基于增强问题重要性表示的答案选择算法研究

Question Based Importance Weighting Network for Answer Selection

摘要:

针对经典的文本匹配模型在问答系统中应用的缺陷和不足，提出了一种基于增强问题重要性表示网络BIWN的答案选择算法.目前，现有的答案选择模型普遍将问题句子和答案句子直接进行匹配，忽略了问题句子和答案句子中的噪声词对匹配的影响.针对这个问题，首先，利用自注意力机制修改问题句子中各个词的权重，生成“干净”的问题句子向量；然后，利用词级交互矩阵捕捉问题句子和答案句子之间的细粒度语义信息，从而有效地弱化了噪声词对正确答案的影响；最后，利用多窗口CNN提取特征信息得到预测结果.基准数据集上的对比实验表明，BIWN模型在答案选择任务的性能优于主流的答案选择算法，MAP值和MRR值提升了约0.7%~6.1%.

Abstract:

According to the defects of the classic text matching model in the question and answer system, a questionbased importance weighting network for answer selection is proposed. At present, the existing answer selection model generally matches the question sentence and the answer sentence directly, ignoring the influence of noise words in the question sentence and the answer sentence on the match. To solve this problem, the selfattention mechanism is firstly used to modify the weight of each word in the sentence to generate a "clean" question sentence vector. The wordlevel interaction matrix is then used to capture the finegrained semantic information between the question sentence and the answer sentence. It weakens the influence of noise words on the correct answer. Finally, the multiwindow CNN is used to extract the feature information to obtain the prediction result. The comparison experiments on benchmark datasets show that the performance of the BIWN model in the answer selection task is better than the mainstream answer selection algorithm, and the MAP value and MRR value are improved by about 0.7%～6.1%.Finally, the multi-window CNN is used to extract the feature information to obtain the prediction result. The comparison experiments on benchmark datasets show that the performance of the BIWN model in the answer selection task is better than the mainstream answer selection algorithm, and the MAP value and MRR value are improved by about 0.7%-6.1%.