Effective Self Attention Modeling for Aspect Based Sentiment Analysis

Aspect Based Sentiment Analysis is a type of fine-grained sentiment analysis. It is popular in both industry and academic communities, since it provides more detailed information on the user generated text in product reviews or social network. Therefore, we propose a novel framework based on neural network to determine the polarity of a review given a specific target. Not only the words close to the target but also the words far from the target determine the polarity of the review given a certain target, so we use self attention to solve the problem of long distance dependence. Briefly, we do multiple linear mapping on the review, do multiple attention and combine them to attend to the information from different representation sub-spaces. Besides, we use domain embedding to get close to the real word embedding in a certain domain, since the meaning of the same word may be different in different situation. Moreover, we use position embedding to underline the target and pay more attention to the words that are close to the target to get better performance on the task. We validate our model on four benchmarks, they are SemEval 2014 restaurant dataset, SemEval 2014 laptop dataset, SemEval 2015 restaurant dataset and SemEval 2016 restaurant dataset. The final results show that our model is effective and strong, which brings a 0.74% boost averagely based on the previous state-of-the-art work.