Enhanced Relation Extraction Based on Pre-Trained Text Generation Model

• Published in: 2025 7th International Conference on Natural Language Processing (ICNLP) pp. 631 - 639
• Main Authors: Li, Dun, Ye, Yawen, Lu, Yuping, Yang, Guowei, Zhao, Hanjie, Wang, Junfeng
• Format: Conference Proceeding
• Language: English
• Published: IEEE 21.03.2025
• Subjects: Accuracy; Adaptation models; Context modeling; Data mining; Fault tolerance; Fault tolerant systems; Labeling; Natural language processing; rare entity; relation extraction; semantic inclusion; Semantics; text generation; Training
• DOI: 10.1109/ICNLP65360.2025.11108673

In the field of information extraction, relation extraction is a critical task. Traditional relation extraction models exhibit collective ignorance when faced with entity combinations not present in the training set, which affects the accuracy and coverage of relation extraction. To address this issue, this paper proposes a text generation-enhanced relation extraction model that employs text generation techniques and integrates text inclusion criteria to comprehensively examine entity combinations, thereby enhancing the extraction of relation triplets. The model utilizes an independent labeling framework for subjects and objects, coupled with a text generation module based on the pretrained T5 model, and improves the text similarity calculation method to accommodate information inclusion criteria. By adjusting weights, the model enhances fault tolerance for unseen entities, while the processes of text generation and information inclusion criteria further ensure extraction accuracy, thereby mitigating the phenomenon of collective ignorance. We evaluate the proposed model against several state-of-the-art models using the NYT and WebNLG datasets. The results demonstrate the superior performance of the proposed model, particularly in handling complex sentence structures and texts with multiple relations.