基于时间卷积网络的机器阅读理解

针对目前机器阅读理解任务中缺乏有效的上下文信息融合方式和丢失文本的整体信息等情况,提出基于时间卷积网络的机器阅读理解模型.首先将文本的单词转化成词向量并加入词性特征;接着通过时间卷积网络获取问题和文章的上下文表示;之后采用注意力机制来计算出问题感知的文章表示;最后由循环神经网络模拟推理过程得到多步预测结果,并用加权和的方式来综合结果得到答案.实验使用了SQuAD2.0数据集,在EM和F1值上相比基准实验分别提升了6.6%和8.1%,证明了该方法的有效性.

Machine read comprehension based on temporal convolutional network

Regarding most current studies in the task of machine read comprehension using recurrent neural network to encode text, lack of effective method for fusing the information of the above and below text, and the fact that the way of sequential processing over time may lose the overall information of the text, this paper proposes a model based on temporal convolutional network for machine read comprehension. The model first transforms words of text into word embeddings and adds part-of-speech features. Then the contextual representations of question and passage are obtained by applying temporal convolution network. Next the attention mechanism is used to get question-aware passage representations. Finally, the recurrent neural network simulates reasoning process to obtain the multi-step prediction results, and the answer is based on the weighted sum of the results.The experiment uses the SQuAD2.0 dataset, and the EM and F1 are 6.6% and 8.1% higher than the baseline experiments respectively on the development set, demonstrating the effectiveness of the proposed method.