tdr-research-team/tdr-dataset

TDR: Tagalog Diacritic Restoration Dataset

TDR is a manually annotated corpus of contemporary Tagalog news text covering 40 frequent homographs, with diacritical marks (tuldik) restored to disambiguate meaning and pronunciation.

Dataset Description

Tagalog uses diacritics (tuldik) to mark syllable stress and final glottal stops, resolving ambiguity among homographs — words that share the same spelling but differ in meaning and pronunciation. For example, puno can refer to punò (tree) or punô (full). Contemporary Tagalog text rarely preserves these marks, creating ambiguity for readers and downstream NLP systems.

TDR addresses this gap by providing diacritized annotations over news articles sourced from the BalitaNLP corpus. The articles used in TDR are a subset of BalitaNLP, and the train/test split is our own.

Source and License

The underlying text comes from BalitaNLP, a collection of 351,755 Tagalog news articles from multiple Filipino news outlets. TDR selects 10,065 articles from BalitaNLP containing target homographs, and adds manual diacritization annotations on top.

The diacritization annotations are original to this work. The underlying article text is sourced from BalitaNLP, and the respective copyrights belong to the individual news outlets. Sources can be traced through the BalitaNLP dataset. This dataset is released for research purposes.

Dataset Statistics

Split	Articles	Sentences	Homograph Instances
Training article pool	9,025	—	—
Test article pool	1,040	—	—
Total	10,065	146,310	48,478

The total homograph count refers to diacritized occurrences in the full corpus, not benchmark questions. A further 82 matched strings were deliberately left undiacritized. The training and test pools are disjoint at the article level.

Homographs Covered

40 homographs grouped by asymmetry type (class imbalance between diacritized variants):

Workable	Moderate	Extreme
bukas, galing, galit, gamit, huli, ika, kilala, kita, pagod, puno, takot, taong, tapos	abala, baka, buhay, butas, dating, huling, kabilang	bagong, bilang, gabi, kaibigan, kasama, kaya, lalaki, lamang, madaling, nasa, paano, pala, pamahalaan, panalo, pangulong, patuloy, sakit, sasakyan, silang, tayo

Dataset Format

Each source file is a plain text article with target homographs replaced by their correct diacritized variant in context. Names retain the original BalitaNLP identifiers, including raw_balitanlp_train_*.txt and raw_balitanlp_test_*.txt. The train or test substring identifies the source corpus split, not the TDR split; use the containing TDR article pool to determine membership.

The evaluation repository also provides data/benchmark.jsonl, with fields id, homograph, sentence (gold reference), bare_sentence (model input), choices, gold_variant, and asymmetry_type. The benchmark contains 7,718 questions across 6,113 distinct bare sentences. Each question targets one homograph in context and provides four candidate forms. Sentences containing conflicting variants of the same homograph are excluded.

Annotation Process

• Three native Tagalog-speaking annotators performed diacritization using a main-annotator/reviewer workflow
• The UP Diksiyonaryong Filipino (UPDF) was consulted alongside speaker intuition
• The 1,040-article test pool underwent independent reannotation and consensus resolution
• A collocation-driven expansion method was used to retrieve additional articles targeting underrepresented minor variants

Limitations

• Coverage is restricted to 40 homographs in news text; performance in other domains is not established.
• Variant imbalance remains substantial, and some dictionary variants have no annotated examples.
• Independent reannotation covers the 1,040-article test pool, not the entire corpus.
• Conflicting-variant sentences are excluded from the benchmark, although practical restoration must handle them.
• The distinction between kilalá and kilála follows an operational annotation rule, not a claim that contemporary usage has an uncontested categorical distinction.

Uses

TDR is designed for:

1. Benchmarking — evaluating language models on Tagalog homograph disambiguation
2. Fine-tuning — improving LLM diacritic restoration capability
3. Training lightweight classifiers — logistic regression, RoBERTa, and other models

Citation

If you use TDR, please cite:

Ivan Yuri De Leon, Lanz Railey Fermin, Zachary Aldridge Africa, and Richell Isaiah Flores. TDR: Tagalog Diacritic Restoration Dataset. O-COCOSDA 2026. Forthcoming.

@inproceedings{deleon2026tdr,
  title     = {{TDR}: Tagalog Diacritic Restoration Dataset},
  author    = {De Leon, Ivan Yuri and Fermin, Lanz Railey and Africa, Zachary Aldridge and Flores, Richell Isaiah},
  booktitle = {Proceedings of O-COCOSDA 2026},
  year      = {2026},
  note      = {Forthcoming}
}

Contact: ivan.deleon@student.ateneo.edu.

Links

• Code and benchmark: tdr-evals (GitHub)
• Finetuned models: tdr-research-team/models