Dataset Card for TaPaCo Corpus

Dataset Summary

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.

Supported Tasks and Leaderboards

Paraphrase detection and generation have become popular tasks in NLP and are increasingly integrated into a wide variety of common downstream tasks such as machine translation , information retrieval, question answering, and semantic parsing. Most of the existing datasets cover only a single language – in most cases English – or a small number of languages. Furthermore, some paraphrase datasets focus on lexical and phrasal rather than sentential paraphrases, while others are created (semi -)automatically using machine translation.

The number of sentences per language ranges from 200 to 250 000, which makes the dataset more suitable for fine-tuning and evaluation purposes than for training. It is 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.


The dataset contains paraphrases in Afrikaans, Arabic, Azerbaijani, Belarusian, Berber languages, Bulgarian, Bengali , Breton, Catalan; Valencian, Chavacano, Mandarin, Czech, Danish, German, Greek, Modern (1453-), English, Esperanto , Spanish; Castilian, Estonian, Basque, Finnish, French, Galician, Gronings, Hebrew, Hindi, Croatian, Hungarian , Armenian, Interlingua (International Auxiliary Language Association), Indonesian, Interlingue; Occidental, Ido , Icelandic, Italian, Japanese, Lojban, Kabyle, Korean, Cornish, Latin, Lingua Franca Nova\t, Lithuanian, Macedonian , Marathi, Bokmål, Norwegian; Norwegian Bokmål, Low German; Low Saxon; German, Low; Saxon, Low, Dutch; Flemish, ]Old Russian, Turkish, Ottoman (1500-1928), Iranian Persian, Polish, Portuguese, Rundi, Romanian; Moldavian; Moldovan, Russian, Slovenian, Serbian, Swedish, Turkmen, Tagalog, Klingon; tlhIngan-Hol, Toki Pona, Turkish, Tatar, Uighur; Uyghur, Ukrainian, Urdu, Vietnamese, Volapük, Waray, Wu Chinese and Yue Chinese

Dataset Structure

Data Instances

Each data instance corresponds to a paraphrase, e.g.:

    'paraphrase_set_id': '1483',  
    'sentence_id': '5778896',
    'paraphrase': 'Ɣremt adlis-a.', 
    'lists': ['7546'], 
    'tags': [''],
    'language': 'ber'

Data Fields

Each dialogue instance has the following fields:

  • paraphrase_set_id: a running number that groups together all sentences that are considered paraphrases of each other
  • sentence_id: OPUS sentence id
  • paraphrase: Sentential paraphrase in a given language for a given paraphrase_set_id
  • lists: Contributors can add sentences to list in order to specify the original source of the data
  • tags: Indicates morphological or phonological properties of the sentence when available
  • language: Language identifier, one of the 73 languages that belong to this dataset.

Data Splits

The dataset is having a single train split, contains a total of 1.9 million sentences, with 200 – 250 000 sentences per language

Dataset Creation

Curation Rationale

[More Information Needed]

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

[More Information Needed]


Annotation process

[More Information Needed]

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

Creative Commons Attribution 2.0 Generic

Citation Information

  author       = {Scherrer, Yves},
  title        = {{TaPaCo: A Corpus of Sentential Paraphrases for 73 Languages}},
  month        = mar,
  year         = 2020,
  publisher    = {Zenodo},
  version      = {1.0},
  doi          = {10.5281/zenodo.3707949},
  url          = {}


Thanks to @pacman100 for adding this dataset.

Models trained or fine-tuned on tapaco