Extract, Transform and Filling: A Pipeline Model for Question Paraphrasing based on Template

Yunfan Gu1, yang yuqiao2, Zhongyu Wei2
1Fudan University, 2School of Data Science, Fudan University


Question paraphrasing aims to restate a given question with different expressions but keep the original meaning. Recent approaches are mostly based on neural networks following a sequence-to-sequence fashion, however, these models tend to generate unpredictable results. To overcome this drawback, we propose a pipeline model based on templates. It follows three steps, a) identifies template from the input question, b) retrieves candidate templates, c) fills candidate templates with original topic words. Experiment results on two self-constructed datasets show that our model outperforms the seq2seq model in a large margin and can generate even better results when the size of training sample is small.