Adaptive Information Seeking for Open-Domain Question Answering

• DOI: 10.48550/arxiv.2109.06747
• Publication Date: 2021-01-01
• Authors: Yunchang Zhu; Liang Pang; Yanyan Lan; Huawei Shen; Xueqi Cheng
• Keywords: Hyperlink

Information seeking is an essential step for open-domain question answering to efficiently gather evidence from a large corpus. Recently, iterative approaches have been proven be effective complex questions, by recursively retrieving new at each step. However, almost all existing use predefined strategies, either applying the same retrieval function multiple times or fixing order of different functions, which cannot fulfill diverse requirements various questions. In this paper, we propose novel adaptive information-seeking strategy answering, namely AISO. Specifically, whole and answer process modeled as partially observed Markov decision process, where three types operations (e.g., BM25, DPR, hyperlink) one operation are defined actions. According learned policy, AISO could adaptively select proper action seek missing step, based on collected reformulated query, directly output when set sufficient question. Experiments SQuAD Open HotpotQA fullwiki, serve single-hop multi-hop QA benchmarks, show that outperforms baseline methods with strategies in terms both evaluations.