Classifying semantic orientation of domain-dependent words with unknown sentiments

Interpretation of semantic orientation of a word depends on the domain topic it describes. Based on Semantic Orientation Pointwise Mutual Information (SO-PMI), we propose a framework for prediction of semantic orientations of words with respect to a domain topic. The framework exploits the number of hits obtained from an available search engine for calculating the SO-PMI value of a given pair of a domain topic and a word with unknown sentiment. For improvement of prediction accuracy, the framework adjusts SO-PMI values by employment of tuning parameters, which are learnt automatically from training data. The framework is evaluated on two different domain topics and the overall accuracy in the range of 66%-78% is obtained.