Unsupervised Contradiction Detection using Sentence Transformations

Abstract

Contradiction detection (CD) is a subfield of Natural Language Inference (NLI) that is relevant to many domains where contradictory statements in texts should be avoided (e.g., in financial or regulatory documents). With the advent of large annotated NLI datasets, there has been an increased focus on supervised deep-learning approaches in this research area. However, since this training data does not necessarily reflect the characteristic properties of the application data, unsupervised CD approaches are still relevant for certain domains or languages. In this paper, we therefore take up a recently published unsupervised NLI approach, reproduce parts of the proposed sentence transformations, extend it with various modifications, and evaluate it for the sole task of contradiction detection. The results show that under the exclusion of certain transformations types, an accuracy of 71.42 can be achieved on the SNLI test dataset.