Reasoning over Logically Interacted Conditions for Question Answering

Anonymous

Abstract: Some questions have answers that are correct only if certain conditions apply. Conditions are used to distinguish answers as well as to provide additional information to support them. To answer questions with conditions, models need to first find eligible answers and conditions from context and then perform logical reasoning to check whether conditions have been satisfied. We propose TReasoner to model this challenging reasoning process. In addition to finding answers, TReasoner can also identify unsatisfied conditions that are required to support the answers, as some answers are constrained by multiple conditions but only one or a subset of the conditions are satisfied. TReasoner consists of an entailment module, a reasoning module, and a generation module (if answers are free-form text spans). TReasoner achieves state-of-the-art performance on two benchmark QA datasets, outperforming the previous state-of-the-art by 3-10 points.

Paper Type: long