Automated Contradiction Detection in Biomedical Literature

Medical literature suffers from inconsistencies between reported findings that answer the same research question. This paper introduces an automated two-phase contradiction detection model that integrates semantic properties as input features to a Learning-to-Rank framework, to accurately identify key findings of a research article. It also relies on negation, antonyms and similarity measures to detect contradictions between findings. The proposed technique is implemented and tested on a publicly available contradiction corpus 259 manually annotated abstracts. The performance is compared based on recall, precision and F-measure. Experimental evaluations prove the utility of the model and its contribution to the contradiction classification and extraction task.