基于多角度注意力机制的单一事实知识库问答方法

Multi-view Attentional Approach to Single-fact Knowledge-based Question Answering

近年来,基于知识库的问答受到了广泛的关注,成为了一个重要的自然语言处理任务。在基于知识库的问答任务中,简单问题是指能够通过知识库的单一事实进行回答的问题。针对简单问题的回答,现有的解决方法主要是将问题和知识库事实映射到同一向量空间中,然后通过计算问题和事实之间的相似度来得到答案,但这种方法会损失原始单词的部分语义交互信息。针对该问题,文中提出了一种基于多角度注意力机制的关系检测模型,从多个角度对问题和知识库关系的相关性进行了建模,从而保留了更多的原始交互信息,并提高了模型的准确率。此外,为了减小噪音的影响并提高实体识别的准确率,在实体链接过程中提出结合基于语言模型的动态词向量和单词的词性特征对问题进行表征。实验结果表明,所提方法在基于FB2M和FB5M的SimpleQuestions数据集上分别获得了78.9%和78.3%的准确率,能够很好地反映问题与知识库关系之间的语义相关性,并提升了单一事实知识库问答的准确率。

Knowledge base question answering (KB-QA) has received extensive attention in recent years,and becomes an important natural language processing task.In the knowledge base question answering task,simple question refers to the question that can be answered by a single-fact of the knowledge base.For this task,the existing approaches mainly map the question and the KB fact into a common vector space and calculate their similarity to get the answer.But this approach would lose part of the semantic interaction information of the original words.To solve this problem,a multi-view attention-based relation detection approach was proposed,which aims to model the correlation between the question and the KB relation from multiple perspectives,and preserves more original interaction information so as to improve the accuracy of the approach.In addition,in order to alleviate the impacts of noisy data and improve the accuracy of entity recognition,this paper also encoded the question by combining the dynamic word vectors based on the language model and the part-of-speech feature of the word during the process of entity linking.Finally,experiments conducted on the SimpleQuestions dataset based on FB2M and FB5M achieve the accuracy results of 78.9% and 78.3%,respectively,which illustrative the effectiveness of the proposed approach for reflecting the semantic correlation between the question and the KB relation,and reflect the improvements of the accuracy for single-fact KBQA.