Detecting and Blocking False Sentiment Propagation

Abstract

Sentiment detection of a given expression involves interaction with its component constituents through rules such as polarity propagation, reversal or neutralization. Such compositionality-based sentiment detection usually performs better than a vote-based bag-ofwords approach. However, in some contexts, the polarity of the adjectival modifier may not always be correctly determined by such rules, especially when the adjectival modifier characterizes the noun so that its denotation becomes a particular concept or an object in customer reviews. In this paper, we examine adjectival modifiers in customer review sentences whose polarity should either be propagated (SHIFT) or not (UNSHIFT). We refine polarity propagation rules in the literature by considering both syntactic and semantic clues of the modified nouns and the verbs that take such nouns as arguments. The resulting rules are shown to work particularly well in detecting cases of ‘UNSHIFT’ above, improving the performance of overall sentiment detection at the clause level, especially in ‘neutral’ sentences. We also show that even such polarity that is not propagated is still necessary for identifying implicit sentiment of the adjacent clauses.