Enhancing Neural Sentiment Analysis with Aspect Weights

Sentiment analysis is a challenging task and has impactful applications, including analyzing customer feedback on social media. In this paper, we propose a novel approach which enhances a neural architecture to predict the overall sentiment of restaurant reviews which may contain multiple aspect-level sentiments. We calculate the weights of different aspects of a restaurant and incorporate them in a neural architecture. We also compare our results with the current state-of-the-art approach (ULMFiT [1]) and show an absolute improvement of 7% in the F-score and 6% in the accuracy. To the best of our knowledge, this is the first work in the line of research investigating the incorporation of aspect weights into a neural architecture for sentiment analysis, culminating in a detector thereof.