基于卷积推理的多跳知识图谱问答算法

Multi-hop Knowledge Graph Question Answering Based on Convolution Reasoning

多跳问题相比于简单问题更符合人们日常的提问方式,同时,研究多跳知识图谱问答(KGQA)算法有助于智能问答系统的推广。然而,现有的多跳KGQA方法在2~3跳问题和不完整知识图谱上的答案推理能力较弱。针对这一问题,本文提出基于卷积推理的多跳KGQA算法。首先,为了获取更具表示能力的问题嵌入向量,本文根据问题与关系的语义相似性提出结合字符特征和语义特征的问题嵌入模型;而后,为了增强算法的长链接推理能力,提出基于卷积神经网络(CNN)的答案推理模型来抽取嵌入向量的高阶信息。实验结果显示,相比于已有的5种算法,本文算法在MetaQA数据集的2跳和3跳问题答案预测准确率分别提高了1.7和1.3个百分点,在不完整知识图谱的2跳和3跳问题上分别提高了9.4和9.3个百分点。

Compared with simple questions,multi-hop questions are more in line with people's daily questioning methods.At the same time,the research on the multi-hop knowledge graph question-answering(KGQA)algorithm is useful to enhance the intelligent question answering system.However,the existing multi-hop KGQA methods show weak answer reasoning ability in 2 and 3-hop questions and incomplete knowledge graph.To solve this problem,a multi-hop KGQA based on convolution reasoning is proposed in this paper.A question embedding model combining character features and semantic features is developed according to the semantic similarity between questions and relationships to obtain more expressive question embedding.Furthermore,to enhance the long link reasoning ability of the algorithm,an answer reasoning model based on convolutional neural network(CNN)is proposed to extract the high-order information of the embedding.The experimental results on MetaQA dataset demonstrate that compared with the five existing methods,the new algorithm improves the prediction accuracy of the 2-hop and 3-hop questions in the complete knowledge graph and incomplete knowledge graph by 1.7%,1.3%,9.4%,and 9.3%,respectively.