融合文本概念化与网络表示的观点检索

Opinion Retrieval Method Combining Text Conceptualization and Network Embedding

观点检索是自然语言处理领域中的一个热点研究课题.现有的观点检索模型在检索过程中往往无法根据上下文将词汇进行知识、概念层面的抽象,在语义层面忽略词汇之间的语义联系,观点层面缺乏观点泛化能力.因此,提出一种融合文本概念化与网络表示的观点检索方法.该方法首先利用知识图谱分别将用户查询和文本概念化到正确的概念空间,并利用网络表示将知识图谱中的词汇节点表示成低维向量,然后根据词向量推出查询和文本的向量,并用余弦公式计算用户查询与文本的相关度,接着引入基于统计机器学习的分类方法挖掘文本的观点.最后,利用概念空间、网络表示空间以及观点分析结果构建特征,并服务于观点检索模型.相关实验结果表明,所提出的检索模型可以有效提高多种检索模型的观点检索性能.其中,基于统一相关模型的观点检索方法在两个实验数据集上相比于基准方法,在MAP评价指标上分别提升了6.1%和9.3%,基于排序学习的观点检索方法在两个实验数据集上相比于基准方法,在MAP评价指标上分别提升了2.3%和14.6%.

Opinion retrieval is a hot topic in the research of natural language processing. Most existing approaches in text opinion retrieval can not extract knowledge and concept from context. They also lack opinion generalization ability and overlook the semantic relations between words. This paper proposes an opinion retrieval method based on knowledge graph conceptualization and network embedding. First, conceptual knowledge graph is used to conceptualize the queries and texts into the correct conceptual space while the nodes in the knowledge graph are embedded into low dimensional vectors space by network embedding technology. Then, the similarity between queries and texts is calculated based on embedding vectors. According to the similarity score, the opinion scores of texts can be captured based on statistical machine learning methods. Finally, the concept space, knowledge representation space, and opinion mining result serve opinion retrieval models. The experiment shows that the retrieval model proposed in this paper can effectively improve the retrieval performance of multiple retrieval models. Compared with referenced method based on unified opinion, the proposed approach improves the MAP scores by 6.1% and 9.3%, respectively. Compared with referenced method based on learning to rank, proposed approach improves the MAP scores by 2.3% and 14.6%, respectively.