Sentiment Aggregation using ConceptNet Ontology

International Joint Conference on Natural Language Processing, pages 570–578, Nagoya, Japan, 14-18 October 2013. Sentiment Aggregation using ConceptNet Ontology Subhabrata Mukherjee and Sachindra Joshi IBM India Research Lab {subhabmu, jsachind}@in.ibm.com Abstract Sentiment analysis of reviews traditionally ig- nored the association between the features of the given product domain. The hierarchical re- lationship between the features of a product and their associated sentiment that influence the polarity of a review is not dealt with very well. In this work, we analyze the influence of the hierarchical relationship between the prod- uct attributes and their sentiments on the over- all review polarity. ConceptNet is used to automatically create a product specific ontol- ogy that depicts the hierarchical relationship between the product attributes. The ontology tree is annotated with feature-specific polari- ties which are aggregated bottom-up, exploit- ing the ontological information, to find the overall review polarity. We propose a weakly supervised system that achieves a reasonable performance improvement over the baseline without requiring any tagged training data. 1 Introduction In recent years there has been a huge surge of activity in the social networking sites, blogs and review sites. The voluminous amount of data generated is a goldmine of information for the re- tail brands to find out the customer needs, con- cerns and potential market segments. Sentiment analysis aims to mine this information to find out the popular sentiment about any product and its associated features. Traditionally sentiment analysis has been posed as a text classification task on features de- rived from the given text. In the product review domain, the initial works in sentiment analysis focused on classifying the entire review as posi- tive or negative using various word-based and phrase-based features (Turney et al., 2003; Tur- ney 2002; Kamps et al., 2002; Hatzivassiloglou et al., 2000; Hatzivassiloglou et al., 2002). The more recent works focused on product feature extraction from a review and performing feature- specific sentiment analysis (Hu et al., 2004; Mukherjee et al., 2012). For example, the re- view, The audio quality of my new phone is ab- solutely awesome but the picture taken by the camera is a bit grainy, is positive with respect to the audio quality and negative with respect to the camera. However, once the feature-specific po- larities are obtained, the works do not describe any systematic approach to aggregate the feature- specific polarities to obtain the overall review polarity. A naïve count-based feature-specific polarity aggregation will not work well for re- views having different features with diverse opinions. A bag-of-words based model will pick up awesome and grainy as the sentiment features and mark the overall review as neutral. One may argue that the audio quality is more important to a cell phone than the camera and hence the over- all review polarity should be positive. While the feature-specific model associates sentiment to features, it cannot do a polarity aggregation in absence of feature association information to find the overall review polarity. Let us consider the following review taken from Amazon.com which …

Content truncated to 500 words