proxectonos/parafrases_gl

Paraphrases GL

Dataset description

Paraphrases GL is a dataset for the evaluation of paraphrase resources in Galician.

Because paraphrase is a complex notion and lacks a fully precise and universally accepted definition, this dataset adopts a three-way annotation scheme designed to capture both clear paraphrases and borderline cases.

Each pair of sentences is labeled with one of the following values:

Label	Category	Definition
0	NON-PARAPHRASE	The two sentences do not clearly and strictly share the same meaning.
1	BORDERLINE PARAPHRASE	The two sentences share a very similar, nearly paraphrastic meaning, and it is difficult to classify them in the other categories.
2	PARAPHRASE	The two sentences clearly and strictly share the same meaning.

Data creation methodology

Text selection

The source texts were selected from materials whose licenses allowed publication. The dataset was designed to include varied text types, with a preference for spoken or conversational language where possible.

The sources include:

• Wikipedia, providing varied encyclopedic texts
• Parliament, using transcripts of political sessions in the Parliament of Galicia
• CRTVG, using transcripts of programs from the Galician public broadcaster
• Hugin e Munin novels, using dialogues from novels made available by Hugin e Munin

Creation of paraphrase pairs

Two types of paraphrase candidates were generated:

• syntactic or full paraphrases, created through backtranslation using a pivot language such as Czech, English, or Spanish
• lexical paraphrases, created through automatic term substitution with BERT

Two annotators assigned labels 0, 1, or 2 to each sentence pair according to the non-paraphrase / borderline paraphrase / paraphrase scheme.

Dataset composition

• Number of syntactic paraphrases created through backtranslation: 1,619
• Number of lexical paraphrases: 1,316
• Total number of sentence pairs: 2,935

Data format

The dataset is distributed in TSV format and includes three splits:

• train
• validation
• test

Each row contains the following columns:

• Dataset: identifier of the paraphrase generation type or subset
• ID: identifier of the original source sentence or example
• Frase: original sentence in Galician
• Paráfrase: candidate paraphrase in Galician
• Avaliación: annotation label on the 0–2 scale

Example

Dataset	ID	Frase	Paráfrase	Avaliación
LEX_12	Parlamento_3572	E entre eses exemplos de boas prácticas atopamos o Plan estratéxico de igualdade de oportunidades entre mulleres e homes da Universidade de Santiago de Compostela, do 25 de marzo de 2009.	E entre eses exemplos de boas prácticas atopamos o Plan integral de igualdade de oportunidades entre mulleres e homes da Universidade de Santiago de Compostela, do 25 de marzo de 2009.	0
COMP_1301	Parlamento_5971	Pero, entón, ¿de que vén falar aquí, señor Rueda?	Mais, entón, de que vén falar aquí, señor Roda?	2

Intended uses

This dataset can be used for:

• evaluation of paraphrase detection systems in Galician
• semantic similarity classification
• research on borderline paraphrase phenomena
• low-resource NLP research for Galician
• development and evaluation of paraphrase generation or filtering methods

Limitations

• The notion of paraphrase is inherently fuzzy, which is why the dataset includes a borderline category.
• Some examples may be difficult even for human annotators to classify consistently.
• The dataset was created partly through automatic generation methods, so generated candidates may reflect artifacts of backtranslation or lexical substitution.
• The dataset is intended primarily for evaluation and analysis, rather than large-scale training.

License

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Users are free to share and adapt the material, provided that appropriate credit is given to the original source.

Usage

Example with datasets:

from datasets import load_dataset

ds = load_dataset(
    "csv",
    data_files={
        "train": "train.tsv",
        "validation": "validation.tsv",
        "test": "test.tsv",
    },
    delimiter="\t"
)

print(ds["train"][0])
print(ds["validation"][0])
print(ds["test"][0])

Acknowledgements

This dataset was compiled within the Nós Project, funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project ILENIA with reference 2022/TL22/00215336.