Learning multiple layers of knowledge representation for aspect based sentiment analysis

Abstract

Sentiment Analysis is the task of automatically discovering the exact sentimental ideas about a product (or service, social event, etc.) from customer textual comments (i.e. reviews) crawled from various social media resources. Recently, we can see the rising demand of aspect-based sentiment analysis, in which we need to determine sentiment ratings and importance degrees of product aspects. In this paper we propose a novel multi-layer architecture for representing customer reviews. We observe that the overall sentiment for a product is composed from sentiments of its aspects, and in turn each aspect has its sentiments expressed in related sentences which are also the compositions from their words. This observation motivates us to design a multiple layer architecture of knowledge representation for representing the different sentiment levels for an input text. This representation is then integrated into a neural network to form a model for prediction of product overall ratings. We will use the representation learning techniques including word embeddings and compositional vector models, and apply a back-propagation algorithm based on gradient descent to learn the model. This model consequently generates the aspect ratings as well as aspect weights (i.e. aspect importance degrees). Our experiment is conducted on a data set of reviews from hotel domain, and the obtained results show that our model outperforms the well-known methods in previous studies.