Pho-BP

Pho-BP (Phonological Benchmark for Brazilian Portuguese) is a complete multitask benchmark for evaluating phonologically informed representations in Brazilian Portuguese. This release contains 46,000 labeled examples distributed across four tasks and fixed train/validation/test partitions.

Tasks and sizes

Configuration	Task	Train	Validation	Test	Primary metrics
rhyme	Binary rhyme detection	10,000	1,000	1,000	Accuracy, macro-F1
syllable	Character-boundary syllable segmentation	8,000	1,000	1,000	Boundary/token F1
initial_phoneme	Initial phoneme classification	15,000	1,500	1,500	Accuracy, macro-F1
reading_error	Reading-error type classification	5,000	500	500	Accuracy, macro-F1

The rhyme splits are exactly balanced. The syllable task has an overall mean word length of 8.70 characters. Initial-phoneme training is exactly balanced across 24 classes; validation and test differ by at most one example per class. Reading errors use a deliberately imbalanced distribution: 45% phoneme substitution, 25% omission, 20% orthographic confusion, and 10% insertion.

Public-release provenance

The lexical tasks were reconstructed from the open Brazilian Portuguese lexicon fserb/pt-br (MIT), ranked by its ICF frequency resource. Pronunciations were generated with the deterministic Pho-BP rule-based Brazilian Portuguese G2P procedure and syllable boundaries with Pyphen pt_BR.

The row-level child oral-reading corpus described in the research manuscript was not available among the supplied files. To avoid fabricating observations attributed to children or publishing potentially identifying educational records, the reading_error configuration is a complete privacy-preserving rule-based public release that instantiates the same four error categories without participant identifiers. The field source_type makes this provenance explicit. The baseline JSON files preserve the values reported in the thesis and identify them as reported results; they were not recomputed on this reconstructed public release.

Loading from Hugging Face

from datasets import load_dataset

rhyme = load_dataset("YOUR_NAMESPACE/pho-bp", "rhyme")
syllable = load_dataset("YOUR_NAMESPACE/pho-bp", "syllable")
initial = load_dataset("YOUR_NAMESPACE/pho-bp", "initial_phoneme")
reading_error = load_dataset("YOUR_NAMESPACE/pho-bp", "reading_error")

The YAML metadata above maps each configuration to its three CSV files, so no custom loading script is required.

Schemas

Rhyme detection

id, word1, word2, ipa1, ipa2, rhyme_key1, rhyme_key2, label, source.

label=1 indicates equal phonetic rhyme keys; label=0 indicates different keys.

Syllable segmentation

id, word, ipa, syllables, boundary_labels, num_syllables, source.

boundary_labels is a JSON array of length len(word)-1; value 1 marks a syllable boundary after that character. Example: casa → [0, 1, 0].

Initial phoneme

id, word, ipa, initial_phoneme, label_id, source.

The 24 classes in label-id order are: a, k, p, e, s, d, m, i, ʁ, t, b, f, l, v, g, o, n, ʒ, ʃ, u, z, ɲ, ʎ, tʃ.

Reading error

id, target_word, observed_reading, target_ipa, observed_ipa, error_type, error_span, generation_rule, source_type.

The four labels are phoneme_substitution, omission, orthographic_confusion, and insertion.

Evaluation

Each file in scripts/ accepts a gold CSV and a prediction CSV. Prediction files must contain id and prediction, except syllable segmentation, which accepts id plus prediction containing either a JSON boundary array or a hyphenated segmentation.

python scripts/evaluate_rhyme.py   --gold test/rhyme_test.csv   --predictions predictions/rhyme_predictions.csv

Baselines

baselines/ contains the results reported for BERTimbau, XLM-R, and Lacium Flower. Classification task entries contain accuracy and macro-F1; syllable segmentation contains token/boundary F1. All values are percentages and correspond to the mean of three runs as reported in the thesis.

Intended use

Pho-BP is intended for research on Brazilian Portuguese phonology, representation learning, multitask learning, literacy-support NLP, probing, and controlled comparison of models. It is not a diagnostic instrument and must not be used to make clinical, educational-placement, or high-stakes decisions about individual learners.

Limitations

Pronunciations and syllabification are generated automatically and may reflect dialect-neutral or tool-specific conventions. Rhyme keys depend on the rule-based G2P stress assignment. The public reading-error subset is rule-based rather than a substitute for ecologically valid, consented oral-reading data. Results reported on the original internal benchmark should not be presented as newly measured results on this release.

Citation

Use CITATION.cff or cite the dataset as:

Gomes da Cruz Junior, G., & Fidalgo, R. N. (2026). Pho-BP: Phonological Benchmark for Brazilian Portuguese (Version 1.0.0) [Dataset].

License

MIT. See LICENSE, including the attribution notice for the source lexicon.