A Temporal Knowledge Graph Generation Dataset Supervised Distantly by Large Language Models.

Area of Science:

• Natural Language Processing
• Information Extraction
• Knowledge Representation

Background:

• Knowledge graphs represent facts as triples (entity, relation, entity).
• Existing methods often overlook temporal information, limiting the understanding of dynamic relationships.
• Document-level relation extraction is crucial for building comprehensive knowledge graphs.

Purpose of the Study:

• To address the gap in temporal knowledge graph (TKG) construction from unstructured text.
• To develop a novel dataset and methodology for extracting temporal information from documents.
• To evaluate the effectiveness of large language models (LLMs) in temporal relation extraction.

Main Methods:

• Constructed a new dataset by mining relation patterns and combining facts with timestamps to form temporal quadruples.
• Utilized LLMs to generate temporal quadruples for triples lacking explicit timestamps.
• Implemented multiple filters and manual annotation for data quality assurance.
• Proposed an LLM-based framework transforming relation extraction into a sequence-to-sequence task for predicting relations with timestamps.

Main Results:

• Demonstrated the feasibility of constructing temporal knowledge graphs from documents.
• Showcased the performance of LLMs in extracting relations with temporal information using the proposed dataset and framework.
• Validated the quality and utility of the newly created temporal dataset.

Conclusions:

• The developed dataset and LLM-based framework significantly advance the field of temporal knowledge graph construction.
• Incorporating temporal information provides a more comprehensive understanding of factual connections within documents.
• LLMs show strong potential for extracting complex temporal relationships from text.