When an aqueous solution of silver nitrate is added to hydrochloric acid, a white precipitate of silver chloride is produced." Here, suppose that triple extraction is performed using information extraction model b. Details of the extraction process are omitted, but when a general information extraction model is used, t2 = (hydrochloric acid, react with silver nitrate solution) t3 = (silver chloride, made from hydrochloric acid) t4 = (silver chloride, made from an aqueous solution of silver nitrate) t5 = (silver chloride, is, white precipitate) This gives us four triples.

[0048] In step S108, the weighting unit 132 assigns a weight to the triple D4. Since the information source reliability list 135 stores reliability for each information source, the reliability corresponding to the information source is acquired. FIG. 5 shows an example of weighting in the first embodiment. A In addition, the reliability of information extraction model a is V a The weighting unit 132 calculates W A and V a The weighted triple s1 that results from assigning a weight to triple t1 is s1 = (Sherlock Holmes, author, Conan Doyle, W A V a )

[0049] Similarly, the reliability of information source B is calculated as W B In addition, the reliability of information extraction model b is V b The weighting unit 132 calculates W B and V b The weights assigned to the triples t2 to t5 are the products of W B V b As an example, the weighted triple s2 resulting from weighting the triple t2 is s2 = (hydrochloric acid, reacting, aqueous solution of silver nitrate, W B V b )

[0050] The result of the processing in step S108 is the weighted triple D5. In step S110, the knowledge graph storage server 140 stores the weighted triple D5 in the knowledge graph database 141.

[0051] (Variation 1) In the first embodiment, the reliability W corresponding to the information source is i and the reliability of the information extraction model Vj The product of these is used as the weight, but the reliability W corresponding to the information source i may be the weight given to the triple. (Variation 2) In the first embodiment, the reliability W corresponding to the information source is i and the reliability of the information extraction model V j The product of these is used as the weight, but the reliability of the information extraction model V j may be the weight given to the triple.

[0052] [Second embodiment] In the first embodiment, when the information extraction model b is applied to the text data obtained from the information source B, the weights of the obtained triples t2 to t5 are all W B V b It was. In the second embodiment, a knowledge graph creation device that assigns different weights to the triples t2 to t5 will be described.

[0053] The second embodiment is the same as the first embodiment up to the point where triple D4 in Fig. 3 is obtained. In the second embodiment, when weighting in step S108, the reliability W corresponding to information source A and the output value V of the softmax function of the final layer of the neural network when a neural network is used as the information extraction model are used. Since the output value V of the softmax function of the final layer differs for each triple, the output values ​​V of the softmax function of the final layer corresponding to triples t1 to t5 will be represented as V1 to V5.

[0054] An example of weighting in the second embodiment is shown in FIG. 6. The weight assigned to triple t1 is W A The weight assigned to triple t2 is W B The weight assigned to triple t3 is W B The weight assigned to triple t4 is W B The weight assigned to triple t5 is W B It becomes V5. In step S110, the knowledge graph storage server 140 stores the weighted triples s1 to s5 in the knowledge graph database 141.

[0055] (Variation 3) In the second embodiment, the weight is the product of the reliability W corresponding to the information source and the output value V of the softmax function in the final layer of the neural network of the information extraction model, but the output value V of the softmax function in the final layer of the neural network of the information extraction model may also be the weight assigned to the triple.

[0056] As described above, according to the first and second embodiments, weights based on reliability are assigned to triples extracted from text data. Therefore, by removing triples with weights below a threshold or by considering the weight of triples when obtaining an embedding of a graph, it is possible to reduce the impact of low-reliability triples on the embedding, thereby reducing the impact of incorrect knowledge on the entire knowledge graph.

[0057] Furthermore, the first and second embodiments make it possible to build, for example, a knowledge graph with reliable edges, which will contribute to achieving Goal 9 of the Sustainable Development Goals (SDGs) led by the United Nations, which is to "Develop resilient infrastructure, promote sustainable industrialization and foster innovation."

[0058] Although the first and second embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments. Furthermore, the effects described in the above-described embodiments are merely a list of the most preferable effects resulting from the present invention, and the effects of the present invention are not limited to those described in the embodiments.

[0059] The knowledge graph generation method by the knowledge graph generation device 10 is realized by software or hardware circuits. When realized by software, the programs constituting this software are installed in the information processing devices (computers) constituting each server. These programs may be recorded on removable media such as CD-ROMs and distributed, or may be distributed by being downloaded to computers via a network. [Explanation of symbols]

[0060] 10 Knowledge graph generator 100 Hardware configuration of knowledge graph generator 101 CPU 102 ROM 103 RAM 104 Storage section 105 Input section 106 Output section 107 Communications Department 120 Information Collection Server 121 Text data acquisition unit 122 Cleansing Processing Unit 123 Collection target domain list 124 Collection target API list 125 API Key List 126 Collection Database 127 Cleansed Database 128 Text Data Cleansing Processing Rule List 130 Model Management and Inference Server 131 Triple Extraction Section 132 Weighting section 133 Information Extraction Model List 134 Inference Log 135 Information Source Reliability List 136 Information Extraction Model Reliability List 140 Knowledge Graph Storage Server 141 Knowledge Graph Database

Claims

1. a text data acquisition unit that acquires text data from an information source; a text data cleansing processing unit that cleanses the text data; a triple extraction unit that applies an information extraction model to the cleansed text data to extract triples; a weighting unit that assigns a weight to the triple based on the reliability of the triple; A knowledge graph generation device including:

2. The knowledge graph generation device according to claim 1 , wherein the reliability of the triple is the reliability of the information source.

3. The knowledge graph generation device according to claim 1 , wherein the reliability of the triple is the reliability of the information extraction model.

4. The knowledge graph generation device according to claim 1 , wherein the reliability of the triple is a product of the reliability of the information source and the reliability of the information extraction model.

5. The knowledge graph generation device according to claim 1 , wherein the reliability of the triple is an output value of a softmax function in a final layer of the neural network of the information extraction model.

6. The knowledge graph generation device according to claim 1 , wherein the reliability of the triple is a product of the reliability of the information source and an output value of a softmax function in a final layer of the neural network of the information extraction model.

7. a text data acquisition step of acquiring text data from an information source; a text data cleansing processing step of cleansing the text data; a triple extraction step of applying an information extraction model to the cleansed text data to extract triples; a weighting step of assigning weights to the triples based on the reliability of the triples; A knowledge graph generation method including:

8. a text data acquisition step of acquiring text data from an information source; a text data cleansing processing step of cleansing the text data; a triple extraction step of applying an information extraction model to the cleansed text data to extract triples; a weighting step of assigning weights to the triples based on the reliability of the triples; A knowledge graph generation program that runs on a computer.

Images