Real-World Application of Machine Translation of Cantonese-to-Written-Chinese Translation

This paper explores integrating a Cantonese-to-Written-Chinese translation model into a real-world application. This translation model is selected amongst various pre-trained models after being trained with a large parallel corpus of Cantonese and written Chinese language pairs and a test set manually translated by the native translators to ensure evaluation reliability. We implemented the best-performing model into a subtitle-generating application with an unsaturated Cantonese market, generating subtitles in written Chinese with Cantonese audio by incorporating speech-to-text and the translation model. In the evaluation, the test set is split into high-similarity cases and low-similarity cases. The low similarity cases are considered more difficult translation tasks since the sequences have relatively more Cantonese proportions. Our model achieved similar performance with the Microsoft Azure system in high-similarity cases. At the same time, it achieved a 30.1 BLEU score and 22.6 CHRF++ score, which outperformed Microsoft Azure with a 28.5 BLEU score and 21.2 CHRF++ score in the low similarity cases. With the combination of Cantonese input from our application, using the model can also reduce processing time while being comparable with existing translation tools.