Optimized Fine-tuning and Pseudo-Data Strategies for Cross-Domain Low-Resource Language Cantonese-English Neural Machine Translation
Bibliographic record
Abstract
Neural machine translation for low-resource languages like Cantonese remains challenging due to limited parallel data and domain adaptation difficulties. Moreover, standardized processing pipelines for minor languages are scarce. Our study presents a rounded framework for Cantonese–English neural machine translation that addresses data scarcity, model optimization, and evaluation rigor. We curate a parallel corpus of 1.1 million sentences from parliamentary records, subtitles, educational texts, and web sources, and implement a multi-stage cleaning pipeline contains LaBSE-based semantic alignment. We fine-tune the NLLB-200-600M model using optimized training configurations. To mitigate data limitations, we develop a domain-adapted reverse translation model to generate pseudo-parallel data and employ a layered sampling strategy to create subsets of varying scales. Our systematic evaluation identifies a 200k pseudo-data to 200k original data ratio as optimal. The best performing model achieves a BLEU score of 29.63 and 56.65 of chrF on in-domain tests, and a BLEU score of 15.87 on legal-domain tests, outperforming NLLB-200-distilled-600M in cross-domain generalization. This work establishes a reproducible pipeline for low-resource translation and underscores the importance of tailored data strategies and rigorous evaluation.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.004 | 0.020 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.001 | 0.003 |
| Open science | 0.002 | 0.003 |
| Research integrity | 0.001 | 0.003 |
| Insufficient payload (model declined to judge) | 0.005 | 0.004 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".