aiqbench analysis
This study reveals a "curse of multilinguality" in lexical normalization, demonstrating that for compact models, jointly training on too many languages causes accuracy to drop by up to forty percent due to internal resource competition. The findings suggest that limiting co-training to just one or four languages often yields superior performance compared to pooling all available data into a single massive model. This insight is critical for optimizing low-resource language processing pipelines where model capacity is constrained.
The Curse of Multilinguality in Lexical Normalization
arXiv:2609.00329v1 Announce Type: cross Abstract: Lexical normalization rewrites the noisy, non-standard words that fill user-generated text (tmrw, u, gr8) into their standard forms. Because labelled data is scarce for most languages, a popular shortcut is to train a single model on many languages at once. We ask a simple question: how many languages should such a model be trained on? Using one fixed-capacity character-level model and twelve languages from a standard benchmark, we vary the number of jointly trained languages from one to twelve and measure per-language accuracy. We find a clear curse of multilinguality: accuracy is highest when a language is trained with only a few others, often just one to four, and then falls steadily and substantially, dropping by about forty percent as the rest are piled on. A control that holds the total amount of training data constant makes the decline arrive sooner and fall further, which points to competition among the languages for one fixed-size model rather than to how much data is available. We also test whether a language's typological distance from the others predicts its ideal number of co-training languages, and find no dependable rule: any apparent relationship rests on a couple of languages and does not hold up. For compact normalization models, less can be more: a few languages beat pooling everything into a single model.
Abstract:Lexical normalization rewrites the noisy, non-standard words that fill user-generated text (tmrw, u, gr8) into their standard forms. Because labelled data is scarce for most languages, a popular shortcut is to train a single model on many languages at once. We ask a simple question: how many languages should such a model be trained on? Using one fixed-capacity character-level model and twelve languages from a standard benchmark, we vary the number of jointly trained languages from one to twelve and measure per-language accuracy. We find a clear curse of multilinguality: accuracy is highest when a language is trained with only a few others, often just one to four, and then falls steadily and substantially, dropping by about forty percent as the rest are piled on. A control that holds the total amount of training data constant makes the decline arrive sooner and fall further, which points to competition among the languages for one fixed-size model rather than to how much data is available. We also test whether a language's typological distance from the others predicts its ideal number of co-training languages, and find no dependable rule: any apparent relationship rests on a couple of languages and does not hold up. For compact normalization models, less can be more: a few languages beat pooling everything into a single model.
Accepted at the 11th Workshop on Natural User-generated Text (W-NUT 2026), EMNLP 2026. Camera-ready version. 7 pages, 3 figures, 3 tables
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2609.00329 [cs.CL]
(or arXiv:2609.00329v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2609.00329 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 340 words
Related coverage
Translation as a Decision Space: A Multi-Agent Perspective on Low-Resource Dialect Generation
arXiv · Sep 5, 2026
IdiomX A Multilingual Benchmark for Idiom Understanding, Retrieval, and Interpretation
arXiv · Jun 3, 2026
Retrieving Biblical Intertextual References in Karen Blixen's Seven Gothic Tales
arXiv · Sep 30, 2026
Cross-Stage Coherence in Hierarchical Driving VQA: A Consistency Protocol and Parameter-Efficient Gated Context Routing
arXiv · Oct 15, 2026