KR-20260064044-A - Knowledge retrieval device and method based on temporal knowledge graph and method for generating temporal knowledge graph

Abstract

A knowledge retrieval device and method based on a knowledge graph with attached time information, and a method for generating a knowledge graph with attached time information are provided. A knowledge retrieval device according to one embodiment of the present invention comprises: a knowledge graph generation module that generates a knowledge graph with attached time information from the latest knowledge; a query analysis module that analyzes an input question to extract entities, relationships, and timestamp information; a knowledge graph search module that searches for content related to the question in the knowledge graph with attached time information using the entities, relationships, and timestamp information analyzed and extracted by the query analysis module; and a generative language model that generates a response to the question based on the search results in the knowledge graph search module.

Inventors

• 이기영
• 권오욱
• 류지희
• 서영애
• 성진
• 신종훈
• 임수종
• 허정

Assignees

• 한국전자통신연구원

Dates

Publication Date

20260507

Application Date

20241031

Claims (20)

1. Knowledge graph generation module that generates a knowledge graph with time information attached from the latest knowledge; A query analysis module that analyzes input questions to extract entity, relationship, and timestamp information; A knowledge graph search module that searches for content related to the question in the time information-attached knowledge graph using entity, relationship, and timestamp information analyzed and extracted by a query analysis module; and A generative language model that generates a response to the above question based on the search results from the above knowledge graph search module A knowledge retrieval device equipped with
2. In paragraph 1, the knowledge graph generation module is, Request the generative language model to generate time-related questions from resources after the deployment date of the above generative language model, and The generated time-related question is input directly into the generative language model to generate a first response using knowledge embedded in the generative language model in the form of parameters, and the generative language model generates a second response using the resources used to generate the time-related question. If the first response and the second response are determined to be different, a knowledge graph is created using the content of the resource used to generate the time-related question and stored in the time-information-attached knowledge graph. Knowledge search device.
3. In paragraph 2, determining that the first response and the second response are different is, Calculating the semantic similarity between a first response and a second response, and determining that the two responses are different if the calculated similarity value is smaller than a predetermined threshold, Knowledge search device.
4. In paragraph 2, the above-mentioned latest knowledge is knowledge that is not internalized in the form of parameters in the generative language model, Knowledge search device.
5. In any one of paragraphs 1 to 4, the knowledge graph is, A device in which timestamps representing time information and text content are stored together Knowledge search device.
6. In any one of paragraphs 1 to 4, the knowledge graph is, A knowledge retrieval device comprising a plurality of entities, relationships between the plurality of entities, and time information attached to an edge representing the relationships.
7. In any one of paragraphs 1 through 4, The device further comprises an external knowledge search module for searching for information sources located outside the device (hereinafter referred to as external knowledge), and The above generative language model generates a response to the above question based on the search results in the knowledge graph search module when matching content is found in the time-information-attached knowledge graph, and generates a response to the above question based on the search results in the external knowledge search module when matching content is not found in the time-information-attached knowledge graph. Knowledge search device.
8. In Paragraph 7, The above generative language model generates a response to the above question based on the search results in the knowledge graph search module when the search results in the knowledge graph search module match the entities, relationships, and timestamp conditions. Knowledge search device.
9. A knowledge graph generation step for generating a knowledge graph with time information attached from the latest knowledge, which is knowledge not internalized in the form of parameters in a generative language model; A step of extracting entity, relationship, and timestamp information from an input question; A step of searching for relevant content in the time information-attached knowledge graph using extracted entity, relationship, and timestamp information; When matching content is found in the above time-information-attached knowledge graph, the step of generating a final response to the above question based on the search results in the time-information-attached knowledge graph using the above generative language model A knowledge search method equipped with
10. In claim 9, the knowledge graph generation step is, A step of requesting a generative language model to generate time-related questions from resources after the deployment date of the generative language model, and A step of inputting the generated time-related question directly into the generative language model to generate a first response using knowledge embedded in the generative language model in the form of parameters, and causing the generative language model to generate a second response using the resources used to generate the time-related question; If the first response and the second response are determined to be different, a step of generating a knowledge graph using the content of the resource used to generate the time-related question and storing it in the time-information-attached knowledge graph. A knowledge retrieval method including
11. In Paragraph 10, determining that the first response and the second response are different is, Calculating the semantic similarity between a first response and a second response, and determining that the two responses are different if the calculated similarity value is smaller than a predetermined threshold, Knowledge search methods.
12. In paragraph 10, the above-mentioned latest knowledge is knowledge that is not internalized in the form of parameters in the above-mentioned generative language model, Knowledge search methods.
13. In any one of paragraphs 9 to 12, the knowledge graph is, A device in which timestamps representing time information and text content are stored together Knowledge search methods.
14. In any one of paragraphs 9 to 12, the knowledge graph is, A knowledge retrieval method comprising a plurality of entities, relationships between the plurality of entities, and time information attached to an edge representing the relationship.
15. In any one of paragraphs 9 through 12, A step of searching for external knowledge related to the user question when a search in a knowledge graph with time information attached fails, and generating a response to the above question using a generative language model based on the external knowledge search results. A knowledge search method that further includes.
16. In any one of paragraphs 9 through 12, A case where a search in a time-attached knowledge graph fails is when the search result in the time-attached knowledge graph does not match any of the above entities, relationships, and timestamp conditions, Knowledge search methods.
17. A step of requesting the generative language model to generate time-related questions from resources after the deployment date of the generative language model, and A step of inputting the generated time-related question directly into the generative language model to generate a first response using knowledge embedded in the generative language model in the form of parameters, and causing the generative language model to generate a second response using the resources used to generate the time-related question; If the first response and the second response are determined to be different, the step of creating and saving a knowledge graph using the content of the resources used to generate the time-related question. A method for generating a time-attached knowledge graph equipped with
18. In Paragraph 17, determining that the first response and the second response are different is, Calculating the semantic similarity between a first response and a second response, and determining that the two responses are different if the calculated similarity value is smaller than a predetermined threshold, Method for generating time-attached knowledge graphs.
19. In paragraph 17 or 18, the knowledge graph above is, A device in which timestamps representing time information and text content are stored together Method for generating time-attached knowledge graphs.
20. In paragraph 17 or 18, the knowledge graph above is, Composed of a plurality of entities, relationships between the plurality of entities, and time information attached to an edge representing the relationship, Method for generating time-attached knowledge graphs.

