关键词: 信息提取, 重叠关系, 三元组提取, BERT, 注意力机制, 深度学习

Abstract:

Aiming at the problems of complex entity overlap situations and difficulties in extracting multiple relational triples in natural language texts， a joint triple extraction model combining pointer network and relational embedding was proposed. Firstly， the BERT （Bidirectional Encoder Representations from Transformers） pre-training model was used to encode and represent the input sentence. Secondly， the head and tail pointer labeling was used to extract all subjects in the sentence， and the attention mechanism guided by subjects and relations was used to distinguish the importance of different relation labels to each word， so that the relation label information was added to the sentence embedding. Finally， for the subjects and each relation， the corresponding object was extracted by using the pointer labeling and cascade structure， and the relational triples were generated. Extensive experiments were conducted on two datasets， New York Times （NYT） and Web Natural Language Generation （WebNLG）， and the results show that the proposed model has better overall performance than the current best Novel Cascade Binary Tagging Framework （CasRel） model by 1.9 and 0.7 percentage points respectively； compared with the Extract-Then-Label method with Span-based scheme （ETL-Span） model， the performance improvements of the proposed model are more than 6.0% and more than 3.7% in the comparison experiments with 1 to 5 triples， respectively. Especially in complex sentences with more than 5 triples， the proposed model has the F1 score improved by 8.5 and 1.3 percentage points respectively. And stable extraction ability of this model is maintained while capturing more entity pairs， which further verifies the effectiveness of this model in triple overlap problem.

Key words: information extraction, overlapping relationship, triple extraction, BERT (Bidirectional Encoder Representations from Transformers), attention mechanism, deep learning