Inducing Domain-specific Noun Polarity Guided by Domain-independent Polarity Preferences of Adjectives

In this paper, we discuss how domainspecific noun polarity lexicons can be induced. We focus on the generation of good candidates and compare two machine learning scenarios in order to establish an approach that produces high precision. Candidates are generated on the basis of polarity preferences of adjectives derived from a large domain-independent corpus. The polarity preference of a word, here an adjective, reflects the distribution of positive, negative and neutral arguments the word takes (here: its nominal head). Given a noun modified by some adjectives, a vote among the polarity preferences of these adjectives establishes a good indicator of the polarity of the noun. In our experiments with five domains, we achieved f-measure of 59% up to 88% on the basis of two machine learning approaches carried out on top of the preference votes.