面向细粒度情感分类任务的双通道分类模型
Dual-Channel Classification Model for Fine-Grained Sentiment Classification Task

句法信息对情感分类任务十分重要，使用GCN来建模这种信息有助于模型关注情感关键词。然而这类模型仅是利用基本语义信息辅助学习句法信息，且单一地从句法依存角度捕获情感关键词，忽略了从语义角度发掘情感关键词。另外，此类模型过于依赖句法信息，没考虑到使用句法提取工具对分类效果造成的负面影响。针对以上问题，提出一种双通道分类模型。该模型利用双通道分类结构减弱对于句法信息的依赖性，同时采用语义情感通道从语义上捕获情感关键词，进而提升模型获取情感信息的能力。在两个常用中文情感分类数据集上的实验表明，该模型的Micro_F值和Macro_F值相较于现有模型均有提升，模型对比和消融实验验证了双通道分类结构在提升模型分类任务性能上的有效性。
Syntactic information plays an important role in sentiment classification, using GCN to model the information can help the model learn sentiment keywords. However, such models only use semantic information to assist learning syntactic information, and capture sentiment keywords from the perspective of syntactic dependency, ignoring the semantic perspective. In addition, such models rely on syntactic information and do not consider the negative impact of using syntax extraction tools on classification results. Giving the aforementioned issues, a dual-channel classification model is proposed. The model uses a dual-channel classification to reduce the dependence on syntactic information, and adopts an attention mechanism to capture semantic sentiment words, thereby improving the ability of the model to obtain sentiment information. Experiments on two commonly used Chinese sentiment classification datasets show that both Micro_F and Macro_F are improved compared with existing model. Comparative and ablation experiments illustrate the effectiveness of dual-channel classification structure to improve the model’s classification performance.