Effective sentence-level relation extraction model using entity-centric dependency tree.

Area of Science:

• Natural Language Processing
• Information Extraction
• Computational Linguistics

Background:

• Traditional dependency-based relation extraction methods face limitations with hard pruning (losing information) and soft pruning (weakening syntax).
• Shift towards language model (LM) based methods increases resource consumption and computational costs with marginal gains.
• Existing methods struggle to balance syntactic information with lexical details effectively.

Purpose of the Study:

• To propose an efficient relation extraction model that overcomes the limitations of existing methods.
• To enhance the capture of syntactic and lexical information simultaneously.
• To reduce the reliance on large language models and additional data.

Main Methods:

• Developed an entity-centric dependency tree, reconstructing dependency trees with entities as root nodes.
• Proposed a novel model integrating entity-centric dependency trees with language models.
• Evaluated the model on standard datasets: TACRED, Re-TACRED, and SemEval 2010 Task 8.

Main Results:

• Achieved state-of-the-art F1-scores: 74.9% on TACRED, 91.2% on Re-TACRED, and 90.5% on SemEval 2010 Task 8.
• Demonstrated effective capture of syntactic information without losing lexical details.
• Showcased efficient relation extraction without requiring larger models or additional data.

Conclusions:

• The entity-centric dependency tree approach offers a robust solution for relation extraction.
• The proposed model provides a computationally efficient and high-performing alternative to current LM-based methods.
• This method advances relation extraction by optimizing the use of syntactic and lexical features.