基于问答交互的答案句选择算法

答案选择任务的精度对问答系统、文本处理等应用的效果具有重要的影响。针对答案选择模型问句与候选答案句语义信息和句子浅层特征利用不充分的问题，提出一种基于问答句交互的答案选择模型。给定问句Q和候选答句A，模型首先使用BiLSTM编码器对它们进行编码，然后针对问句Q使用Feed-Forward注意力机制得到句子编码；针对答句A，将问句Q和答句A的所有时间步输出两两进行匹配,根据匹配结果计算出答句的每个单词的权重，进而加权计算出答句的句子编码。最后，将问答句的句子编码经过聚合操作后输入全连接层，并与词共现特征相融合输出最终判断结果。在DBQA数据集上的实验结果表明，该模型与主流的Siamese结构的神经网络相比，能够有效地提升答案选择任务的效果。

Algorithm of Answer Sentence Selection Based on Q and A Interaction

The accuracy of answer selection task has an important influence on the application of question answering system, text processing and so on. In order to solve the problem that the semantic information and the shallow features of question and the candidate answer are not fully utilized in the answer selection model, an answer selection model based on the interaction of question and answer is proposed. Given question Q and candidate answer A, the model first uses BiLSTM encoder to encode them, concatenates the two directions encode results of each time step of Q and A, and then uses Feed-Forward attention to encode question sentence Q; all the time steps of question Q are matched with all the time steps of A. According to the matching results, the weight of each word of the answer sentence is calculated, and then the sentence code of the answer sentence is calculated according to the weight of each word of the answer sentence. Finally, the sentence coding of Q and A sentences is input into the full connection layer after the aggregation operation, and the final judgment result is output by fusion with the co-occurrence feature of words. The experimental results on the DBQA data set show that this model can effectively improve the effect of answer selection task compared with the mainstream Siamese neural network.