Deep neural networks (DNNs) have become a critical tool in natural language processing (NLP) for a wide variety of language technologies, from syntax to semantics to pragmatics. In particular, in the field of natural language inference (NLI), DNNs have become the de-facto model, providing significantly better results than previous paradigms. Their power lies in their ability to embed complex language ambiguities in high dimensional spaces coupled with non-linear compositional transformations learned to directly optimize task-specific objective functions. We propose to adapt Deep NLI techniques to the biomedical domain, specifically investigating question answering, information extraction and synthesis. The biomedical domain presents many key challenges and a critical impact that standard NLI challenges do not posses. First, while standard NLI data sets requires a system to model basic world knowledge (e.g., that ‘soccer’ is a ‘sport’), they do not presume a rich domain knowledge encoded in various and often heterogeneous resources such as scientific articles, textbooks and structured databases. Second, while standard NLI data sets presume that the answer/inference is encoded in a single utterance, the ability to reason and extract information from biomedical domains often requires information synthesis from multiple utterances, paragraphs, and even documents. Finally, whereas standard NLI is a broad challenge aimed at testing whether computers can make general inferences in language, biomedical texts are a grounded and impactful domain where progress in automated reasoning will directly impact the efficacy of researchers, physicians, publishers and policy makers.
Fields of science
Call for proposal
See other projects for this call