Intention-Aware Neural Networks for Question Paraphrase Identification

We tackle Question Paraphrasing Identification (QPI), a task of determining whether a pair of interrogative sentences (i.e., questions) are paraphrases of each other, which is widely applied in information retrieval and question answering. It is challenging to identify the distinctive instances which are similar in semantics though holding different intentions. In this paper, we propose an intention-aware neural model for QPI. Question words (e.g., " when") and blocks (e.g., " what time") are extracted as features for revealing intentions. They are utilized to regulate pairwise question encoding explicitly and implicitly, within Conditional Variational AutoEncoder (CVAE) and multi-task VAE frameworks, respectively. We conduct experiments on the benchmark corpora QQP, LCQMC and BQ, towards both English and Chinese QPI tasks. Experimental results show that our method yields generally significant improvements compared to a variety of PLM-based baselines (BERT, RoBERTa and ERNIE), and it outperforms the state-of-the-art QPI models. It is also proven that our method doesn't severely reduce the overall efficiency, which merely extends the training time by 12.5% on a RTX3090. All the models and source codes will be made publicly available to support reproducible research.