Submodular Subset Selection for Long-Document Question Answering

Submodular Optimization is a useful tool in Optimization, as it has properties of both convex and concave functions, and is efficiently optimizable. When answering questions on long documents, we want to (1) select spans of text relevant to the quesiton, but also (2) retain co-references in the text as well as the overall narrative, so that the supporting information is sufficient. This project involved building optimization objectives that captured these desiderata that were submodular. Find more details here.