(FAT) Fact Augmented Text

This was Vidhisha Balachandran's summer 2019 research intern project. It is no longer being used.

Answering Natural Questions with Background Knowledge

Natural Questions (NQ) is a newly-release question-answering (QA) dataset. We will focus on the problem of extracting short-answer questions from passages. Many of these answers can also be supported with external knowledge graphs (KGs), and we conjecture that state-of-the-art QA systems can be improved by aligning external KG triples with the passage text. The summer project is to (1) use the SLING entity-linking system and Wikidata repository, combined with RWR on the Wikidata KG, to find KG facts that are loosely-aligned with the text in the NQ passages, or potentially a larger corpora (2) fine-tune the BERT-based masked language model on the fact-aligned text and (3) further fine-tune that model as part of a short-answer QA model for NQ.