Paraphrase Generation as Zero-Shot Multilingual Translation: Disentangling Semantic Similarity from Lexical and Syntactic Diversity

Proceedings of the 5th Conference on Machine Translation (WMT), pages 561–570 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 561 Paraphrase Generation as Zero-Shot Multilingual Translation: Disentangling Semantic Similarity from Lexical and Syntactic Diversity Brian Thompson Johns Hopkins University brian.thompson@jhu.edu Matt Post Johns Hopkins University post@cs.jhu.edu Abstract Recent work has shown that a multilingual neural machine translation (NMT) model can be used to judge how well a sentence para- phrases another sentence in the same language (Thompson and Post, 2020); however, attempt- ing to generate paraphrases from such a model using standard beam search produces trivial copies or near copies. We introduce a sim- ple paraphrase generation algorithm which dis- courages the production of n-grams that are present in the input. Our approach enables paraphrase generation in many languages from a single multilingual NMT model. Further- more, the amount of lexical diversity between the input and output can be controlled at gen- eration time. We conduct a human evalua- tion to compare our method to a paraphraser trained on the large English synthetic para- phrase database ParaBank 2 (Hu et al., 2019c) and find that our method produces paraphrases that better preserve meaning and are more gra- matical, for the same level of lexical diver- sity. Additional smaller human assessments demonstrate our approach also works in two non-English languages. 1 Introduction Paraphrase generation is the task of producing a fluent output sentence which is semantically simi- lar to the input sentence while being syntactically and/or lexically different from it (Bhagat and Hovy, 2013). Paraphrasing has been of longstanding in- terest in the NLP community (McKeown, 1983) and has been used for data augmentation in ques- tion answering (Dong et al., 2017; Gan and Ng, 2019), machine translation (MT) (Hu et al., 2019a; Khayrallah et al., 2020), task oriented dialog (Niu and Bansal, 2018, 2019), and MT metrics (Baner- jee and Lavie, 2005; Zhou et al., 2006; Denkowski and Lavie, 2010; Thompson and Post, 2020). Thompson and Post (2020) recently released the Prism MT metric, which uses a multilingual neural MT (NMT) model as a paraphraser to score para- phrastic pairs; they treat paraphrasing as a zero-shot translation task (e.g., “translation” from English to English) and force-decode and score MT sys- tem outputs conditioned on their respective human translations. They denote their paraphraser as lex- ically/syntactically unbiased as it does not prefer output that differs lexically or syntactically from the input; this is advantageous for an MT metric as it assigns the highest score to an MT output which matches or nearly matches a human reference, but generating from the Prism model using standard beam search produces trivial copies or near copies. We introduce a simple method to enable para- phrase generation from a multilingual NMT model.1 Our method discourages the model from producing n-grams that match n-grams in the input sentence. This serves to lexically bias the output away from the input sentence, resulting in non- trivial paraphrases. When considered together with Prism model of Thompson and Post (2020), …

Content truncated to 500 words