Unsupervised Improving of Sentiment Analysis Using Global Target Context

Proceedings of Recent Advances in Natural Language Processing, pages 122–128, Hissar, Bulgaria, 7-13 September 2013. Unsupervised Improving of Sentiment Analysis Using Global Target Context Tom´aˇs Brychc´ın Ivan Habernal NTIS – New Technologies for the Information Society, and Department of Computer Science and Engineering, Faculty of Applied Sciences, University of West Bohemia, Univerzitn´ı 8, 306 14 Plzeˇn, Czech Republic {brychcin,habernal}@kiv.zcu.cz Abstract Current approaches to document-level sentiment analysis rely on local informa- tion, e.g., the words within the given doc- ument. We try to achieve better perfor- mance by incorporating global context of the sentiment target (e.g., a movie or a product). We assume that sentiment la- bels of reviews about the same target are often consistent in some way. We model this consistency by Dirichlet distribution over sentiment labels and use it together with Maximum entropy classifier to gain significant improvement. This unsuper- vised extension increases the classification F-measure by almost 3% absolute on both Czech and English movie review datasets and outperforms the current state of the art. 1 Introduction Sentiment analysis on the document level has been one of the most targeted research topic in the past decade (Liu and Zhang, 2012). Given a document (e.g. a review, a blog post, or a tweet), the goal is to automatically obtain its sentiment which is mostly considered as a binary value (positive and negative) or is more granular (e.g. positive, neg- ative, and neutral or a number on the pre-defined scale). Since the pioneering research by Pang et al. (2002), movie reviews have represented a very popular domain for evaluating sentiment analy- sis systems, mainly because of abundance of la- beled data from existing on-line movie databases.1 1One might argue that if movie or product databases al- ready contain reviews labeled with e.g. number of stars, it is useless to try to estimate it automatically; however, not all databases are alike, e.g., the Polish movie database has no such star rating and contains only pure text reviews. Large datasets are crucial for employing machine learning approaches. Both approaches to sentiment analysis (machine learning-based and vocabulary-based) attempt to estimate the polarity of the document taking into account only its content (e.g. words, morphology patterns, syntax, and other features). Other exter- nal information, such as the sentiment target, the author, and others, are mostly ignored in the po- larity estimation step. This means that the distri- bution of sentiment for each target is considered as random. We assume that sentiment labels for each tar- get are not independent of each other. This means that given a movie with the majority of positive reviews, there is a chance that the next unknown review will be positive as well. We model this as- sumption as a Dirichlet distribution over sentiment labels for each target. In summary, our approach to sentiment analysis consists of two steps. In the first step, we employ a supervised Maximum en- tropy classifier in order to estimate sentiment la- bel probabilities for each review. In the second (unsupervised) step, these …

Content truncated to 500 words