Making Machine Translation Better: A Cantonese-English Case Study

“

The research focused on data augmentation techniques (back- and forward-translation) for synthetic data generation and model-switch mechanisms for Cantonese-to-English Neural Machine Translation (NMT). Additionally, the project resulted in the open-source release of the CANTONMT toolkit and the underlying corpora. The four primary objectives outlined in the initial section were successfully accomplished: A new parallel dataset was constructed by merging existing corpora, offering a valuable resource for researchers in Cantonese Natural Language Processing (NLP). Additionally, a substantial Cantonese monolingual dataset was scraped and processed from an online forum, including anonymization and data cleaning. This dataset is particularly valuable for future research, such as training Cantonese word embeddings for downstream NLP tasks, given its unique nature and limited public availability. Model Investigations: Novel Methods for Cantonese-English Translation: Several models were developed and trained using back-translation to generate a synthetic parallel corpus for fine-tuning. The best-performing models achieved comparable results to state-of-the-art commercial translators like Microsoft Bing and Baidu Translators, despite limitations in computational resources and corpus size. This work is pioneering in applying back-translation-generated synthetic data and model switch mechanisms to enhance Cantonese-English Neural Machine Translation (NMT) in the field. Experimental Evaluations: Both Automatic and Human Evaluations: Extensive experiments were conducted to evaluate the performance of trained models and commercial translators, employing a variety of automatic evaluation metrics from lexicon-based (BLER, hLEPOR) and embedding-space (COMET, BERTscore) categories. Moreover, a modified HOPE framework was used for human evaluation, providing a clearer understanding of each model's strengths and weaknesses. CANTONMT User Interface: A highly modular, full-stack web application was designed and developed as an open-source translation platform. This platform serves as a toolkit for researchers to add different models and language pairs, fulfilling the investigation's objective of creating an open-source Cantonese-to-English translation tool.

”