Outperforming State-of-the-Art Systems for Aspect-Based Sentiment Analysis

Aspect-Based Sentiment Analysis (ABSA) is a very important problem with numerous applications. The three editions of SemEval's ABSA Shared Task have been instrumental in fostering the development in this field. One of its sub-tasks is the sentence-level ABSA. This sub-task has received a lot of attention and new techniques and better results are reported on it frequently. The purpose of this work is to achieve the highest accuracy for this problem. We follow a state-of-the-art (SOTA) approach that is based on multi-grain attention networks and infuse it with better embedding mechanisms in order to improve the results. For the famous SemEval's ABSA Shared Task, the results of the SOTA approaches reach 81.25 accuracy and 71.94 F1 score, whereas our approach surpasses them with 83.75 accuracy and 75.75 F1 score.