A survey of the evolving role of transliteration in natural language processing.
The central problem
Related languages can share meaning, grammar, and sound—yet appear completely unrelated to a model when their writing systems differ.
Transliteration converts text into a shared script, increasing lexical overlap and creating anchor points for cross-lingual transfer. It can unlock low-resource languages, normalize code-mixed text, and make inference cheaper.
But every bridge has a toll: collapsed meanings, spelling variation, and the risk of severing useful native-script connections.
One romanization, two meanings
A single Latin form can collapse distinct words in the same language into one ambiguous surface.
Two decades of change
The paper identifies five motivations that emerged as NLP itself evolved.
Pre-neural era
Transliteration first served as a repair tool for names missing from statistical translation dictionaries.
Social web
Romanized language became everyday user text, making normalization essential for real-world NLP.
Multilingual pretraining
A shared script exposes similarities between related languages hidden behind different writing systems.
Scaling era
Lower token fertility can shorten sequences, unify preprocessing, and reduce inference cost.
Now
The goal shifts from bypassing the barrier to teaching models to understand multiple scripts directly.
How to integrate transliteration
Transliteration can change the data, the input, the architecture, or only the inference path.
Train on transliterated corpora, mix them with native text, or adapt the vocabulary.
Concatenate both forms or fuse their embeddings before the model processes them.
Use multiple encoders, script adapters, or objectives that explicitly align representations.
Ensemble outputs across scripts, add script prefixes, or prompt LLMs with multiple forms.
Put the survey to work
There is no universal best method. Choose based on the linguistic anchor, output type, and engineering constraints—not habit.
Bridge to the related high-resource language with a shared script. Validate that the conversion preserves the semantic features your NLU task needs.
Form over fidelity
Not because it is always linguistically faithful—because today’s model ecosystem makes it pragmatic.
Many LLMs have far more Latin-script data in pretraining, making that surface easier for them to process.
Romanization can mean shorter sequences, faster inference, and lower API cost.
Loanwords, code-switching, and informal online text already create large amounts of naturally romanized language.
General-purpose romanizers are easier to find than high-quality converters for every language and script pair.
The emerging clue
Recent work identifies “latent romanization”: middle-to-upper layers of English-centric models can represent non-Latin tokens as phonetic Latin approximations before resolving back into native script.
The evidence is early, but it reframes transliteration as a possible internal bridge—not only an external preprocessing trick.
The paper’s bottom line
Where it helps
Where it breaks
Use transliteration as scaffolding while models learn to cross the script barrier themselves.
— distilled from the survey’s conclusionNilekani Centre at AI4Bharat · Indian Institute of Technology Madras · Indian Institute of Engineering Science and Technology, Shibpur