Contradiction Detection Approach Based on Semantic Relations and Evidence of Uncertainty

Abstract

The automatic detection of contradictions or the detection of contradictory statements consists in identifying the discrepancies, inconsistencies, conflicts and disagreement in a given text. This technique has several applications in the real world; as in question-and-answer systems, multi-document synthesis, and contradictions of opinions analysis in social networks. The automatic detection of contradictions is a technically difficult natural language processing problem, given the variety of ways in which contradictions occur between texts. Indeed, even if brutal negations, antonyms and numerical mismatches are obvious characteristics to diffuse contradictions, there can be a contradiction which also comes from an inconsistent domain knowledge, from uncertain coreferences or from differences in the structures of the assertions.

In this paper, we address the problem of detecting contradictions for uncertain statements when the user (author) not only provides factual information, but also clues about its plausibility. Therefore, we propose a contradiction detection approach which introduces additional criteria for detecting linguistic semantic features due to natural language ambiguity and evidence of uncertainty. The idea is to build a model to detect contradictions using a joint analysis of semantic relations and evaluations of uncertainties. In order to validate our proposed approach, we are carrying out experiments on four data sets. The results based on the experiments indicate the effectiveness of our approach.