As part of my deeplearning course, my team mate and I have worked on deep QA models attempting SQUAD dataset. We have tried summarizing the answer conditioned on the question into a hidden state initially using RNN variants. Realising these models inability to capture important part of the supporting paragraph, we have incorporated attention layer as suggested in BiDAF paper.

As part of a coursework, few of my amazing classmates and I collaborated with IBM Watson QA team on building a tool for analysing the reasoning and knowledge abilities required for an AI model to tackle ARC, a dataset of science domain based multiple choice questions. We analysed the importance of an effective IR system able to get the facts needed for answering a scientific question by changing an open-doman problem to a closed domain variant and testing the reading comprehension model’s performance; we showed that model performance increases drastically.

We were awarded the BestPaper award at ACL MRQA workshop’ 18.