Idiomatic Expression Paraphrasing without Strong Supervision

Idiomatic expressions (IEs) play an essential role in natural language. In this paper, we study the task of idiomatic sentence paraphrasing (ISP), which aims to paraphrase a with IE by replacing its literal paraphrase. The lack large-scale corpora idiomatic-literal parallel sentences is primary challenge for task, consider two separate solutions. First, propose unsupervised approach ISP, leverages IE's contextual information and definition does not require training set. Second, weakly supervised using back-translation jointly perform generation IEs enlarge small-scale dataset. Other significant derivatives include model that replaces phrase generate expression large scale dataset idiomatic/literal pairs. effectiveness proposed solutions compared competitive baselines seen relative gains over 5.16 points BLEU, 8.75 METEOR, 19.57 SARI when generated are empirically validated on automatic manual evaluations. We demonstrate practical utility ISP as preprocessing step En-De machine translation.