摘要： 现有基于序列标注或文本生成的三元组抽取模型通常未考虑完整文本片段级别的交互，且忽略了句法知识的应用。为解决上述问题，提出一种基于依存图卷积与文本片段搜索的深度学习模型来联合抽取方面情感三元组。通过预训练语言模型BERT编码层学习句子中每个单词的上下文表达，同时利用图卷积神经网络学习句子单词之间的依存关系和句法标签信息，以捕获远距离的方面词与观点词之间的语义关联关系，并采用文本片段搜索构造候选方面词与观点词及其特征表示，最终使用多个分类器同时进行方面词与观点词抽取及情感极性判断。在ASTE-Data-V2数据集上的实验结果表明，该模型在14res、14lap、15res和16res子集上的F1值相比于JET模型提升了10.61、10.54、4.91和8.48个百分点，具有较高的方面情感三元组抽取效率。

关键词: 方面情感三元组抽取, 图卷积神经网络, 深度学习, 依存句法分析, 文本片段搜索

Abstract: Existing studies on the extraction of aspect sentiment triples mostly employ methods based on sequence tagging or text generation. These methods consider neither the interactions at the span pair level nor the application of syntactic knowledge. To address these problems, this study proposes a deep learning model based on dependency graph convolution and text span search to jointly extract aspect sentiment triples. The model first learns the contextual representation of each word in a sentence through the Bidirectional Encoder Representations from Transformers(BERT) coding layer. It then utilizes a graph convolutional neural network to learn the dependencies and syntactic label information in words to capture the semantic associations between distant aspect terms and opinion terms. It also uses text span search to construct candidate aspect and opinion terms and their feature representations. Finally, the model extracts aspect and opinion terms and sentiment polarity simultaneously using multiple classifiers. Experiments on the Aspect Sentiment Triple Extraction(ASTE)-Data-V2 dataset show that the model improves F1 scores by 10.61, 10.54, 4.91, and 8.48 percentage points on the 14res, 14lap, 15res, and 16res subsets, respectively, as compared with the JET model, thus demonstrating its effectiveness.

Key words: Aspect Sentiment Triple Extraction(ASTE), graph convolution neural network, deep learning, dependency syntactic parsing, text span search

中图分类号: 

• TP391