一种基于关键词扩展的答案块提取模型

Answer Chunk Extraction Model Based on Key Words’ Extension

针对问答式机器阅读理解中非定长答案的提取问题,本文提出了一种基于关键词扩展的答案块提取模型.该模型首先确定答案所在区块的中心词,即将文本与问题进行联合处理后计算问题关于联合向量的注意力值并按列输入softmax函数,将此概率分布矩阵逐列相加后遍历全文,检索出答案所在区块的中心词.然后,以该词为中心进行答案块扩展,并在每次扩展后计算答案块与问题向量之间的相似程度,相似度开始减小时停止扩展以优化候选答案块的质量.相较于以往的答案块提取模型,该模型一方面不再依赖于词性标注,另一方面大大提高了答案块的生成效率,在简化模型的同时提高了机器阅读理解的准确性.实验结果表明,该模型在SQuAD测试数据集上的EM(Exact Match)和F1值均表现优异,分别获得了65. 7%和74. 3%的准确度.

For answering a question who has non-fixed length in machine reading comprehension,this paper proposes an answer chunk extraction model based on keywords’ extension. Firstly,the model determines the central word of the chunk where the answer exists. It calculates the joint vector of the passage and question,and applies a column-wise softmax function to get probability distributions in each column,where each column is an individual joint vector-level attention when considering a single query word. After adding this probability distribution matrix column by column,it traverses the full passage and retrieves the central word. Then,the model extends the answer chunk around the center word. Afterwards,it calculates the similarity between the answer chunk and the question,and stops extension when similarity starts to decrease. The purpose of this step is to optimize the quality of candidate answer chunks. Compared with the previous models which extract answer chunk,our approach no longer depends on part-of-speech tagging,but greatly improves the generation efficiency of answer chunks. Our approach improves the accuracy of machine reading comprehension while simplifying the model. Experiments on the SQuAD dataset show that our model achieves an excellent performance in both the EM( exact match) and F1 values,and the accuracy reaches 65. 7% and 74. 3%,respectively.