We have created domain-adaptable rankers fine-tuned using knowledge distillation in order to re-rank the passages retrieved using BM25. We propose a novel difficulty prediction heuristic which dynamically determines the number of paragraphs to be fed to the reader by utilising the ranker scores and the remaining time. Finally, we use signals from reader, ranker as well as the retriever to determine the answerability of the question.
IIT-Patna-Inter-IIT-Tech-Meet/Domain-Specific-Question-Answering-DevRev
Domain adaptation in open domain question answering is tackled through theme specific rankers. We also propose a novel resource allocation algorithm to select the number of paragraph to be examined for extracting the answering. Finished 1st among participating IITs in Inter IIT tech Meet 11.0
Jupyter Notebook