Abstract: Fine-grained aspect-based sentiment analysis involves aspect term extraction and aspect sentiment classi-fication. Most existing research methods address them in an independent fashion, which lack a mechanism to account for the relevant information between each other, resulting in training redundancy and waste of resources. To solve the above problems, a joint model based on position embedding and graph convolutional network (PE-GCN) under the framework of multi-task learning is proposed, which is an end-to-end approach to the overall solution of aspect term extraction and aspect sentiment classification. Firstly, the model learns the semantic feature representation of sentence through a bidirectional gated recurrent unit network. Then, it exploits positional embedding to enhance the recognition of aspect terms in sentence, and uses the graph convolutional network to generate a contextual representation containing syntactic information. Finally, interactive attention network is used to model the semantic relationship between context and aspect terms, and the sentiment polarity of aspect terms is output through softmax. Experimental results on the SemEval-2014 public datasets show that the performance of the proposed model has a significant improvement compared with other existing models.

Key words: aspect term extraction, aspect sentiment classification, position embedding, graph convolutional network, interactive attention

摘要： 基于方面的细粒度情感分析包括方面术语抽取和方面情感分类两项任务，以独立方式解决以上两项任务的研究方法无法利用彼此之间的关联信息，同时也会造成训练冗余和资源浪费。针对上述问题，在多任务学习框架下提出一种基于位置嵌入和图卷积网络的联合模型（PE-GCN），以端到端方式整体解决方面术语抽取和方面情感分类。该模型首先通过双向门控循环单元网络学习句子的语义特征表示；随后利用位置嵌入增强句子中方面术语的识别，同时使用图卷积网络生成包含句法信息的上下文表示；最后通过交互注意力网络建模上下文和方面术语之间的语义关系，并通过softmax输出方面术语的情感极性。在SemEval-2014公开数据集上的实验结果表明，提出的模型与其他现有模型相比性能有显著的提升。

关键词: 方面术语抽取, 方面情感分类, 位置嵌入, 图卷积网络, 交互注意力