Few-Shot Complex Knowledge Base Question Answering via Meta Reinforcement Learning

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 5827–5837, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 5827 Few-Shot Complex Knowledge Base Question Answering via Meta Reinforcement Learning Yuncheng Hua†,§, Yuan-Fang Li♦, Gholamreza Haffari♦, Guilin Qi†,‡∗and Tongtong Wu† †School of Computer Science and Engineering, Southeast University, China ♦Faculty of Information Technology, Monash University, Australia §Southeast University-Monash University Joint Research Institute, China ‡Key Laboratory of Computer Network and Information Integration, Southeast University †{devinhua, gqi, wutong8023}@seu.edu.cn ♦{yuanfang.li, gholamreza.haffari}@monash.edu Abstract Complex question-answering (CQA) involves answering complex natural-language ques- tions on a knowledge base (KB). However, the conventional neural program induction (NPI) approach exhibits uneven performance when the questions have different types, harboring inherently different characteristics, e.g., dif- ficulty level. This paper proposes a meta- reinforcement learning approach to program induction in CQA to tackle the potential distri- butional bias in questions. Our method quickly and effectively adapts the meta-learned pro- grammer to new questions based on the most similar questions retrieved from the training data. The meta-learned policy is then used to learn a good programming policy, utilizing the trial trajectories and their rewards for simi- lar questions in the support set. Our method achieves state-of-the-art performance on the CQA dataset (Saha et al., 2018) while using only five trial trajectories for the top-5 re- trieved questions in each support set, and meta- training on tasks constructed from only 1% of the training set. We have released our code at https://github.com/DevinJake/MRL-CQA. 1 Introduction Knowledge-base question-answering (KBQA) in- terrogates a knowledge-base (KB) (Yin et al., 2016; Yu et al., 2017; Jin et al., 2019) by interpreting natural-language questions as logical forms (anno- tations), which can be directly executed on the KB to yield answers (denotations) (Pasupat and Liang, 2016). KBQA includes simple questions that re- trieve answers from single-hop triples (“what is Donald Trump’s nationality”) (Berant et al., 2013; Yih et al., 2014), multi-hop questions that infer an- swers over triple chains of at least 2 hops under specific constraints (“who is the president of the European Union 2012”) (Yih et al., 2016; Liang ∗Corresponding Author. et al., 2017), and complex questions that involve set operations (“how many rivers flow through India and China”) (Saha et al., 2019). In particular, com- plex question answering (CQA) (Saha et al., 2018) is a sophisticated KBQA task in which a sequence of discrete actions—e.g., set intersection and union, counting, comparison—needs to be executed, and is the subject of this paper. Consider the complex question “How many rivers flow through India and China?”. We first form a set of entities whose type is river and flow in China from the KB. We then form another set for rivers that flow through India. The answer is then gen- erated by counting the entities in the intersection of the two sets. More concretely, the question is trans- formed into the action sequence “Select (China, flow, river), Intersection (India, flow, river), Count”, which is executed on the KB to yield the answer. As such, the CQA …

Content truncated to 500 words