📄 Incorporating Multiword Expressions in Galician Neural Machine Translation: Compositionality, Efficiency, and Performance

Summary

Neural MT systems often mistranslate multiword expressions (MWEs), especially idiomatic ones. Adding such knowledge to an existing model is hard: retraining is expensive, and fine-tuning can cause catastrophic forgetting. The authors address this for Galician, a low-resource language, translating into English (a distant language) and Spanish (a closely related one). They build two parallel corpora from an existing set of 240 Galician noun-adjective MWEs covering 322 senses and 1,858 context sentences. Professional translators rendered every sentence into English and Spanish, and a second translator reviewed the work. The sentences carry annotations for compositionality class, sense ambiguity and frequency. Using the Transformer-based Proxecto Nós models built on OpenNMT, the team fine-tuned on small MWE subsets while varying the number of training steps and the corpus size. Moderate fine-tuning helps MWE handling without much forgetting, but the benefit depends on the language pair. Idiomatic MWEs remain the hardest case, particularly into English. The work matters because it offers the first curated Galician MWE translation benchmark and practical guidance for adapting MT to new expressions cheaply.

Key points

• Resource: two new parallel datasets (GL-EN and GL-ES), each with 1,858 sentence pairs covering 240 noun-adjective MWEs and 322 senses. They are released on GitHub, with a private held-out subset to prevent contamination.
• Fine-tuning setup: fine-tuning of the Proxecto Nós NMT models worked best at roughly 600–1000 steps. Enlarging the MWE fine-tuning corpus brought little or no consistent gain.
• Compositionality effect (GL→EN accuracy):
  • Idiomatic MWEs: 56.25% for the base model and 52.08% after fine-tuning.
  • Partial MWEs: about 77%.
  • Compositional MWEs: 82.54%.
  • GL→ES stays between about 93% and 99% across all classes.
• Frequency effect: high-frequency MWEs translate better than rare ones. The gap is close to 10 points for GL-EN but under 1 point for GL-ES.
• Error analysis: most errors are literal translations (e.g., montaña rusa → "Russian mountain") or wrong sense choices. For GL-EN, fine-tuning improved sense disambiguation but increased literal renderings.