鞍山师范学院学报

文本观点检索旨在从特定文档集中检索出与查询相关的观点文本，过程中涉及查询-文本相关性及文本观点特征的提取.现有模型在文档倾向性特征提取时，主要采用传统的机器学习方法进行观点特征挖掘，该方法忽视了文本中词语节点间关系对判断观点倾向的影响，导致了模型在文本语义层面的泛化能力受限，进而影响了观点检索的性能.提出一种基于图卷积网络的观点检索方法，旨在将文档中的词语映射到高维语义空间中，自适应地学习文档词语节点间的连接关系，以获取更准确的文档语义信息表述，应用于模型之中.实验表明，与当前最优的模型相比，改进后的模型在两个Twitter数据集上，检索主指标分别提升1.2%和0.8%,验证了本文所提方法的有效性.

Opinion retrieval text aims to retrieve the opinion documents related to a query from a specific document set, which involves the combination of query and document relevance as well as the orientation of the document itself.The existing models focus on mining opinion features based on traditional machine learning when extracting document opinion features.These methods ignore the information representation of adjacent word nodes in the document, which leads to limited generalization ability at the semantic level of the text and cannot obtain more accurate opinion features, thus affecting the opinion retrieval performance.To solve the above problems, this paper proposes an opinion retrieval method based on graph convolutional network, which aims to map the words in the document into a high-dimensional semantic space, adaptively learn the connection relationship between the document word nodes, and obtain more accurate semantic information representation of the document.Experiments show that, compared with the current best model, the improved model improves the main index by 1.2% and 0.8% respectively on two Twitter public data, which well verifies the effectiveness of the proposed method.