Abstract:

Aspect-Based Sentiment Analysis(ABSA)is a fine-grained sentiment analysis task that aims to analyze the sentiment polarity of multiple specific aspects in a given text. Most of the current models based on syntactic analysis rely heavily on the single parsing result of the dependency tree. Since this structure only reveals the dependency relationship between two words, it cannot capture the connection between the aspect word and the context. In addition, most studies ignore the relationship types in dependency parsing, resulting in the possible loss of important feature information. In response to the above problems, a Multi-Channel-based Graph Convolutional Network (MCGCN) model is proposed. The proposed model mainly includes a component graph module, a dependency graph module (DepGCN) and a self-attention mechanism, which aims to mine the rich syntactic and grammatical information in the sentence structure. In addition, in order to reduce dependency parsing errors, the dependency types and dependency probabilities are introduced in DepGCN to obtain a more precise grammatical representation. MCGCN first uses the pre-trained model to embedding words to obtain text initialization vectors, which are respectively input to the graph convolutional network that integrates different syntactic analysis for learning of different nodes. The self-attention mechanism can effectively associate the global semantic information of aspect words. Finally, the output of the three modules is interactively learned through affine, thereby improving the representation ability of the model. The experimental results show that the ACC value and F1 value of the model on the Laptop and Twitter datasets are improved compared with the baseline model, reaching 78.48%, 75.03% and 75.92%, 74.53%, respectively, and the F1 value on the three datasets has at least a 1%-3% improvement.

Key words: Aspect-Based Sentiment Analysis(ABSA), syntactic analysis, Graph Convolutional Network(GCN), dependency tree, self-attention

摘要：

方面级情感分析是一项细粒度的情感分析任务，旨在对给定文本中多个特定方面进行情感极性分析。大多数基于句法分析的模型严重依赖依存树的单一解析结果，由于这种结构仅揭示2个单词之间的依存关系，因此无法捕捉方面词与上下文之间的联系。此外，大部分研究忽略了依存解析中的关系类型，可能丢失重要特征信息。提出一种基于多通道的图卷积网络（MCGCN）模型，主要包含成分图模块、依存图模块（DepGCN）以及自注意力机制，旨在挖掘句子结构中丰富的句法和语法信息。此外，为减少依存解析错误，在DepGCN中引入依存关系类型和依存概率以获取更精准的语法表示。利用预训练模型进行词嵌入得到文本初始化向量，分别输入到融合不同句法分析的图卷积网络进行不同节点的学习，利用自注意力机制能有效地对方面词进行全局的语义信息关联。最后将3个模块的输出通过仿射交互层进行交互学习，进而提升模型的表征能力。实验结果表明，MCGCN模型在Laptop数据集上的ACC值和F1值分别为78.48%和75.03%，在Twitter数据集上的ACC值和F1值分别为75.92%和74.53%，相较于基准模型均有提升，在3个数据集上的F1值都有1%~3%的提升。

关键词: 方面级情感分析, 句法分析, 图卷积网络, 依存树, 自注意力