Not All Words Are Created Equal: Extracting Semantic Orientation as a Function of Adjective Relevance

Semantic orientation (SO) for texts is often determined on the basis of the positive or negative polarity, or sentiment, found in the text. Polarity is typically extracted using the positive and negative words in the text, with a particular focus on adjectives, since they convey a high degree of opinion. Not all adjectives are created equal, however. Adjectives found in certain parts of the text, and adjectives that refer to particular aspects of what is being evaluated have more significance for the overall sentiment of the text. To capitalize upon this, we weigh adjectives according to their relevance and create three measures of SO: a baseline SO using all adjectives (no restriction); SO using adjectives found in ontopic sentences as determined by a decision-tree classifier; and SO using adjectives in the nuclei of sentences extracted from a high-level discourse parse of the text. In both cases of restricting adjectives based on relevance, performance is comparable to current results in automated SO extraction. Improvements in the decision classifier and discourse parser will likely cause this result to surpass current benchmarks.