Effective Attention Networks for Aspect-level Sentiment Classification

This paper deals with the aspect-level sentiment classification which identifies the sentiment polarity of a specific aspect of its context. We introduce novel attention networks by using the benefits of Long Short-Term Memory (LSTM), Attention mechanisms and Lexicons to form an aspect-specific representation. Though a variety of neural network models have been proposed recently, however, previous models have captured the importance of aspects in their contexts and developed various methods by modeling their contexts via generating aspect representations. In this paper, aspects and their contexts are treated separately and learned their own representations. Additionally, the purpose of lexicons is to highlight the important sentiment words of aspects and their contexts. The relation between aspects and their contexts are explored by concentrating on different parts of a sentence when different aspects are taken as input. We evaluate our models on Laptop and Restaurant datasets and show that our approaches improve classification accuracy in aspect-level sentiment classification.