Towards Recognition of Textual Entailment in the Biomedical Domain

The medical literature suffers from disagreements among authors discussing the same topic or treatment. With thousands of articles published daily, there is a need to detect inconsistent and often contradictory findings. Natural language inference (NLI) gained a lot of interest in the past years, however, domain-specific NLI systems are yet to be examined in depth. In this paper, we conduct several experiments on sentence pairs extracted from PubMed abstracts, to infer whether they express entailment, contradiction or neutral meanings. The main focus of this research is to recognize textual entailment in published evidence-based medicine findings. We explore popular NLI models and sentence embeddings, adapted to the biomedical domain. We further investigate improving the inference detection abilities of the models by incorporating traditional machine learning (ML) features with deep learning (DL) architecture. The proposed model serves in capturing biomedical language’s representations by combining lexical, contextual and compositional semantics.