Polarity-aware deep attention network for aspect-based sentiment analysis

Deep Attention Neural Networks have revolutionized the way we approach complex learning tasks such as Sentiment Analysis. However, existing methods mainly focus on coarse-grained sentiment classification, i.e., identifying emotions by assigning text to a dominating discrete sentiment category. This research is an effort to tackle a Sentiment Analysis task to improve sentiment granularity at the aspect level. Meanwhile, Deep Learning techniques have emerged as a potential solution for accomplishing Aspect-based Sentiment Analysis (ABSA). The proposed two-stage model is implemented in two consecutive stages: (1) Aspect word extraction and (2) Sentiment classification based on aspect words. Unlike traditional Machine Learning methods, Deep Learning techniques can capture syntactic and semantic features without high-level feature engineering. The first task, aspect word extraction, is achieved through a Bi-LSTM-CRF model, which resembles a Named Entity Recognition (NER) system. However, the majority of existing NER solutions are specific to a particular domain and rely on hand-crafted features to train Supervised Learning models. Therefore, a novel neural architecture for aspect word extraction that combines Bi-LSTM and Conditional Random Fields is presented. However, the second task, i.e., sentiment classification, is accomplished through a novel Polarity-aware Deep Attention Network (PDAN) model implemented using Bi-GRU. The notion behind the proposed PDAN model is that when an aspect word appears in a sentence, it can be replaced by its corresponding polarity embedding. Furthermore, this representation is utilized for performing ABSA. Two benchmark datasets are employed to demonstrate the performance of the proposed model, and the results are compared with other baseline techniques. The results show that the proposed model outperforms many state-of-the-art models.