基于图卷积和自然实体知识的问答式情感分析研究

Question-answering Sentiment Analysis Based on Graph Convolution and Natural Entity Knowledge

该文针对前人研究中依存语法的学习建模过程与文本情感特征提取模块完全独立的问题,考虑到很难根据新数据进行整体更新,结合引入自然实体知识至神经网络的新方式,为问答式情感分析任务提出了结合自然语言知识和图卷积网络的双重网络模型E-QAGCN。模型考虑了词语间的相互依赖关系,并基于依存语法树的结构特征,将依赖句法关系融入语义向量表示中,联合自然语言知识和多层图卷积网络模型对文本进行了深层次的情感特征提取。此为基于依存语法树和图卷积操作结构的方法第一次在问答式情感分析任务中被应用,在三个领域的中文问答式情感分析数据集上的实验结果验证了该模型的有效性,达到了同期的最优效果。

To deal with the issue that dependency grammar learning process and text sentiment feature extraction is independent,considering the difficulty in making overall updates based on new data.This paper put forward a dual network model E-QAGCN to introduce the natural entity knowledge into the neural network.Based on the structural features of text dependency tree,the model takes into account the inter-relationship between words,and combines the knowledge of natural language to extract the deep sentiment features of the text through multi-layer graph convolution network.This is the first time that the dependency tree-based sentiment feature extraction from text via GCN has been applied in question-answering Sentiment Analysis task.The experimental results on three Chinese datasets prove the effectiveness of the model by achieving the best result.