metadata

YAML tags: null
annotations_creators:
  - expert-generated
language_creators:
  - found
language:
  - ca
multilinguality:
  - monolingual
pretty_name: teca
size_categories:
  - unknown
source_datasets: []
task_categories:
  - text-classification
task_ids:
  - natural-language-inference

Dataset Card for XNLI-ca

Dataset Description

Website:
Paper:
Point of Contact:

Dataset Summary

XNLI-ca is a dataset of textual entailment in Catalan, which contains 21,163 pairs of premises and hypotheses, annotated according to the inference relation they have (implication, contradiction or neutral).

This dataset was developed by the Language Technologies Unit at BSC as part of Projecte AINA, to enrich the Catalan Language Understanding Benchmark (CLUB).

Supported Tasks and Leaderboards

Textual entailment, Text classification, Language Model

Languages

The dataset is in Catalan (ca-CA).

Dataset Structure

Data Instances

Three JSON files, one for each split.

Example:

    
   {
    "label": "contradiction",
    "premise": "Bé, ni tan sols estava pensant en això, però estava molt frustrat i vaig acabar tornant a parlar amb ell.",
    "hypothesis": "No he tornat a parlar amb ell."
  },
  {
    "label": "entailment",
    "premise": "Bé, ni tan sols estava pensant en això, però estava molt frustrat i vaig acabar tornant a parlar amb ell.",
    "hypothesis": "Estava tan molest que vaig començar a parlar amb ell de nou."
  },
  {
    "label": "neutral",
    "premise": "Bé, ni tan sols estava pensant en això, però estava molt frustrat i vaig acabar tornant a parlar amb ell.",
    "hypothesis": "Vam tenir una gran xerrada."
  }

Data Fields

premise: text
hypothesis: text related to the premise
label: relation between premise and hypothesis:
- 0: entailment
- 1: neutral
- 2: contradiction

Data Splits

dev.json: 2116 examples
test.json: 2117 examples
train.json: 16930 examples

Dataset Creation

Curation Rationale

We created this dataset to contribute to the development of language models in Catalan, a low-resource language.

Source Data

Source sentences are extracted from the Catalan Textual Corpus and from VilaWeb newswire.

Initial Data Collection and Normalization

12000 sentences from the BSC Catalan Textual Corpus, together with 6200 headers from the Catalan news site VilaWeb, were chosen randomly. We filtered them by different criteria, such as length and stand-alone intelligibility. For each selected text, we commissioned 3 hypotheses (one for each entailment category) to be written by a team of native annotators.

Some sentence pairs were excluded because of inconsistencies.

Who are the source language producers?

The Catalan Textual Corpus corpus consists of several corpora gathered from web crawling and public corpora. More information can be found here.

VilaWeb is a Catalan newswire.

Annotations

Annotation process

We commissioned 3 hypotheses (one for each entailment category) to be written by a team of annotators.

Who are the annotators?

Annotators are a team of native language collaborators from two independent companies.

Personal and Sensitive Information

No personal or sensitive information included.

Considerations for Using the Data

Social Impact of Dataset

We hope this dataset contributes to the development of language models in Catalan, a low-resource language.

Discussion of Biases

[N/A]

Other Known Limitations