Open Domain Question Answering with A Unified Knowledge Interface

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1605 - 1620 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Open Domain Question Answering with A Unified Knowledge Interface Kaixin Ma♣†∗, Hao Cheng♠∗, Xiaodong Liu♠, Eric Nyberg♣, Jianfeng Gao♠ ♣ Carnegie Mellon University ♠ Microsoft Research {kaixinm,ehn}@cs.cmu.edu {chehao,xiaodl,jfgao}@microsoft.com Abstract The retriever-reader framework is popular for open-domain question answering (ODQA) due to its ability to use explicit knowledge. Al- though prior work has sought to increase the knowledge coverage by incorporating struc- tured knowledge beyond text, accessing het- erogeneous knowledge sources through a uni- fied interface remains an open question. While data-to-text generation has the potential to serve as a universal interface for data and text, its feasibility for downstream tasks remains largely unknown. In this work, we bridge this gap and use the data-to-text method as a means for encoding structured knowledge for ODQA. Specifically, we propose a verbalizer-retriever- reader framework for ODQA over data and text where verbalized tables from Wikipedia and graphs from Wikidata are used as aug- mented knowledge sources. We show that our Unified Data and Text QA, UDT-QA, can effectively benefit from the expanded knowl- edge index, leading to large gains over text- only baselines. Notably, our approach sets the single-model state-of-the-art on Natural Ques- tions. Furthermore, our analyses indicate that verbalized knowledge is preferred for answer reasoning for both adapted and hot-swap set- tings. 1 Introduction Pretrained language models (Devlin et al., 2019; Brown et al., 2020) have been shown to store cer- tain knowledge (linguistic or factual) implicitly in parameters (Manning et al., 2020; Petroni et al., 2019; Roberts et al., 2020), partially explaining the superior generalization abilities over downstream tasks. However, besides the well-known hallucina- tion issue, the implicit knowledge learned through language modeling objective over text struggles at reflecting up-to-date knowledge from text and struc- tured data for answering open-domain questions. † Work done during an internship at Microsoft Research ∗ Equal contribution To overcome this, recent work on open domain question answering (ODQA) focuses on the semi- parametric method (Karpukhin et al., 2020; Guu et al., 2020) where the pretrained language models can leverage external explicit knowledge sources for reasoning. For example, in the retriever-reader framework (Min et al., 2021, inter alia), the reader produces answers by grounding on the relevant evidence from the retriever, the interface to the ex- plicit knowledge source (Wikipedia text passages). In this work, we focus on the semi-parametric ap- proach for ODQA going beyond textual knowledge. Specifically, we are interested in the question: Can we develop a viable unified interface over a real- istic heterogeneous knowledge source containing both data and text? Recent retriever-reader models (Oguz et al., 2020; Agarwal et al., 2021) have demonstrated that expanding the textual knowledge source with more structured data is beneficial. However, only knowl- edge base (KB) is considered in (Agarwal et al., 2021), limiting the applicability of their method to other structured data. In (Oguz et al., 2020), both …

Content truncated to 500 words