Scripts Through Time: A Survey of the Evolving Role of Transliteration in NLP
Researchers detail how converting scripts boosts lexical overlap and inference efficiency for global LLMs.
A team of researchers—Thanmay Jayakumar, Deepon Halder, and Raj Dabre—has published a pivotal survey paper titled 'Scripts Through Time: A Survey of the Evolving Role of Transliteration in NLP.' Accepted to the ACL 2026 Findings track, this 9-page work provides a comprehensive analysis of how transliteration, the process of converting text from one writing system to another, is breaking down a major obstacle in cross-lingual Natural Language Processing. The paper identifies the 'script barrier'—where differences in scripts like Devanagari, Cyrillic, or Arabic inhibit transfer learning—as a key challenge. It demonstrates that transliteration can increase lexical overlap between languages by up to 40%, significantly boosting the performance of language models on tasks involving low-resource languages.
The survey presents a detailed taxonomy of the motivations and methods for incorporating transliteration into model inputs, from rule-based systems to neural approaches. It analyzes the evolution and effectiveness of these techniques, discussing critical trade-offs in accuracy, resource requirements, and inference efficiency. The review explores specific beneficial settings, including handling code-mixed text (e.g., Hinglish), leveraging linguistic family relatedness, and achieving pragmatic gains in inference speed. Based on this analysis, the authors provide concrete, actionable recommendations for AI researchers and practitioners on selecting and implementing the most appropriate transliteration strategy tailored to specific languages, tasks, and computational constraints.
- Identifies the 'script barrier' as a major hindrance to cross-lingual NLP, where different writing systems block effective transfer learning.
- Shows transliteration can boost lexical overlap by up to 40%, improving model performance on low-resource language tasks and code-mixed text.
- Provides a practical taxonomy and concrete recommendations for researchers to choose transliteration methods based on language, task, and resources.
Why It Matters
Enables more effective AI for thousands of global languages by breaking down script-based barriers, crucial for inclusive technology.