datasheet.md

Datasheet for Pho-BP

Motivation

Pho-BP was created to provide a standardized evaluation resource for phonological representation learning in Brazilian Portuguese, especially for tasks connected to phonological awareness and literacy-support NLP.

Composition

The dataset contains four configurations and 46,000 rows. Each row is a lexical or word-pair example; no row represents a named person. Splits are fixed and their sizes follow the benchmark specification.

Collection process

Lexical items were selected from an open Brazilian Portuguese lexicon using frequency-oriented ICF ordering. Phonetic transcriptions were generated automatically. Syllable boundaries were generated with language-specific hyphenation patterns. Rhyme and initial-phoneme labels were derived from the generated IPA.

The public reading-error rows are privacy-preserving rule-based realizations. They operationalize four error categories but are not represented as direct transcriptions of identifiable children.

Preprocessing

• Unicode NFC normalization
• Lowercasing
• Alphabetic filtering for Brazilian Portuguese characters
• Word-length filtering
• deterministic Brazilian Portuguese rule-based G2P conversion
• Pyphen pt_BR syllabification
• Deterministic split assignment and duplicate-pair prevention

Labels

• Rhyme: 0, 1
• Syllable: binary boundary sequence
• Initial phoneme: 24 IPA labels (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: phoneme_substitution, omission, orthographic_confusion, insertion

Quality controls

The build validation checks exact row counts, unique IDs, rhyme balance, 24-class training balance, validation/test near-balance, reading-error distributions, boundary-vector lengths, label/IPA consistency, and absence of duplicate rows within each file.

Uses

Recommended uses include probing, multitask learning, phonological representation evaluation, controlled baseline comparison, and educational NLP research at aggregate level.

Prohibited or unsupported uses include diagnosis, learner ranking, automated educational placement, surveillance, or claims that rule-generated reading errors are authentic child observations.

Distribution

The release is distributed as UTF-8 CSV files with Markdown documentation, Python evaluation scripts, machine-readable baseline JSON files, a citation file, and a license.

Maintenance

Version 1.0.0 is a deterministic public release. Future revisions should retain versioned splits, publish change logs, document tool-version changes, and separate any governed authentic oral-reading extension from the open lexical benchmark.