An approach to the Winograd Schema Challenge based on semantic classification of events and adjectives

The Winograd Schema Challenge is an interesting problem that can be considered as an alternative to the renowned Turing test for measuring the intelligence of artificial agents. Originally proposed by Hector Levesque in 2011, the challenge gained significant popularity among the researchers of Natural Language Understanding (NLU). It is a special class of coreference resolution problem and the challenge involves finding the correct referent for a given pronoun or possessive adjective in a short text. The Winograd Schema was proposed to encourage human-like reasoning and therefore the problems were carefully designed so that the statistical methods which solely depend on lexical correctness and word associativity without any logical reasoning will not be able to solve these problems reliably. The designers encouraged the use of commonsense based logical methods to solve the Winograd Schema. However, at present, in terms of accuracy, the most advanced logical methods are far from human-like accuracy and also significantly outperformed by the statistical methods. The major challenge for the logical methods is the difficulty of finding a general model for the Winograd Schema. The objective of this research is to introduce a commonsense based logical method that can achieve competitive accuracy with the statistical methods on a specific form of the Winograd Schema problems. Our approach is based on semantic classification of common verbs and adjectives, creating a commonsense knowledge base, and using the knowledge base to align the pronoun with the correct referent. We have evaluated the performance of our approach and compared the accuracy with the state of the art statistical approach on two benchmark problem sets: WSCL and WSCR.