基于多粒度双向注意力机制的词义消歧深度学习方法

DEEP LEARNING METHOD FOR WORD SENSE DISAMBIGUATION BASED ON MULTI-GRANULARITY BI-DIRECTIONAL ATTENTION

词义消歧的目标是在特定的上下文中识别歧义词的正确词义。传统的监督方法主要是利用上下文的数据,而忽略了丰富的词义定义等词汇资源。最近的研究发现将词义定义整合到神经网络对于词义消歧具有显著的改进效果。提出引入词义定义的基于多粒度双向注意力机制的词义消歧模型,该模型采用字符级、词级和上下文嵌入的表示,使用双向注意力机制获取上下文和词义定义之间的交互关系,消融实验验证了模型中每个组成的重要性。实验结果表明,这种建模方式可以有效地识别歧义词的正确词义,在SemEval-13-task#12和SemEval-15-task#13公开数据集上进行了测试,F1值分别可达到68.9%和73.1%。

The goal of word sense disambiguation is to identify the correct meaning of an ambiguous word in particular context.Traditional supervised methods mainly rely on massive context data,ignoring lexical resources like rich sense definitions.Recent studies have shown that incorporating sense definitions into neural networks for word sense disambiguation has made significant improvement.This paper proposes a word sense disambiguation model that integrates senses definitions with multi-granularity bi-directional attention.It used character-level,word-level,and context embedding representations,and used bi-directional attention to obtain the interaction between the context and sense definitions.The ablation study verified the importance of each component in the model.The experiment results show that the modeling method can effectively distinguish the correct word meanings of the disambiguation words,and the accuracy can reach 68.7%and 73.1%respectively when tested on the public data sets of the semeval-13-task#12 and the semeval-15-task#13.