Joint Passage Retrieval and Answer Prediction for Open-Domain Question Answering

Abstract:
Open-domain question answering (ODQA), which requires to answer questions asked in natural language with a vast document collection, has gained increasingly wide attention recently. As machine reading comprehension (MRC) has similarity with ODQA, successful achievement in MRC can be extended to ODQA. However, there are two challenges: (1) passage retrieval performance will affect answer prediction in open domain setting and (2) successive questions are common in ODQA and coreference phenomena need to be resolved. In order to tackle these problems, MRC-based models are extended to ODQA with training answer prediction and question-passage match jointly based on pre-trained BERT. Moreover, we split multiple questions and introduce heuristic coreference resolution methods. Extensive experiments have been conducted on a ODQA dataset Les MMRC 2.0. Results demonstrate that our proposed model outperforms other models in both passage retrieval recall and answer prediction Rouge-L.