Automatic Generation of Multiple-choice Cloze-test Questions for Lao Language Learning

Xinying Qiu, Haiwei Xue, Lingfeng Liang, Zexin Xie, Shuxuan Liao, Guofeng Shi

Abstract: We propose a framework and a set of distractor generation strategy for multiple-choice cloze test question generation for low-resource language such as Lao. We hypothesize that a neural model trained with high predictive performance on MC cloze test could serve as question filters to provide the instructors with a model-verified pre-selected question set for easy question selection and adaptation. Therefore we evaluate a BERT baseline model to analyze the neural language model's predictive performance on MC cloze test. We observe that our distractor generation strategies can increase the problem difficulty for adjective and adverb question answering significantly more than for the others. The neural model could be a good question filter for close word class questions of conjunctions and preposition by achieving high accuracy and true positive rate. Its performance needs to be further improved for answering open word class questions such as adjectives and adverbs.