CN-121998057-A - Space-time knowledge graph construction method based on large language model and Beidou grid coding

CN121998057ACN 121998057 ACN121998057 ACN 121998057ACN-121998057-A

Abstract

The invention discloses a space-time knowledge graph construction method based on a large language model and Beidou grid coding, which comprises the steps of carrying out semantic segmentation on an input text by utilizing a large language model LLM, extracting a fact statement sentence from each segmentation, and respectively extracting a triplet, a time entity and a space geographic entity from the fact statement sentence; the method comprises the steps of carrying out time attribute coding on a time entity, constructing a meshed Beidou satellite geographic map, obtaining a Beidou grid coding data set, carrying out associated grid matching coding on the time entity to form a corresponding space attribute, and gathering the time entity containing the time attribute, the space geographic entity containing the space attribute and a triplet to construct a space-time knowledge map. According to the method, attribute coding is carried out on the time entity and the space geographic entity, and the coded time entity, the space geographic entity and the triples are assembled and constructed to obtain the space-time knowledge graph, so that space-time information in the entity can be conveniently and deeply mined, and the space-time reasoning capability is improved.

Inventors

ZHANG BO
PENG XIAOBO
MA XIAOFEI
DAI XUQIANG
Qiao Hanyang

Assignees

浙江时空智子大数据有限公司

Dates

Publication Date: 20260508
Application Date: 20260126

Claims (9)

1. A space-time knowledge graph construction method based on a large language model and Beidou grid coding is characterized by comprising the following steps: S1, carrying out semantic segmentation on an input text by utilizing a large language model LLM, extracting a fact statement sentence from each segmentation, and respectively extracting a triplet, a time entity and a space geographic entity from the fact statement sentence, wherein the triplet comprises an entity and an entity relation; S2, performing time attribute coding on the time entity, constructing a gridded Beidou satellite geographic map, wherein all grids in the Beidou satellite geographic map correspond to geographic codes to form a Beidou grid coding dataset, and performing associated grid matching coding on the space geographic entity by using the Beidou grid coding dataset to form corresponding space attributes; S3, assembling the time entity containing the time attribute, the space geographic entity containing the space attribute and the triples to construct a space-time knowledge graph.
2. The space-time knowledge graph construction method based on the big language model and the Beidou grid coding of claim 1 is characterized by further comprising the following steps: And S4, performing time sequence association on the time entities in all the blocks according to the time attribute codes, and performing space association on the space attribute codes of the space geographic entities in all the blocks based on the Beidou satellite geographic map.
3. The method for constructing a space-time knowledge graph based on big language model and Beidou grid coding according to claim 1, wherein in the method S1, statement extraction is carried out on the blocks, and the method adopts the following steps of Function extraction and atomization fact statement sentence set The expression is as follows: Wherein In the case of a large language model, For the purpose of semantic blocking, To extract the hint words.
4. The method for constructing space-time knowledge graph based on big language model and Beidou grid coding according to claim 1, wherein the method is characterized by comprising the steps of The encoded temporal attribute expression is as follows: , Is an entity The unique identification is provided with a unique identification, Representing the description information of the entity, As a matter of time-type, 、 Respectively representing the time start boundaries.
5. The space-time knowledge graph construction method based on the big language model and the Beidou grid coding of claim 1 is characterized in that in the method S2, the spatial attribute expression of the spatial geographic entity is as follows: Wherein Is an entity The unique identification is provided with a unique identification, Representing the description information of the entity, In the form of a spatial geographic type, Corresponding the longitude and latitude and the altitude of the geographic coordinates for the entity; And And respectively representing the Beidou grid code set and the Beidou grid level corresponding to the space geographic entity.
6. The space-time knowledge graph construction method based on the big language model and the Beidou grid coding according to claim 1 or 5, wherein the method S2 is further characterized by comprising the following steps: s21, through Function acquisition of spatially or geographically correlated sets of spatial geographic locations By means of Function judgment and acquisition of space geographic position The spatial geographic type and Beidou grid level of (C) are obtained Corresponding longitude and latitude and height data are combined with the Beidou grid level to carry out corresponding coding of the Beidou grid coding set; s22, constructing a space geographic position As a parent node, a spatial geographic entity And performing space geographic association as child nodes.
7. The space-time knowledge graph construction method based on the big language model and the Beidou grid coding is characterized in that grids in the Beidou satellite geographic map are subjected to space position resolution and hierarchical division, and a Beidou grid coding data set is correspondingly divided into Beidou grid levels belonging to different scales up and down.
8. The method for constructing the space-time knowledge graph based on the big language model and the Beidou grid coding of claim 6 is characterized in that the big language model LLM is utilized for carrying out space geographic position or/and space geographic entity for carrying out space geographic association, and the space geographic association comprises space attribute association and semantic association.
9. The method for constructing a space-time knowledge graph based on big language model and Beidou grid coding of claim 6, wherein the space geographic position and space geographic entity comprise space geographic information with space geographic types of points, lines and planes, the points represent space geographic position points, the lines represent linear sets of the space geographic position points, and the planes represent planar sets of the space geographic position points.

