Expanding Opinion Lexicon with Domain Specific Opinion Words Using Semi-Supervised Approach

Opinion words as well as opinion phrases and idioms are very useful in sentiment analysis. All these terms together build opinion or sentiment lexicons. Therefore, opinion lexicons are large lists of terms that encode the sentiment of each phrase within it. Generally, to create such a lexicon automatically, high-precision classifiers use known sentiment vocabulary, e.g. the prior polarity of an adjective at word-level, to separate corresponding phrases from a non-annotated text collection. Most unsupervised approaches try to determine prior polarity, also called semantic orientation, of adjectives. However, adjective phrases or verb phrases are useful indicators of sentiment as well. To build domain independent opinion lexicons classifiers need to be applied to a high number of corpora regarding different text categories. This introduces the challenge of ambiguity, as opinion terms or phrases often show different sentiment when used in various sorts of texts. Therefore, a tradeoff which takes the most applicable sentiment in regards of a general domain has to be developed in such a case. In this paper we show a novel approach to extract domain specific adjectives from the Twitter corpus and expand the general lexicon. We build an undirected weighted graph of the adjective pairs, and use the weighted adjacency matrix as input of the clustering algorithm.