Refining Entity Descriptions with Relation Embeddings for Scientific Relation Classification

In recent years, the task of Relation Classification (RC) in scientific domains has received widespread attention. During the fine-tuning phase of Pre-trained Language Models (PLMs), the method of knowledge enhancement through the injection of entity description into PLMs can significantly improve their performance in Relation Classification(RC) task, particularly in scientific domains. However, current knowledge enhancement methods fail to effectively emphasize crucial information in external entity description knowledge that is most relevant to the target relations. A substantial amount of extraneous information may serve as noise, impairing the PLMs’ performance. This paper proposes a new knowledge enhancement method based on PLMs, which uses target Relation embeddings to Refine external Entity Description knowledge (RRED) for targeted knowledge enhancement. We initially introduce a relation embedding generation method based on masked language modeling objectives, yielding more precise relation embeddings. Subsequently, a gated collaborative attention module is meticulously designed. This innovative module integrates gating mechanisms with attention strategies, leveraging target relation embeddings to accentuate relevant information within entity descriptions, while concurrently mitigating the impact of irrelevant information. The experimental results demonstrate that the RRED method significantly improves the performance of PLMs across seven RC datasets in scientific domains.