基于ACSA-CapNet的方面级情感分析研究

A Study on Aspect-Level Sentiment Analysis Based On ACSA-CapNet

摘要: 方面情感分析(Aspect Based Sentiment Analysis, ABSA)是指预测给定方面的情感极性，已引起广泛关注。在现有的ABSA数据集中，大多数句子只包含一个方面或具有相同情感极性的多个方面，这使得ABSA任务退化为句子层面的情感分析。本文中，我们采用了一个新的大规模多角度多情绪(MAMS)数据集，其中每个句子至少包含两个具有不同情感极性的不同方面，但该数据集类别分布极其不平衡，导致模型准确率受到某个类别的影响。为了解决这一问题，我们在原数据集的基础上进行了数据增强，并提出了一个新的MAMS的数据集，并将其命名为ACSA-V2。此外，我们在CapsNet模型的基础上，对压缩函数进行了修改，使得模型在提升准确率的同时可以更快地收敛。实验结果显示，相较于江等人的胶囊网络模型，基于改进的胶囊模型在方面级情感分析任务上准确率提高了约5个点。

Abstract: Aspect Based Sentiment Analysis (ABSA), which refers to the prediction of the sentiment polarity of a given aspect, has attracted a great deal of attention. In existing ABSA datasets, most sentences contain only one aspect or multiple aspects with the same sentiment polarity, which degrades the ABSA task to sentence-level sentiment analysis. In this paper, we use a new large-scale multi-angle multi-emotional (MAMS) dataset in which each sentence contains at least two different aspects with different sentiment polarity, but the distribution of categories in this dataset is extremely unbalanced, resulting in model accuracy being influenced by a particular category. To address this issue, we augmented the original dataset with data and proposed a new dataset for MAMS, which was named ACSA-V2. In addition, we modified the compression function based on the CapsNet model so that the model could converge faster while improving accuracy. The experimental results show that the improved CapsNet-based model improves accuracy by about 5 points on the aspect-level sentiment analysis task compared to Jiang et al.’s CapsNet model.