pagantibet/Tibetan-abbreviation-dictionary

Tibetan Normalisation - Abbreviation Dictionary

A custom-built dictionary of ~10,400 Classical Tibetan abbreviation–expansion pairs, mapping abbreviated or contracted diplomatic forms to their Standard Classical Tibetan equivalents. This dictionary is used as the basis for the rule-based normalisation component of the PaganTibet inference pipeline, and also as a data augmentation resource during training.

Abbreviations are one of the most systematic and frequent sources of deviation from standard orthography in historical Tibetan manuscripts. This dictionary encodes scribal conventions — contracted syllables, merged characters, elided particles, and shorthand forms — enabling both direct rule-based replacement and neural model training on a wide range of abbreviation types.

This dataset is part of the PaganTibet project and accompanies the paper:

Meelen, M. & Griffiths, R.M. (2026) 'Historical Tibetan Normalisation: rule-based vs neural & n-gram LM methods for extremely low-resource languages' in Proceedings of the AI4CHIEF conference, Springer.

Please cite the paper and the code repository when using this dataset.

---

Dataset Description

The dictionary contains 10,442 rows (including a header row), each representing a mapping from a diplomatic abbreviated form to its Standard Classical Tibetan expansion. Both forms are enclosed in square brackets. The file is tab-separated with a header row:

Diplomatic	Normalised
[རིལུ་]	[རིལ་བུ་]
[ཀུནསྱུད]	[ཀུན་ནས་འཁྱུད་]
[ཀུནགའ]	[ཀུན་དགའ་]
[ཁྖེྃས་]	[ཁ་ཆེམས་]
[སུྃ]	[སུམ་ཅུ་]
[སུྃ]	[སུམ་]

Key Properties

One diplomatic form may have multiple expansions. The same abbreviated form can be a contraction of different standard forms depending on context. These are represented as separate rows, giving the inference script multiple candidates to evaluate. For example:

[སུྃ]	[སུམ་ཅུ་]
[སུྃ]	[སུམ་]
[སུྃ]	[སུམ]

Abbreviations reflect diverse scribal conventions, including: syllable merging (multiple syllables compressed into one character sequence), diacritic elision (vowel marks omitted), particle contraction (grammatical particles absorbed into the preceding syllable), and shorthand character substitutions specific to certain manuscript traditions.

Square brackets delimit the diplomatic and normalised forms in each entry, as expected by the inference and augmentation scripts.

---

Uses

The dictionary serves two distinct roles in the PaganTibet normalisation pipeline:

1. Rule-Based Inference (Modes 1, 4, 5, 6)

In four of the six inference modes, the dictionary is used at runtime to perform direct abbreviation expansion via string replacement. Rules are applied either as preprocessing (before the neural model) or postprocessing (after the neural model and KenLM ranking), depending on the chosen mode. In addition to abbreviation expansion, the rule-based component also fixes punctuation: replacing ༑ and ༎ with །, adding spaces after །, and removing double tshegs (་་ → ་). See the Inference ReadMe for full usage details on all six inference modes.

The inference modes that use the dictionary are:

Mode	Dictionary role
rules	Rule-based only — abbreviation expansion as the sole normalisation method
neural+lm+rules	Postprocessing — rules applied after neural + KenLM (recommended)
rules+neural+lm	Preprocessing — rules applied before neural + KenLM
rules+neural	Preprocessing — rules applied before neural model only

To use the dictionary in inference:

python3 tibetan-inference-flexible.py \
    --mode neural+lm+rules \
    --model_path tibetan_model_nontokenized_allchars.pt \
    --kenlm_path model_5gram_char.arpa \
    --lm_backend python \
    --rules_dict abbreviations.txt \
    --input_file input.txt

The dictionary file should be passed as the --rules_dict argument. The S2S models (tokenised and non-tokenised) and KenLM rankers (tokenised and non-tokenised) used in the PaganTibet normalisation pipeline are available on HuggingFace.

2. Data Augmentation During Training

The dictionary is also used during the construction of the training dataset pagantibet/normalisation-S2S-training via the dictionary-based augmentation script. Entries are injected into random training lines, exposing the neural model to a wide range of abbreviation–expansion pairs and helping it learn to resolve abbreviated forms even in contexts not covered by the gold-standard data.

python3 dictionary-augmentation.py input.txt abbreviation-dictionary.txt

See the Dictionary_Augmentation ReadMe for full details.

---

Intended Use

This dictionary is intended for:

• Rule-based Classical Tibetan normalisation, either as a standalone normalisation method or as pre-/post-processing in combination with a neural Seq2Seq model.
• Data augmentation for training sequence-to-sequence normalisation models on Classical Tibetan.
• Research on Tibetan manuscript abbreviation systems, scribal conventions, and low-resource historical text processing.
• Extension: the dictionary can be expanded with additional abbreviation–expansion pairs from other manuscript traditions and used with the same inference and augmentation scripts.

---

Related Models and Resources

Resource	Link
Training dataset	pagantibet/normalisation-S2S-training
Test datasets	pagantibet/Tibetan-normalisation-testdata
Non-tokenised Seq2Seq model	pagantibet/normalisationS2S-nontokenised
Tokenised Seq2Seq model	pagantibet/normalisationS2S-tokenised
Non-tokenised KenLM ranker	pagantibet/5gram-kenLM_char
Tokenised KenLM ranker	pagantibet/5gram-kenLM_char-tok
Inference scripts	github.com/pagantibet/normalisation/Inference
Inference ReadMe	Inference_ReadMe.md
Data augmentation scripts	Data_Augmentation
Dictionary augmentation ReadMe	Dictionary_ReadMe.md
PaganTibet project	pagantibet.com

---

License

This dataset is released under CC BY-NC-SA 4.0. It may be used freely for non-commercial research and educational purposes, with attribution and under the same licence terms.

---

Funding

This work was partially funded by the European Union (ERC, Pagan Tibet, grant no. 101097364). Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency.