Intra-Attention and Inter-Attention for Aspect-level Sentiment Classification

Aspect-level sentiment classification is an important tool for e-business platforms to help the user and suppliers to summary the fine-grained aspect-based sentiment opinions. The same review document may have different sentiment polarity for each aspect. Existing aspect-level sentiment classification usually learn the aspect-based sentiment polarity with traditional supervised learning and deep learning models with inter-attention, which may be insufficient. In this paper, we propose an intra- and inter-attention network model to predict the aspect-level of sentiment category. The core of our model is the representation module, intra-attention module and inter-attention module. In the representation module, we use the character-level semantic and word-level semantic information, respectively. In the intra-attention module, we use the intra-attention to capture the sentiment features relevant among each word. In addition, we use the aspect-sentiment pair attention to represent the sentiment features of the document under the specific aspect information and sentiment features in the inter-attention module. Experiments on four real-world datasets show that our proposed model outperforms the state-of-the-art baselines in the prediction accuracy.