Elamur Research⌕
☰
home / papers / 2026 / dialect2vec: a vector-based method for dialect transcription of Portuguese from ALiB questionnaires

dialect2vec: a vector-based method for dialect transcription of Portuguese from ALiB questionnaires

Original (Portuguese): dialect2vec: Um método baseado em vetores para transcrição dialetal do português a partir de questionários do ALiB

Paper · 2026 · Laila Mota, Daniela Barreiro Claro, Eloize R. Marques Seno et al.

Summary

Subword language models handle phonetic transcriptions poorly. IPA symbols missing from their vocabularies become [UNK] tokens or get split into many fragments, so pronunciation detail is lost. That is a problem for modelling dialect variation. The authors propose dialect2vec, a hybrid embedding for Brazilian Portuguese that concatenates three components: a lexical vector from BERTimbau, a phonetic vector from the token-free, byte-level ByT5 encoder applied to IPA transcriptions, and a one-hot vector for the informant's location. They test it on manually revised answers to the phonetic-phonological questionnaires of the Atlas Linguístico do Brasil (ALiB), using informants from Salvador, Três Rios and Curitiba. The evaluation isolates the phonetic component, removing semantic and geographic cues, and compares it with mBERT and BERTimbau baselines in unsupervised clustering. With KMeans, the byte-level phonetic vectors reach agreement close to BERTimbau. This suggests that pronunciation cues alone can recover dialect groupings. The authors describe it as the first attempt to encode dialect variation in vector representations for Portuguese.

Key points

  • The embedding concatenates three vectors: BERTimbau lexical (768 dims), ByT5 phonetic from mean-pooled encoder states (1472 dims), and one-hot geographic (N dims, one per locality).
  • The data comes from ALiB phonetic-phonological questionnaires, with four informants each from Salvador (BA), Três Rios (RJ) and Curitiba (PR).
  • Clustering used KMeans and HDBSCAN and was scored with precision, recall, ARI and Silhouette.
  • The phonetic-only dialect2vec reached an ARI of 0.602 with KMeans, near the lexical BERTimbau baseline. HDBSCAN performed poorly (ARI 0.013), merging classes and producing many noise points.
  • Confusion matrices separate Curitiba clearly, while Salvador and Três Rios overlap. Extrinsic evaluation of the full hybrid embedding and more ALiB localities are left for future work.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

Cite

Cite as Laila Mota, Daniela Barreiro Claro, Eloize R. Marques Seno and Rerisson Cavalcante de Araújo. 2026. dialect2vec: Um método baseado em vetores para transcrição dialetal do português a partir de questionários do ALiB. Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{mota2026dialect2vec,
  title = {{dialect2vec: Um método baseado em vetores para transcrição dialetal do português a partir de questionários do ALiB}},
  author = {Laila Mota, Daniela Barreiro Claro, Eloize R. Marques Seno and Rerisson Cavalcante {de Araújo}},
  booktitle = {Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1},
  year = {2026},
  url = {https://aclanthology.org/2026.propor-1.79/}
}

References

8 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ricardo J. G. B. Campello, Davoud Moulavi, and Joerg Sander. 2013. Density-based clustering based on hierarchical density estimates. In Advances in Knowledge Discovery and Data Mining, pages 160–172, Berlin, Heidelberg. Springer Berlin Heidelberg. find
  2. Suzana Cardoso and Jacyra Mota. 2014. Atlas Linguístico do Brasil. Addison-Wesley Longman Publishing Co., Inc. find
  3. Yi-Chen Chen, Sung-Feng Huang, Chia-Hao Shen, Hung-yi Lee, and Lin-shan Lee. 2018. Phonetic-and-semantic embedding of spoken words with applications in spoken content retrieval. In 2018 IEEE Spoken Language Technology Workshop (SLT), pages 941–948. find
  4. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805.
  5. Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. 2022. Byt5: Towards a token-free future with pre-trained byte-to-byte models. Preprint, arXiv:2105.13626. find
  6. Zixiaofan Yang and Julia Hirschberg. 2019. Linguistically-informed training of acoustic word embeddings for low-resource languages. In Interspeech 2019, pages 2678–2682. find
  7. Vilém Zouhar, Kalvin Chang, Chenxuan Cui, Nate B. Carlson, Nathaniel Romney Robinson, Mrinmaya Sachan, and David R. Mortensen. 2024. PWESuite: Phonetic word embeddings and tasks they facilitate. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resource… find
  8. Anjie Fang, Simone Filice, Nut Limsopatham, and Oleg Rokhlenko. 2020. Using phoneme representations to build predictive models robust to asr errors. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’20, page 699–708, New find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mota-2026-dialect2vec-um-metodo-baseado-em/