NLP

Adaptive Batch Scheduling for Open-Domain Question Answering

최동현(카카오엔터프라이즈), 신명철(카카오엔터프라이즈), 김응균(카카오엔터프라이즈), 신동렬(성균관대학교)

IEEE Access

2021-08-11

Abstract

Open-domain question answering aims to get answers for given questions from a set of documents. Recently, dual encoder architecture is widely adopted to dense passage retrieval for question answering. In-batch negative sampling is typically used to gather extra negative samples during training. In this paper, we propose adaptive batch scheduling to enhance the performance of in-batch negative sampling. The proposed algorithm schedules training batches to increase the difficulty of the sampled negatives by in-batch negative sampling during training. We evaluated the proposed approach on the two well-known document retrieval benchmark datasets MSMARCO and Natural Questions. The evaluation result shows that the proposed adaptive batch scheduling could significantly improve the document retrieval performances of dual encoder architecture document retrieval systems.