Description

Space-time knowledge graph construction method based on large language model and Beidou grid coding Technical Field The invention relates to the technical field of space-time knowledge graph in artificial intelligence and big data processing, in particular to a space-time knowledge graph construction method based on big language model and big dipper grid coding. Background Along with the rapid development of artificial intelligence and big data technologies, knowledge maps are used as core technologies for integrating data and semantic association, and are widely applied to the fields of intelligent transportation, urban management, environmental monitoring and the like. Knowledge graph is an information extraction technology for extracting important entities and relations from mass information and constructing a graph net structure. Due to the fact that deep associated information is good at being mined and global semantics are understood, the output accuracy and reliability of retrieval enhancement (RETRIEVAL-Augmented Generation, RAG) are remarkably improved through the application of the knowledge graph, and the problems that knowledge of a large language model (Large Language Model, LLM) is outdated, illusion is generated and the like are solved. The traditional knowledge graph technology has two defects that firstly, the construction process (from unstructured text) depends on the traditional natural language processing (Natural Language Processing, NLP) method, fuzzy descriptions such as 'upstream in the Yangtze river', 'emergency site close to a science and technology garden', 'post-flood river inspection' and the like cannot be well resolved, and 'upstream in the Yangtze river' comprises an upstream section of the Yangtze river (particularly a Yangtze river section from southwest side of each radan winter peak of Qinghai-Tibet plateau to Hubei Yichang river section of the Yangtze river) and a middle section of the Yangtze river (particularly a Yangtze river section from Hubei Yichang to Hubei river outlet of the river) by taking the fuzzy description of 'upstream in the Yangtze river' as an example. Secondly, static knowledge is expressed, dynamic change information cannot be well perceived in time and space dimensions, space-time semantic understanding capability under an actual service scene is lacking, for example, fuzzy description of the upper part of the Yangtze river is lacked, and natural language processing can only be recognized from semantics and cannot be understood and recognized in the space dimensions. In recent years, a space-time knowledge graph (Spatio-Temporal Knowledge Graphs) technique has been developed. The research shows that constructing the space-time knowledge graph with space-time characteristic information can enhance the space-time semantic understanding capability and realize complex space-time reasoning (such as predicting the future spatial distribution of the entity and tracing time sequence association), thereby providing a direct decision basis for the scene service. However, the traditional space-time knowledge graph has two short plates in technical floor and business adaptation, namely, the first short plate has weak space-time reasoning capability and poor generalization. Traditional space-time reasoning is limited by simple time sequence association (for example, an event A occurs after an event B) or space proximity judgment (for example, a place C approaches to a place D), depth mining of complex space-time dependency relations of cross-scale (for example, urban planning and neighborhood level management and control), cross-scene (for example, daily monitoring and emergency scheduling) is lacked, and an existing space-time reasoning model is mainly based on fixed rule design or specific data set training, and is remarkably insufficient in model suitability and generalization capability when facing dynamically-changing actual business requirements (for example, resource scheduling scenes after sudden disasters). Secondly, the space identifiers are not uniform and the entity suitability is poor. On one hand, the traditional space-time knowledge graph does not introduce a unified space identification standard, lacks a unified space specification, has different space identification formats of different data sources, and causes the problems of easy coordinate conflict, repeated labeling, large data fusion difficulty and the like when the multi-source data are in butt joint, on the other hand, the traditional space-time knowledge graph is difficult to realize multi-scale adaptation for different types of geographic entities such as points (such as single monitoring equipment), lines (such as roads and rivers), faces (such as administrative areas and lakes) and the like, and particularly cannot accurately describe the space range and association relation of the geographic entities in irregular forms such as rivers and complex administrative areas and the like. In