PerPaDa: A Persian Paraphrase Dataset based on Implicit Crowdsourcing Data Collection

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5090–5096 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5090 PerPaDa: A Persian Paraphrase Dataset based on Implicit Crowdsourcing Data Collection Salar Mohtaj1,2, Fatemeh Tavakkoli3, Habibollah Asghari4 1 Technische Universit¨at Berlin, Berlin, Germany 2 German Research Centre for Artificial Intelligence (DFKI), Labor Berlin, Germany 3 Freie Universit¨at Berlin, Berlin, Germany 4 ICT Research Institute of ACECR, Tehran, Iran salar.mohtaj@tu-berlin.de, f.tavakkoli@fu-berlin.de, habib.asghari@ictrc.ac.ir Abstract In this paper we introduce PerPaDa, a Persian paraphrase dataset that is collected from users’ input in a plagiarism detection system. As an implicit crowdsourcing experience, we have gathered a large collection of original and paraphrased sentences from Hamtajoo; a Persian plagiarism detection system, in which users try to conceal cases of text re-use in their documents by paraphrasing and re-submitting manuscripts for analysis. The compiled dataset contains 2446 instances of paraphrasing. In order to improve the overall quality of the collected data, some heuristics have been used to exclude sentences that don’t meet the proposed criteria. The introduced corpus is much larger than the available datasets for the task of paraphrase identification in Persian. Moreover, there is less bias in the data compared to the similar datasets, since the users did not try some fixed predefined rules in order to generate similar texts to their original inputs. Keywords: Paraphrase, Persian, Implicit crowdsourcing, Plagiarism detection 1. Introduction Paraphrase identification is the task of investigating if a piece of text is a paraphrase of another one (Hunt et al., 2019). It is a basic task for different Natural Lan- guage Processing (NLP) tasks like Plagiarism Detec- tion (PD) and question answering. Paraphrase identifi- cation could improve the overall performance of these tasks by enabling systems to detect semantically sim- ilar or related sentences or phrases. For instance, a plagiarism detection system that empowered by a para- phrase identification module not only can detect cases of verbatim text re-use, but also the cases in which peo- ple try to conceal plagiarism by using different word- ings and the other paraphrasing techniques. There are different approaches for identification of paraphrases. A possible approach for measuring if two sentences are semantically similar (i.e., paraphrased) is to measure cosine similarity between two sen- tences (Mahmoud and Zrigui, 2017). However, more recent approaches are using Machine Learning (ML) models to train a classifier to detect if two pieces of text are paraphrased. Moreover, pre-trained language mod- els (e.g., BERT (Devlin et al., 2019)) have been using recently for the task of paraphrase identification (Wang et al., 2020). The neural networks could also be used to generate paraphrases from textual data (Wahle et al., 2021). A training dataset is an essential part of all super- vised learning approaches in ML to train a model. A paraphrase identification dataset is an important piece of puzzle to train an outstanding paraphrase detection model. In this paper we introduced PerPaDa, a new Persian Paraphrase Dataset in Persian. The data is …

Content truncated to 500 words