ParaAMR: A Large-Scale Syntactically Diverse Paraphrase Dataset by AMR Back-Translation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8047–8061 July 9-14, 2023 ©2023 Association for Computational Linguistics ParaAMR: A Large-Scale Syntactically Diverse Paraphrase Dataset by AMR Back-Translation Kuan-Hao Huang† Varun Iyer⊕ I-Hung Hsu⋄ Anoop Kumar‡ Kai-Wei Chang†‡ Aram Galstyan‡⋄ †University of California, Los Angeles ⊕University of Illinois Chicago ⋄Information Science Institute, University of Southern California ‡Amazon Alexa AI {khhuang, kwchang}@cs.ucla.edu, viyer9@uic.edu ihunghsu@isi.edu, {anooamzn, argalsty}@amazon.com Abstract Paraphrase generation is a long-standing task in natural language processing (NLP). Super- vised paraphrase generation models, which rely on human-annotated paraphrase pairs, are cost- inefficient and hard to scale up. On the other hand, automatically annotated paraphrase pairs (e.g., by machine back-translation), usually suf- fer from the lack of syntactic diversity — the generated paraphrase sentences are very similar to the source sentences in terms of syntax. In this work, we present PARAAMR, a large-scale syntactically diverse paraphrase dataset cre- ated by abstract meaning representation back- translation. Our quantitative analysis, qualita- tive examples, and human evaluation demon- strate that the paraphrases of PARAAMR are syntactically more diverse compared to existing large-scale paraphrase datasets while preserv- ing good semantic similarity. In addition, we show that PARAAMR can be used to improve on three NLP tasks: learning sentence embed- dings, syntactically controlled paraphrase gen- eration, and data augmentation for few-shot learning. Our results thus showcase the poten- tial of PARAAMR for improving various NLP applications. 1 Introduction Paraphrase generation is a long-standing task in nat- ural language processing (NLP) (McKeown, 1983; Barzilay and Lee, 2003; Kauchak and Barzilay, 2006). It has been applied to various downstream applications, such as question answering (Yu et al., 2018), chatbot engines (Yan et al., 2016), creative generation (Tian et al., 2021), and improving model robustness (Huang and Chang, 2021). Most exist- ing paraphrase generation models require a large amount of annotated paraphrase pairs (Li et al., 2019; Gupta et al., 2018; Kumar et al., 2020). Since human-labeled instances are expensive and hard to scale up (Dolan et al., 2004; Madnani et al., 2012; Iyer et al., 2017), recent research has explored the possibility of generating paraphrase pairs automat- ically. One popular approach is back-translation (Wieting and Gimpel, 2018; Hu et al., 2019a,b), which generates paraphrases of a source sentence by translating it to another language and translat- ing back to the original language. Although back- translation creates large-scale automatically anno- tated paraphrase pairs, the generated paraphrases usually suffer from the lack of syntactic diversity — they are very similar to the source sentences, es- pecially in syntactic features. Consequently, super- vised paraphrase models trained with those datasets are also limited in their ability to generate syntac- tically diverse paraphrases. Furthermore, not all words can be perfectly translated into another lan- guage. As we will show in Section 4.3, this mis- match may produce subpar paraphrases. In this work, we leverage abstract meaning rep- resentation (AMR) (Banarescu et al., 2013) to gen- erate syntactically diverse paraphrase pairs. We present PARAAMR, a large-scale syntactically …

Content truncated to 500 words