mariadelcarmenramirez/metaphor-catalan

META4XNLI-cat

Dataset Summary

META4XNLI-cat is a token-level metaphor detection dataset for Catalan derived from the multilingual XNLI (Cross-lingual Natural Language Inference) corpus.

The dataset was created by projecting metaphor annotations from English and Spanish into Catalan, followed by manual validation and correction. The English and Spanish annotations come from the Meta4XNLI corpus, which provides manually labeled metaphors aligned across languages.

Using multilingual contextual embeddings, metaphor labels were automatically projected into the Catalan translations of the same sentences. The projected annotations were then manually reviewed and corrected to ensure annotation quality and linguistic consistency.

To our knowledge, META4XNLI-cat is the first token-level metaphor detection dataset specifically designed for Catalan.

---

Supported Tasks

This dataset is intended for:

• Token classification
• Metaphor detection
• Figurative language processing
• Cross-lingual transfer learning
• Multilingual NLP research

Primary task:

• Token classification (metaphor identification)

---

Languages

• Catalan (ca)

---

Dataset Structure

Data Instances

Each example corresponds to a tokenized sentence with token-level metaphor annotations.

Example:

{
  "id": 1,
  "tokens": ["La", "veu", "tremolava", "d'indignació"],
  "tags": ["O", "O", "B-METAPHOR", "O"]
}

Fields:

Field	Description
id	Sentence identifier
tokens	Tokenized sentence
tags	Token-level metaphor labels

---

Data Fields

• id (int64)
  Unique identifier for each sentence.

• tokens (list[string])
  Tokenized sentence using whitespace tokenization.

• tags (list[class_label])
  Token-level metaphor annotations.

Label set:

Label	Meaning
O	Literal token
B-METAPHOR	Beginning of a metaphorical expression

Although BIO schemes usually include I-METAPHOR, the final dataset contains only single-token metaphors, so I-METAPHOR does not appear in the labels.

---

Data Splits

Split	Sentences
Train	525
Validation	66
Test	66

Total sentences: 658

---

Dataset Creation

Source Data

The dataset combines two main resources:

Metaphor annotations

• https://huggingface.co/datasets/HiTZ/meta4xnli
• Sanchez-Bayona & Agerri (2025)

This dataset contains English and Spanish sentences annotated with metaphor labels.

Catalan sentence translations

• https://huggingface.co/datasets/projecte-aina/xnli-ca

Provided by the AINA project, which released Catalan translations of the XNLI dataset.

---

Label Projection

Metaphor annotations were projected from English and Spanish into Catalan using multilingual contextual embeddings.

The embeddings were obtained using the bert-base-multilingual-cased model introduced in:

Devlin et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018).

During projection:

• Tokens across languages were aligned using cosine similarity.
• Labels were transferred only when similarity exceeded 0.7.
• BIO constraints were enforced to maintain valid label sequences.

Due to the conservative alignment strategy, the resulting annotations mostly consist of single-token metaphor spans.

---

Manual Validation

All projected annotations were manually reviewed and corrected using a custom annotation interface.

The validation process included:

• correcting projection errors
• identifying missed metaphors caused by tokenization differences
• fixing translation issues in the Catalan sentences

Additionally, 7 sentences in the original Catalan XNLI subset were corrected, improving the overall linguistic quality of the dataset.

---

Dataset Statistics

Metric	Value
Total sentences	658
Literal tokens (O)	7,717
Metaphor tokens (B-METAPHOR)	403
Sentences containing metaphors	335 (50.99%)
Corrected XNLI-cat sentences	7

---

Limitations

Users should consider the following limitations:

• The dataset is relatively small (<1K sentences).
• Most metaphor annotations correspond to single-token spans due to the projection strategy.
• Sentences originate from NLI data, which may differ from natural conversational or literary text.

Despite these limitations, the dataset provides a high-density metaphor resource for Catalan.

---

Licensing Information

The dataset is derived from:

• Meta4XNLI
• XNLI Catalan translations

Users should consult the licensing terms of the original datasets before redistribution or commercial use.

---

Citation

If you use this dataset, please cite the following works.

META4XNLI

@misc{sanchezbayona2025meta4xnlicrosslingualparallelcorpus,
      title={Meta4XNLI: A Crosslingual Parallel Corpus for Metaphor Detection and Interpretation}, 
      author={Elisa Sanchez-Bayona and Rodrigo Agerri},
      year={2025},
      eprint={2404.07053},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2404.07053}, 
}

AINA XNLI Catalan

@inproceedings{gonzalez-agirre-etal-2024-building,
    title = "Building a Data Infrastructure for a Mid-Resource Language: The Case of {C}atalan",
    author = "Gonzalez-Agirre, Aitor  and
      Marimon, Montserrat  and
      Rodriguez-Penagos, Carlos  and
      Aula-Blasco, Javier  and
      Baucells, Irene  and
      Armentano-Oller, Carme  and
      Palomar-Giner, Jorge  and
      Kulebi, Baybars  and
      Villegas, Marta",
    editor = "Calzolari, Nicoletta  and
      Kan, Min-Yen  and
      Hoste, Veronique  and
      Lenci, Alessandro  and
      Sakti, Sakriani  and
      Xue, Nianwen",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.231/",
    pages = "2556--2566",
}

---

Acknowledgements

This dataset builds upon resources from:

• HiTZ Center for Language Technology
• The AINA Catalan NLP Project (Barcelona Supercomputing Center)

Their work on multilingual NLP resources made this dataset possible.