Deep Latent Variable Models for Semi-supervised Paraphrase Generation

This paper explores deep latent variable models for semi-supervised paraphrase generation, where the missing target pair is modelled as a latent paraphrase sequence. We present a novel unsupervised model named variational sequence auto-encoding reconstruction (VSAR), which performs latent sequence inference given an observed text. To leverage information from text pairs, we introduce a supervised model named dual directional learning (DDL). Combining VSAR with DDL (DDL+VSAR) enables us to conduct semi-supervised learning; however, the combined model suffers from a cold-start problem. To combat this issue, we propose to deal with better weight initialisation, leading to a two-stage training scheme named knowledge reinforced training. Our empirical evaluations suggest that the combined model yields competitive performance against the state-of-the-art supervised baselines on complete data. Furthermore, in scenarios where only a fraction of the labelled pairs are available, our combined model consistently outperforms the strong supervised model baseline (DDL and Transformer) by a significant margin.