@InProceedings{bennett:2026:wmt,
  author    = {Bennett, Eric R.},
  title     = {Synthetic Slang Generation for Benchmarking Machine Translation},
  booktitle      = {Proceedings of the Eleventh Conference on Machine Translation},
  month          = {October},
  year           = {2026},
  address        = {Budapest, Hungary},
  publisher      = {Association for Computational Linguistics},
  pages     = {33--46},
  abstract  = {Slang presents a persistent challenge in machine translation (MT) because its informal, socially situated meanings can depend heavily on context, especially when a system has not encountered the form or sense in training. We introduce a language-adaptable pipeline for generating and validating realistic synthetic slang, and use it to create a benchmark of 1,087 items across Chinese, Russian, and Farsi. We evaluate ten large language model (LLM)-based MT systems under three conditions: translation from a single usage, translation with an additional example of usage in context, and translation conditioned on a gold definition. We find additional context consistently improves translation quality, especially on slang types which are most difficult without assistance. The benefit of providing the gold definition generally grows as model size shrinks. We also test the ability for models to recover the synthetic slang definition, rising to 99\% accuracy with 2 examples in context. These results demonstrate how synthetic slang can provide a controlled stress test of meaning inference from context in MT systems.},
  url       = {https://aclanthology.org/2026.wmt-1.3}
}