基于可分解注意力机制的医疗问句语义匹配研究

Research on semantic matching of medical questions based on a decomposable attention model

问句语义匹配旨在判定给定的两个语句的语义信息是否匹配,在信息检索、自动问答、机器翻译等领域应用广泛,是自然语言处理研究的一个关键问题。现有基于机器学习或深度学习的问句语义匹配任务大多采用对整个句子构建语义信息表示,而忽视了语句各组成部分所蕴含的具体细节信息。提出一种基于可分解注意力机制的语义匹配模型(Decomposable Attention based Semantic Matching,DASM),该模型首先使用软注意力机制将整个序列问句分解为可以独立解决的子问句,使得子问句间权重计算可以并行;然后结合注意力机制充分捕获问句中潜在的语义信息,从而提高问句匹配任务的性能。实验结果表明,本文方法提高了问句语义匹配的准确性和模型性能。

The purpose of question semantic matching is to determine whether the semantic information of two given sentences matches,It is widely used in information retrieval,automatic question and answer,knowledge extraction and other fields.It is a key problem in natural language processing research.Most of the existing sentence semantic matching based on machine learning or deep learning adopts to construct semantic information representation of the whole sentence,but ignores the specific details of the sentence.Therefore,we propose a decomposable attention based semantic matching model.Firstly,this model uses the soft attention mechanism to decompose the whole sequence of questions into sub-questions that can be solved independently,so that the weight calculation between sub-questions can be parallel.Then the attention mechanism is combined to fully capture the potential semantic information in the question sentences,so as to improve the performance of the question matching task.Experimental results show that this method is superior to the existing modeling methods and enhances the accuracy and performance of semantic matching.