Language as a latent sequence: 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 unlabelled data is modelled as a sequence. We present novel unsupervised model named variational sequence auto-encoding reconstruction (VSAR), which performs inference given an observed text. To leverage information from text pairs, we additionally introduce supervised call dual directional learning (DDL), designed to integrate with our proposed VSAR model. Combining DDL (DDL+VSAR) enables us conduct learning. Still, combined suffers cold-start problem. further combat this issue, propose improved weight initialisation solution, leading two-stage training scheme knowledge-reinforced-learning (KRL). Our empirical evaluations suggest that yields competitive performance against state-of-the-art baselines on complete data. Furthermore, in scenarios only fraction of labelled pairs are available, consistently outperforms strong baseline (DDL) by significant margin (p<.05; Wilcoxon test). code publicly available at https://github.com/jialin-yu/latent-sequence-paraphrase.