Unsupervised Multi-hop Question Answering by Question Generation

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5866–5880 June 6–11, 2021. ©2021 Association for Computational Linguistics 5866 Unsupervised Multi-hop Question Answering by Question Generation Liangming Pan1 Wenhu Chen2 Wenhan Xiong2 Min-Yen Kan1 William Yang Wang2 1School of Computing, National University of Singapore, Singapore 2University of California, Santa Barbara, CA, USA e0272310@u.nus.edu {wenhuchen, xwhan, william}@cs.ucsb.edu kanmy@comp.nus.edu.sg Abstract Obtaining training data for multi-hop question answering (QA) is time-consuming and resource-intensive. We explore the possibility to train a well-performed multi-hop QA model without referencing any human-labeled multi- hop question-answer pairs, i.e., unsupervised multi-hop QA. We propose MQA-QG, an unsupervised framework that can generate human-like multi-hop training data from both homogeneous and heterogeneous data sources. MQA-QG generates questions by first selecting/generating relevant information from each data source and then integrating the multiple information to form a multi-hop question. Using only generated training data, we can train a competent multi-hop QA which achieves 61% and 83% of the supervised learning performance for the HybridQA and the HotpotQA dataset, respectively. We also show that pretraining the QA system with the generated data would greatly reduce the demand for human-annotated training data. Our codes are publicly available at https: //github.com/teacherpeterpan/ Unsupervised-Multi-hop-QA. 1 Introduction Extractive Question Answering (EQA) is the task of answering questions by selecting a span from the given context document. Works on EQA can be divided into the single-hop (Rajpurkar et al., 2016, 2018; Kwiatkowski et al., 2019) and multi- hop cases (Yang et al., 2018; Welbl et al., 2018; Perez et al., 2020). Unlike single-hop QA, which assumes the question can be answered with a sin- gle sentence or document, multi-hop QA requires combining disjoint pieces of evidence to answer a question. Though different well-designed neu- ral models (Qiu et al., 2019; Fang et al., 2020) have achieved near-human performance on the multi-hop QA datasets (Welbl et al., 2018; Yang et al., 2018), these approaches rely heavily on the When was the driver ranked in the 4-th position in 2004 United States Grand Prix born? Jenson Button is the driver ranked in the 4-th position in 2004 United States Grand Prix Jenson Button: Jenson Alexander Lyons Button MBE (born 19 January 1980) is a British racing driver and former Formula One driver. Question Generation Operator When was Jenson Button born? Bridge Entity: Jenson Button Table-to-Text Operator Bridge Blending Operator 2004 United States Grand Prix Pos Driver Constructor Time 1 Rubens Barrichello Ferrari 1:10.223 2 Michael Schumacher Ferrari 1:10.400 3 Takuma Sato BAR - Honda 1:10.601 4 Jenson Button BAR - Honda 1:10.820 Figure 1: An overview of our approach for generat- ing bridge-type multi-hop questions from table and text. The full set of supported input types and question types are described in Section 3.2. . availability of large-scale human annotation. Com- pared with single-hop QA datasets (Rajpurkar et al., 2016), annotating multi-hop QA datasets is signifi- cantly more costly and time-consuming because a human worker needs to read multiple data sources …

Content truncated to 500 words