基于跨度和边界探测的实体关系联合抽取模型

Joint Extraction Model for Entity Relationships Based on Span and Boundary Detection

针对大多数跨度模型将文本分割成跨度序列时,产生大量非实体跨度,导致了数据不平衡和计算复杂度高等问题,提出了基于跨度和边界探测的实体关系联合抽取模型(joint extraction model for entity relationships based on span and boundary detection,SBDM)。SBDM首先使用训练Transformer的双向编码器表征量(bidirectional encoder representations from Transformer,BERT)模型将文本转化为词向量,并融合了通过图卷积获取的句法依赖信息以形成文本的特征表示;接着通过局部信息和句子上下文信息去探测实体边界并进行标记,以减少非实体跨度;然后将实体边界标记形成的跨度序列进行实体识别;最后将局部上下文信息融合到1个跨度实体对中并使用sigmoid函数进行关系分类。实验表明,SBDM在SciERC(multi-task identification of entities,relations,and coreference for scientific knowledge graph construction)数据集、CoNLL04(the 2004 conference on natural language learning)数据集上的关系分类指标S F1分别达到52.86%、74.47%,取得了较好效果。SBDM用于关系分类任务中,能促进跨度分类方法在关系抽取上的研究。

In view of the fact that when most span models divided text into span sequences,a large number of non-entity spans were generated,which led to problems such as data imbalance and high computational complexity.This paper proposed a joint extraction model of entity relationships based on span and boundary detection(SBDM).The model first used the bidirectional encoder representations from Transformer(BERT)model to convert text into word vectors,and integrated syntactic dependency information obtained through graph convolution to form a feature representation of the text.Next,it used local information and sentence context information to detect and mark entity boundaries,thereby reducing non-entity spans.Then,the span sequence formed by entity boundary markers was used for entity recognition.Finally,local context information was fused into a span pair and the sigmoid function was used for relationship classification.The experiment showed that SBDM achieved good results in relation classification S F1 values of 52.86%and 74.47%on the multi-task identification of entities,relations,and coreference for scientific knowledge graph construction(SciERC)dataset and the 2004 conference on natural language learning(CoNLL04)dataset,respectively.The use of SBDM in relation classification tasks promotes the research of spanning classification methods in relation.