ParaSCI: A Large Scientific Paraphrase Dataset for Longer Paraphrase Generation

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 424–434 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 424 ParaSCI: A Large Scientific Paraphrase Dataset for Longer Paraphrase Generation Qingxiu Dong1,3,4, Xiaojun Wan1,2,3 and Yue Cao1,2,3 1Wangxuan Institute of Computer Technology, Peking University 2Center for Data Science, Peking University 3The MOE Key Laboratory of Computational Linguistics, Peking University 4College of Science, Minzu University of China qingxiudong@icloud.com,{wanxiaojun,yuecao}@pku.edu.cn Abstract We propose ParaSCI, the first large-scale para- phrase dataset in the scientific field, including 33,981 paraphrase pairs from ACL (ParaSCI- ACL) and 316,063 pairs from arXiv (ParaSCI- arXiv). Digging into characteristics and com- mon patterns of scientific papers, we construct this dataset though intra-paper and inter-paper methods, such as collecting citations to the same paper or aggregating definitions by sci- entific terms. To take advantage of sentences paraphrased partially, we put up PDBERT as a general paraphrase discovering method. The major advantages of paraphrases in ParaSCI lie in the prominent length and textual diver- sity, which is complementary to existing para- phrase datasets. ParaSCI obtains satisfactory results on human evaluation and downstream tasks, especially long paraphrase generation. 1 Introduction A paraphrase is a restatement of meaning with dif- ferent expressions (Bhagat and Hovy, 2013). Being very common in our daily language expressions, it can also be applied to multiple downstream tasks of natural language processing (NLP), such as gen- erating diverse text or adding richness to a chatbot. At present, paraphrase recognition or paraphrase generation are largely limited to the deficiency of paraphrase corpus. Especially, due to the perma- nent vacancy of paraphrase corpus in the scien- tific field, scientific paraphrase generation advances slowly. Scientific paraphrases can not only be help- ful for data augmentation of challenging scientific machine translation, but is also effective for polish- ing scientific papers. However, existing paraphrase datasets are mainly from news, novels, or social media platforms. Most of them remain short sen- tences and interrogative or oral style. As a result, none of such training data can train out a scien- tific paraphrase generator. Taking the sentence “we used pos tags predicted by the stanford pos tag- ger” as an example, the generated sentences from Transformer (Vaswani et al., 2017) models trained on existing paraphrase datasets1 are “level basic topics : what is the basic purpose of stanford tradi- tional hmo” and “a picture of a street sign with a sign on it”, far from ground-truth paraphrases. We have noticed that the structure of scientific papers is nearly fixed. Paraphrase sentence pairs appear not only within a paper (intra-paper) but also across different papers (inter-paper), which makes it possible to construct a paraphrase dataset in the scientific field. For example, repetitions of the same crucial contribution in a paper or explanations of the same term in different papers are potential paraphrases. Based on such characteristics, we design different methods to extract paraphrase pairs (shown in Section 4). In terms of the construction methods, existing methods merely focus on …

Content truncated to 500 words