Description

Knowledge retrieval device and method based on temporal knowledge graph and method for generating temporal knowledge graph The present invention relates to a knowledge retrieval device and method based on a knowledge graph with attached time information, and a method for generating a knowledge graph with attached time information. With the advancement of artificial intelligence technology, various technologies utilizing large language models are delivering high performance across numerous domains and tasks. In fact, high-quality results employing language models are pouring out, ranging from language comprehension to language generation. However, since updating language models in real-time is costly, most new models are released at specific intervals. For this reason, language models have not been trained on the latest knowledge since their release, leading to issues such as hallucinations or the generation of outdated answers to questions requiring up-to-date knowledge. Several approaches have been proposed to address the above problem, the most representative of which is directly editing the parameters of the language model. However, this method has the disadvantage that modifying the parameters may also modify related content that should not be modified. Additionally, although continual learning techniques that continuously learn based on new data are used, this method also suffers from the disadvantage of catastrophic forgetting, where previously learned content is forgotten. FIG. 1 is a functional block diagram showing the configuration of a knowledge search device based on a knowledge graph with attached time information according to one embodiment of the present invention. FIG. 2 is a flowchart showing the process of generating a final response through a generative language model by referring to a knowledge graph with time information attached to a user question input according to one embodiment of the present invention. FIG. 3 is a flowchart showing the process of generating a knowledge graph with attached time information according to one embodiment of the present invention. Figure 4 shows one example of a knowledge graph with attached time information being generated from a specific document. The aforementioned objectives of the present invention, as well as other objectives, advantages, and features, and the methods for achieving them, will become clear from the embodiments described in detail below together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but can be implemented in various different forms, and the following embodiments are provided merely to easily inform those skilled in the art of the purpose, structure, and effects of the invention, and the scope of the rights of the present invention is defined by the description in the claims. Meanwhile, the terms used in this specification are for describing the embodiments and are not intended to limit the invention. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. As used in this specification, "comprises" and/or "comprising" do not exclude the presence or addition of one or more other components, steps, actions, and/or elements to the mentioned components, steps, actions, and/or elements. Generally, language models are not trained on new knowledge information generated after the time of their development and deployment. The time-information-attached knowledge graph proposed in this invention stores the latest knowledge available since the deployment of the language model, along with timestamp information. Subsequently, when a question required by the language model is input, the timestamp information attached to the question and the query content are analyzed, and a more accurate response is generated by first searching the time-information-attached knowledge graph. FIG. 1 is a functional block diagram showing the configuration of a knowledge search device based on a knowledge graph with attached time information according to one embodiment of the present invention. The knowledge search device (100) of FIG. 1 comprises a knowledge graph generation module (101) that generates a knowledge graph with time information attached from the latest knowledge, a query analysis module (102) that analyzes an input question and extracts entity, relationship, and timestamp information, a knowledge graph search module (103) that searches for content related to the question in a knowledge graph (104) with time information attached using the entity, relationship, and timestamp information analyzed and extracted by the query analysis module (102), an external knowledge search module (105) that searches for information sources (hereinafter referred to as external knowledge) (200) located outside the device, such as the internet, and a generative language model (106) that generates a response to the question based on the search r