Interactive Language Learning by Question Answering

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2796–2813, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 2796 Interactive Language Learning by Question Answering Xingdi Yuan♥∗ Marc-Alexandre Cˆot´e♥∗ Jie Fu♣♠ Zhouhan Lin♦♠ Christopher Pal♣♠ Yoshua Bengio♦♠ Adam Trischler♥ ♥Microsoft Research, Montr´eal ♣Polytechnique Montr´eal ♦Universit´e de Montr´eal ♠ Mila eric.yuan@microsoft.com macote@microsoft.com Abstract Humans observe and interact with the world to acquire knowledge. However, most existing machine reading comprehension (MRC) tasks miss the interactive, information-seeking com- ponent of comprehension. Such tasks present models with static documents that contain all necessary information, usually concentrated in a single short substring. Thus, models can achieve strong performance through simple word- and phrase-based pattern matching. We address this problem by formulating a novel text-based question answering task: Ques- tion Answering with Interactive Text (QAit)1. In QAit, an agent must interact with a par- tially observable text-based environment to gather information required to answer ques- tions. QAit poses questions about the exis- tence, location, and attributes of objects found in the environment. The data is built using a text-based game generator that defines the un- derlying dynamics of interaction with the en- vironment. We propose and evaluate a set of baseline models for the QAit task that includes deep reinforcement learning agents. Experi- ments show that the task presents a major chal- lenge for machine reading systems, while hu- mans solve it with relative ease. 1 Introduction The research community has defined the task of machine reading comprehension (MRC) to teach machines to read and understand text. In most MRC tasks, given a knowledge source (usually a text document) and a question on its content, a model is required to answer the question either by pointing to words in the source or by generating a text string. Recent years have seen a flourishing of MRC works, including the release of numerous ∗ Equal contribution. 1The dataset and implementation of our baseline agents are publicly available at https://github.com/ xingdi-eric-yuan/qait_public. Game You find yourself in a backyard. You make out a patio table. You see a patio chair. It’s stylish. In a corner you can see a BBQ. There’s an open screen door leading south. Quest Is ghargh edible? Agent > Qapla’ Game That’s not a verb I recognize. Agent > eat ghargh Game You can’t see any such thing. Agent > go south Game You’ve just entered a kitchen. You can see an open fridge here. It is empty! You see a closed oven. Look over there, it’s a counter. On the counter you can see a diced potato, a red ghargh and a knife. A an open door is leading north. Agent > eat ghargh Game You need to take the red ghargh first. Agent > take red ghargh Game You take the red ghargh from the counter. Agent > eat ghargh Game You eat the red ghargh. Not bad. Agent > wait Agent > answer: yes Table 1: …

Content truncated to 500 words