Unsupervised Paraphrasing under Syntax Knowledge

The soundness of syntax is an important issue for the paraphrase generation task. Most methods control paraphrases by embedding and semantics in process, which cannot guarantee syntactical correctness results. Different from them, this paper we investigate structural patterns word usages termed as composable knowledge integrate it into to explicit way. This pretrained on a large corpus with dependency relationships formed probabilistic functions word-level soundness. For sentence-level correctness, design hierarchical structure loss quantitatively verify against given template. Thus, process can select appropriate words consideration both syntax. proposed method evaluated few datasets. experimental results show that quality our outperforms compared methods, especially terms correctness.