TaPaCo: A Corpus of Sentential Paraphrases for 73 Languages

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6868–6873 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6868 TaPaCo: A Corpus of Sentential Paraphrases for 73 Languages Yves Scherrer University of Helsinki Department of Digital Humanities, Faculty of Arts yves.scherrer@helsinki.fi Abstract This paper presents TaPaCo, a freely available paraphrase corpus for 73 languages extracted from the Tatoeba database. Tatoeba is a crowdsourcing project mainly geared towards language learners. Its aim is to provide example sentences and translations for particular linguistic constructions and words. The paraphrase corpus is created by populating a graph with Tatoeba sentences and equivalence links between sentences “meaning the same thing”. This graph is then traversed to extract sets of paraphrases. Several language-independent filters and pruning steps are applied to remove uninteresting sentences. A manual evaluation performed on three languages shows that between half and three quarters of inferred paraphrases are correct and that most remaining ones are either correct but trivial, or near-paraphrases that neutralize a morphological distinction. The corpus contains a total of 1.9 million sentences, with 200 – 250 000 sentences per language. It covers a range of languages for which, to our knowledge, no other paraphrase dataset exists. The dataset is available at https://doi.org/10.5281/zenodo.3707949. Keywords: Multilingual corpus, Paraphrases, Crowdsourcing 1. Introduction Paraphrases are different textual realizations of the same meaning within a single language (Bhagat and Hovy, 2013; Ganitkevitch and Callison-Burch, 2014). Paraphrase de- tection and generation have become popular tasks in NLP and are increasingly integrated into a wide variety of com- mon downstream tasks such as machine translation, infor- mation retrieval, question answering, and semantic pars- ing (Federmann et al., 2019). Although the availability of large datasets for training and evaluation has facilitated research on paraphrase detection and generation, most of these datasets cover only a single language – in most cases English – or a small number of languages (see Section 4.). Furthermore, some paraphrase datasets focus on lexical and phrasal rather than sentential paraphrases, while others are created (semi-)automatically using machine translation. This paper describes the creation of a paraphrase corpus for 97 languages with a total of 1.9 million sentences. It consists of entire sentences produced by crowdsourcing within the Tatoeba project (Section 2.). The paraphrase matching process is entirely automatic and is based on the multilingual pivoting approach introduced by Bannard and Callison-Burch (2005). The number of sentences per lan- guage ranges from 200 to 250 000, which makes the dataset more suitable for fine-tuning and evaluation purposes than for training. It covers languages for which, to our knowl- edge, no other paraphrase dataset exists. In contrast to some previous work, we organize paraphrases as sets rather than pairs: all sentences in a paraphrase set are considered paraphrases of each other. This representa- tion is especially well-suited for multi-reference evaluation of paraphrase generation models, as there is generally not a single correct way of paraphrasing a given input sentence. This paper is structured as follows: Section 2. …

Content truncated to 500 words