Extracting and filtering paraphrases by bridging natural language inference and paraphrasing

Paraphrasing is a useful natural language processing task that can contribute\nto more diverse generated or translated texts. Natural language inference (NLI)\nand paraphrasing share some similarities and can benefit from a joint approach.\nWe propose a novel methodology for the extraction of paraphrasing datasets from\nNLI datasets and cleaning existing paraphrasing datasets. Our approach is based\non bidirectional entailment; namely, if two sentences can be mutually entailed,\nthey are paraphrases. We evaluate our approach using several large pretrained\ntransformer language models in the monolingual and cross-lingual setting. The\nresults show high quality of extracted paraphrasing datasets and surprisingly\nhigh noise levels in two existing paraphrasing datasets.\n