Latent Reasoning for Low-Resource Question Generation

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3008–3022 August 1–6, 2021. ©2021 Association for Computational Linguistics 3008 Latent Reasoning for Low-Resource Question Generation Xinting Huang,1 Jianzhong Qi,1 Yu Sun,2 Rui Zhang3∗ 1The University of Melbourne, 2Twitter Inc., 3www.ruizhang.info {xintingh@student., jianzhong.qi@}unimelb.edu.au, ysun@twitter.com, rui.zhang@ieee.org Abstract Multi-hop question generation requires com- plex reasoning and coherent language realiza- tion. Learning a generation model for the problem requires extensive multi-hop question answering (QA) data, which are limited due to the manual collection effort. A two-phase strategy addresses the insufficiency of multi- hop QA data by first generating and then composing single-hop sub-questions. Learn- ing this generating and then composing two- phase model, however, requires manually la- beled question decomposition data, which is labor intensive. To overcome this limitation, we propose a novel generative approach that optimizes the two-phase model without ques- tion decomposition data. We treat the unob- served sub-questions as latent variables and propose an objective that estimates the true sub-questions via variational inference. We further generalize the generative modeling to single-hop QA data. We hypothesize that each single-hop question is a sub-question of an unobserved multi-hop question, and propose an objective that generates single-hop ques- tions by decomposing latent multi-hop ques- tions. We show that the two objectives can be unified and both optimize the two-phase gen- eration model. Experiments show that the pro- posed approach outperforms competitive base- lines on HOTPOTQA, a benchmark multi-hop question answering dataset. 1 Introduction Question generation aims to automatically gener- ate valid and coherent questions based on given context, which is widely applied to enrich ques- tion answering (QA) datasets, facilitate text com- prehension (Ko et al., 2020), seek clarification in conversation (Rao and Daum´e III, 2019), etc. Re- cently, neural encoder-decoder based approaches ∗Rui Zhang is the corresponding author. Table 1: Multi-Hop Question Reasoning Example Supportive Evidence Paragraph A. Dario Franchitti [1] George ..., known professionally as Dario Franchitti, is a retired Scottish racing driver. [2] After Franchitti did not secure a single-seater drive in 1995, he was contracted by the AMG team to compete in touring cars in the DTM and its successor — the International Touring Car Championship. Paragraph B. Mercedes-AMG [1] Mercedes-AMG GmbH (AMG)... is the high performance division of Mercedes-Benz. [2] Mercedes-AMG is headquartered in Affalterbach, Baden W¨urttemberg, Germany. Reasoning Progress < Dario Franchitti , contracted by , AMG > < Dario Franchitti , competed in , DTM > < AMG , headquartered in , Affalterbach, Baden W¨urttemberg, Germany > Multi-hop Question After he was contracted by the team that is headquartered in Affalterbach, Baden W¨urttemberg, Germany, Dario Franchitti competed in what series? Sub- Questions Which team is headquartered in Affalterbach, Baden W¨urttemberg, Germany? After contracted by AMG, Dario Franchitti competed in what series? Answer DTM have shown promising results for simple, single- hop question generation (Du and Cardie, 2018). Such approaches directly maps context (e.g., text passages) to questions without reasoning, and thus struggle when generating multi-hop questions (Pan et al., 2020). Here, reasoning refers to identify- ing …

Content truncated to 500 words