Paraphrase Generation: A Survey of the State of the Art

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5075–5086 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 5075 Paraphrase Generation: A Survey of the State of the Art Jianing Zhou and Suma Bhat University of Illinois at Urbana-Champaign {zjn1746, spbhat2}@illinois.edu Abstract This paper focuses on paraphrase generation, which is a widely studied natural language gen- eration task in NLP. With the development of neural models, paraphrase generation research has exhibited a gradual shift to neural meth- ods in the recent years. This has provided architectures for contextualized representation of an input text and generating fluent, diverse and human-like paraphrases. This paper sur- veys various approaches to paraphrase genera- tion with a main focus on neural methods. 1 Introduction Paraphrases are texts that convey the same meaning while using different words or sentence structures. The generation of paraphrases is a longstanding problem for natural language learning. For exam- ple, the question How do I improve my English could be equivalently phrased as What is the best way to learn English. Paraphrasing can be play an important role in language understanding tasks, such as question answering (Dong et al., 2017; Zhu et al., 2017), machine translation (Seraj et al., 2015; Thompson and Post, 2020a), and semantic pars- ing (Berant and Liang, 2014; Cao et al., 2020). And it is also a good way for data augmentation (Kumar et al., 2019; Gao et al., 2020). Given a sentence, paraphrase generation aims to create its paraphrases that can have a different wording or different structure from the original sentence, while preserving the original meaning. The focus of paraphrase generation has exhibited a gradual shift from classical approaches to more advanced neural approaches in the recent years with the rapid development of various neural mod- els. Neural models have changed the traditional way paraphrase generation is performed and also provided new directions and architectures for the NLP community. While several surveys on the traditional methods and limited neural methods for paraphrase gener- Sentences Paraphrases How do I improve my En- glish What is the best way to learn English How far is Earth from Sun What is the distance be- tween Sun and Earth if at any time in the prepa- ration of this product the in- tegrity of this container is compromised it should not be used . this container should not be used if the product is compromised at any time in preparation . Table 1: Examples of paraphrases from available datasets for paraphrase generation. ation have been published (Metzler et al., 2011; Gupta and Krzy˙zak, 2020), there is no thorough and comprehensive survey on neural methods for paraphrase generation. To our best knowledge, this is the first survey on neural methods for paraphrase generation. Therefore, our goal in this paper is to provide a timely survey on paraphrase generation, with a main focus on neural methods. In the following section, we will first introduce the most frequently used datasets for paraphrase generation (Section 2). Then we list …

Content truncated to 500 words