关键词: 远程监督, 关系抽取, 语义增强, 感知注意力, 降噪

Abstract: Focused on the issues that text feature extraction lacks consideration of the contextual discriminative features of sentences and fails to fully utilize the associative information between instances and relation labels, a relation extraction method combining semantic enhancement and perception attention was proposed. Firstly, during the sentence feature encoding phase, a semantic enhancement mechanism was constructed to extract salient semantic features of sentences, and an enhanced sentence representation was obtained through entity-aware word embeddings and salient feature perception. Then, a perception attention mechanism was designed to integrate sentence features. It evaluated the matching degree between sentences and relation labels by perceiving the semantic information between sentences and relation labels, the consistency information between the entity types of sentences and the corresponding relations, and the similarity information between sentences, so as to fully utilize the dependencies between instances and relation labels in a bag, further enhancing the noise reduction capability of the method. Finally, a classifier was used for relation prediction, and the network parameters were adjusted according to the cross-entropy between the predicted results and the actual results. Extensive experiments were conducted on the NYT-10 and GDS datasets. The experimental results show that on the NYT-10 dataset, compared with the BERT-based(Bidirectional Encoder Representations from Transformers) relation extraction method PARE (Passage-Attended Relation Extraction), the proposed method achieves an AUC value increase of 2.1 percentage points and an average Precision@N (P@M) increase of 2.4 percentage points for the top 100, 200, and 300 data entries ranked in descending order of confidence. On the GDS dataset, the AUC value and P@M also reach the highest values of 90.5% and 97.8%, respectively. The proposed method significantly outperforms other mainstream distant supervision relation extraction methods on both datasets, verifying its effectiveness. In mainstream distant supervision relation extraction tasks, the proposed method can effectively enhance the model's ability to learn data features.

Key words: distant supervision, relation extraction, semantic enhancement, perception attention, noise reduction