HCPG: a highlighted contrastive learning framework for exemplar-guided paraphrase generation

Exemplar-guided Paraphrase Generation aims to use an exemplar sentence to guide the generation of a paraphrase that retains the semantic content of the source sentence, along with the syntax of the exemplar. Some methods use syntax structure extracted from exemplars to guide generation, but the preprocesses may cause information loss. The other methods directly use the natural exemplar sentences (NES) as syntactic guidance, which avoids the loss of information but fails to capture and integrate the exemplar's syntax and source sentence's semantics effectively. In this paper, we propose a Highlighted Contrastive learning framework for exemplar-guided Paraphrase Generation (HCPG), which solves the shortcomings of using NES as syntactic guidance. The "highlight" refers to a continuous process of supplementing and refining, which effectively captures both the semantic and syntactic information of the sentences. HCPG also includes a contrastive loss layer to help the decoder fully integrate the highlighted semantic and syntactic information to generate final paraphrases. Experiments on ParaNMT and QQP-Pos show that HCPG is comparable to several state-of-the-art models, including SAGP and GCPG, and achieves an average 3.19% improvement compared with CLPG.