Computing relative polarity for textual inference

Semantic relations between main and complement sentences are of great significance in any system of automatic data processing that depends on natural language. In this paper we present a strategy for detecting author commitment to the truth/falsity of complement clauses based on their syntactic type and on the meaning of their embedding predicate. We show that the implications of a predicate at an arbitrary depth of embedding about its complement clause depend on a globally determined notion of relative polarity. We, moreover, observe that different classes of complement-taking verbs have a different effect on the polarity of their complement clauses and that this effect depends recursively on their own embedding. A polarity propagation algorithm is presented as part of a general strategy of canonicalization of linguistically-based representations, with a view to minimizing the demands on the entailment and contradiction detection process.