A generalized knowledge hunting framework for the winograd schema challenge

Abstract

The Winograd Schema Challenge is a popular alternative Turing test, comprising a binary-choice coreference-resolution task that requires significant common-sense and world knowledge to solve. In this paper, we propose a novel framework that successfully resolves many Winograd questions while imposing minimal restrictions on their form and difficulty. Our method works by (i) generating queries from a parsed representation of a Winograd question, (ii) acquiring relevant knowledge using Information Retrieval, and (iii) reasoning on the gathered knowledge. Our approach improves the F1 performance by 0.16 over previous works, without task-specific supervised training.