基于双重匹配注意力网络的观点型阅读理解

提出一种基于双重匹配注意力网络的方法.先用动态匹配机制迭代综合获取全局观点信息，同时利用多维度匹配机制在不同特征空间上计算全局语义信息，然后交互式多路注意力机制通过两路注意力之间的交互计算对上述全局的观点与语义信息进行融合，最后与选项表示结合预测答案的观点倾向.在观点型阅读理解数据集ReCO和Dureader上面的实验表明，该方法相对于基准模型在准确率上提升了1.18%和0.84%,在加权宏F1上提升了1.16%和0.75%.

Opinion reading comprehension based on double matching attention network

Opinion reading comprehension aims to judge the opinion of the article to the question according to the given question and article. The existing works mainly focus on the bidirectional matching of questions and articles using attention mechanism, which can only capture the matching relationship of local text fragments between questions and articles. It is difficult to pay attention to the opinion polarity in global semantic information, which limits the model performance. Therefore, a method based on double matching attention network is proposed. First, the dynamic matching mechanism fuses information iteratively to obtain the global opinion information, and the multidimensional matching mechanism calculates the global semantic information in different feature spaces. Then, Interactive Multiple Attention (IMA) mechanism fuses the above global opinion and semantic information through two ways of interactive attention. Finally, the result of IMA is combined with embedding of options to predict the opinion tendency. Experimental result on the opinion reading comprehension dataset ReCO and Dureader demonstrates that, this method improves the accuracy by 1.18% and 0.84%, and the weighted macro F1 by 1.13% and 0.75% compared with the benchmark model.