CLJun 21

ROMEVA: Geometry-Preserving Vocabulary Expansion for Roman Urdu Language Models

arXiv:2606.224784.3
Predicted impact top 99% in CL · last 90 daysOriginality Synthesis-oriented
AI Analysis

For low-resource NLP practitioners working on morphologically inconsistent languages, this work highlights that embedding preservation may not always benefit downstream tasks.

ROMEVA proposes a vocabulary expansion method for Roman Urdu that preserves pretrained embeddings, but naive fine-tuning achieves better downstream sentiment classification, revealing a trade-off between embedding stability and performance.

Multilingual Language Models like mBERT are widely used for low-resource NLP, yet their adaptation to morphologically inconsistent languages such as Roman Urdu remains underexplored. Roman Urdu spelling variation causes severe sub-word fragmentation, averaging 1.50 sub-words per token. We propose \textit{ROMEVA} (Roman Urdu Embedding-preserving Vocabulary Adaptation), which combines sub-word-average initialization and a PCA-guided anchor loss to stabilize embeddings during vocabulary expansion. Using a 36,130-comment Roman Urdu corpus, we add 500 highly fragmented tokens to mBERT and compare naive fine-tuning, sub-word-aware fine-tuning, and \textit{ROMEVA}. While \textit{ROMEVA} most effectively preserves the pretrained embedding space, naive fine-tuning achieves the strongest downstream sentiment classification performance. These findings reveal a disconnect between embedding stability and downstream performance, suggesting that stronger adaptation may be preferable to strict embedding preservation in morphologically inconsistent languages.

Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes