基于协同注意力和自适应调整的阅读理解模型

Reading comprehension model based on coattention and adaptive adjustment

为解决机器阅读理解模型中存在语义向量表示不准确、信息冗余、长距离依赖等问题,提出一种基于协同注意力和自适应调整的阅读理解模型。基于余弦相似度计算问题与文档的相似度权重,根据相似度权重自适应调整文档的词嵌入,解决信息冗余问题;引入协同注意力机制,捕获文档和问题的交互信息,生成感知向量;利用自注意力机制学习文本内部的依赖关系,增强问题和文档的语义向量表示,解决长距离依赖问题,提升模型性能。实验结果表明,该模型在精确匹配和模糊匹配指标上均得到提升。

A reading comprehension model based on coattention and adaptive adjustment was proposed to solve the problems of inaccurate semantic vector representation,information redundancy and long-distance dependence in the machine reading comprehension model.The similarity weight between the problem and the document was calculated based on the cosine similarity degree,and the word embedding of the document was adaptively adjusted according to the similarity weight to solve the information redundancy problem.The coattention mechanism was introduced to capture the interaction information of the document and the problem and a perceptual vector was generated.The self-attention mechanism was used to learn the internal dependence of the text to enhance the semantic vector representation of the problem and the document,solve the long-distance dependence problem,and improve the performance of the model.Experimental results show that the proposed model is significantly improved on both accurate matching and fuzzy matching indicators.