Long-Text Relation Extraction Model Based on Evidence Sentence Selection and Multi-attention Networks

Long-text relation extraction constitutes a critical natural language processing task with the objective of extracting structured semantic relations between entities, from lengthy textual documents. Since lengthy texts encompass semantic units of varying granularity, including mentions, entities, and sentences, their contributions to relation inference outcomes exhibit variability. To leverage and integrate this multi-granularity semantic information, this paper proposes a novel evidence-enhanced multiple attention network to explicitly indicate the differential impacts of various levels of semantic units on relation inference. An evidence judgment task is also introduced to guide the model's focus toward sentences containing more pertinent evidence. The proposed method significantly outperforms common baseline models in terms of evaluation metrics based on our experiments on three publicly available long-text relation extraction datasets.