Fine-Tuned mT5 Series for Global Fish Species Dual Latin-Chinese Translation
This repository contains fine-tuned versions of the mT5 series models (
mT5-large,
mT5-base, and
mT5-small) for the specialized task of fish species name translation. These models have been adapted specifically for translating between Chinese and Latin species names.
Overview
By comparing the BLEURT and COMET scores of various models, our experiments demonstrate that the fine-tuned mT5 models significantly outperform general-purpose language models such as DeepSeek-R1, Qwen-Plus, and GLM-Plus. The evaluation covers three translation tasks:
- Chinese-to-Latin
- Latin-to-Chinese
- Dual Translation
The results, summarized in Table 1 below, indicate that:
- The mT5 series consistently achieves higher scores on both evaluation metrics (BLEURT and COMET) across all translation tasks.
- fish-mT5-large stands out, outperforming its counterparts (fish-mT5-base and fish-mT5-small) by achieving BLEURT and COMET scores that are 2 to 3 times higher than those of the traditional models.
This substantial performance gap underscores the advantages of the mT5 architecture in handling the nuances of fish species name translation.
Evaluation Results
| Models | Chinese to Latin (BLEURT) | Chinese to Latin (COMET) | Latin to Chinese (BLEURT) | Latin to Chinese (COMET) | Dual Translation (BLEURT) | Dual Translation (COMET) |
|---|
| fish-mT5-large | 0.89 | 0.91 | 0.87 | 0.93 | 0.90 | 0.93 |
| fish-mT5-base | 0.80 | 0.87 | 0.77 | 0.87 | 0.80 | 0.88 |
| fish-mT5-small | 0.66 | 0.80 | 0.75 | 0.86 | 0.71 | 0.84 |
| DeepSeek-R1 | 0.44 | 0.66 | 0.56 | 0.74 | 0.45 | 0.67 |
| Qwen-Plus | 0.29 | 0.58 | 0.35 | 0.74 | 0.33 | 0.60 |
| GLM-Plus | 0.26 | 0.55 | 0.33 | 0.60 | 0.30 | 0.58 |
Table 1: BLEURT and COMET scores of six models in dual Latin-Chinese translation for global fish species.
Key Takeaways
- Superior Performance: The fish-mT5 models, particularly fish-mT5-large, demonstrate a clear advantage over traditional models, achieving markedly higher BLEURT and COMET scores.
- Robustness Across Tasks: The models perform consistently well in Chinese-to-Latin, Latin-to-Chinese, and dual translation tasks.
- Specialized Adaptation: The fine-tuning process has enabled the fish-mT5 architecture to excel in the niche task of fish species name translation, making it a valuable tool for researchers and practitioners in the field.
Usage
We provide an online demo, which can be accessed at
https://huggingface.co/spaces/MHBS-IHB/fishmt5.
Feel free to open issues or contribute to the repository if you have suggestions or improvements!
This model card introduces the models, summarizes the evaluation results with the provided table, and highlights the key performance advantages of the fish-mT5-large model for fish species name translation.
