多通道特征融合的实体链接模型泛化性能优化

Generalization Performance Optimization of Entity Link Models Based on Multi-Channel Feature Fusion

实体链接是知识库问答和知识图谱构建的关键环节,中文语料库的语义表达稀疏,存在大量难以区分的相似实体,一般模型过于依赖除原始问答以外的特征信息,很难完全学习文本特征,使得实体链接准确率难以提高,进而限制了问答等上层应用的性能上限。针对这些问题,聚焦问答系统实体链接的候选生成和候选消歧,将实体消歧视为分类任务,构建了一种基于Bi-LSTM和CNN的多通道网络模型,提出阈值权重拼接策略融合CNN和LSTM通道提取的多维特征。引入双向注意力机制,充分挖掘问句提及表征和知识库实体描述之间的深层语义关系,有效降低问答对额外特征规则的依赖,以便应用在多领域知识库中。实验结果表明,在仅依靠问答原始信息的情况下,提出的实体链接模型显著提高了问答系统的整体性能,并具有较强的泛化性,在公开数据集CCKS2019-CKBQA和NLPCC-2016KBQA中取得了最优的Acc@1和F1值。

Entity linking is a key link in knowledge base question answering and knowledge graph construction.The semantic expression of the Chinese corpus is sparse,and there are a large number of similar entities that are difficult to distinguish.The general model is too dependent on the feature information other than the original question and answer.It is difficult to fully learn text features,which makes it difficult to improve the accuracy of entity linking,in turn limits the performance upper limit of upper-level applications such as question answering.Aiming at these problems,this paper focuses on the candidate generation and candidate disambiguation of entity links in question answering systems,and regards entity disambiguation as a classification task.A multi-channel network model based on Bi-LSTM and CNN is constructed,and a threshold weight splicing strategy is proposed to fuse the multidimensional features extracted by CNN and LSTM channels.The bidirectional attention mechanism is introduced to fully mine the deep semantic relationship between the question mention representation and the knowledge base entity description,and effectively reduce the dependence of question answering on additional feature rules,so that it can be applied in multi-domain knowledge bases.The experimental results show that the proposed entity linking model significantly improves the overall performance of the question answering system,and has strong generalization under the condition of only relying on the original information of question answering.The optimal Acc@1 and F1 values are obtained on CCKS2019-CKBQA and NLPCC-2016KBQA datasets.