Cross-Lingual Training for Automatic Question Generation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4863–4872 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 4863 Cross-Lingual Training for Automatic Question Generation Vishwajeet Kumar1,2, Nitish Joshi2, Arijit Mukherjee2, Ganesh Ramakrishnan2, and Preethi Jyothi2 1IITB-Monash Research Academy, Mumbai, India 2IIT Bombay, Mumbai, India {vishwajeet, nitishj, ganesh, pjyothi}@cse.iitb.ac.in {arijitmukh007}@gmail.com Abstract Automatic question generation (QG) is a chal- lenging problem in natural language under- standing. QG systems are typically built as- suming access to a large number of training in- stances where each instance is a question and its corresponding answer. For a new language, such training instances are hard to obtain mak- ing the QG problem even more challenging. Using this as our motivation, we study the reuse of an available large QG dataset in a sec- ondary language (e.g. English) to learn a QG model for a primary language (e.g. Hindi) of interest. For the primary language, we assume access to a large amount of monolingual text but only a small QG dataset. We propose a cross-lingual QG model which uses the fol- lowing training regime: (i) Unsupervised pre- training of language models in both primary and secondary languages and (ii) joint super- vised training for QG in both languages. We demonstrate the efficacy of our proposed ap- proach using two different primary languages, Hindi and Chinese. We also create and release a new question answering dataset for Hindi consisting of 6555 sentences. 1 Introduction Automatic question generation from text is an im- portant yet challenging problem especially when there is limited training data (i.e., pairs of sen- tences and corresponding questions). Standard se- quence to sequence models for automatic question generation have been shown to perform reason- ably well for languages like English, for which hundreds of thousands of training instances are available. However, training sets of this size are not available for most languages. Manually cu- rating a dataset of comparable size for a new lan- guage will be tedious and expensive. Thus, it would be desirable to leverage existing question answering datasets to help build QG models for a Sentence : िव�ा के ये सभी �प हमारे रा�ट्रीय �ान के िविवध अंग ह� (All these forms of education are diverse aspects of our national knowledge system.) Question (ground truth) : िव�ा के सभी �प हमारे रा�ट्रीय �ान के �या ह� � (What is the relationship between different forms of education and our national knowledge systems?) Question (predicted) : िव�ा के सभी �प �या ह� � (What are all the forms of education?) Sentence : स�यता का अथ� है संपि� की िनरंतर वृि� , �यव�था और र�ा अपनी संपि� की र�ा औजारों के �ारा की जाती है (Civilization means continuous growth of prosperity, the system and its security are facilitated by the defense mechanism of the civilization.) Question (ground truth) : स�यता का �या अथ� है � (What is the meaning of civilization?) Question (predicted) : स�यता का �या अथ� है � (What is the meaning of …

Content truncated to 500 words