Assisting Access to COVID-19 Information Through Deep Learning Based Machine Translation: Attention Mechanism Via Bidirectional GRU
Daniel Chang · American Journal of Data Mining and Knowledge Discovery · 2021
Due to the recent COVID-19 crisis, there is an increasing need for effective communication and sharing of information internationally in various fields. One of the obstacles that these needs face are language: In texts such as COVID-19 related research, currently existing machine translations which are effective in normal texts because they are trained with normal-context data are often inaccurate, and manual translation is slow and laboursome. So, the exchange of information is being delayed. To overcome this language barrier, this project aimed to create a model that is effective for translating COVID-19 crisis related data specifically. In the research, there are two models created: one is trained with TAUS English-French Corona Crisis Corpus, and another used transfer learning by Kaggle English-French corpus and then trained with TAUS corpus. The model consisted of four bidirectional GRU layers, and used rmsprop as optimizer. The project evaluated the model using the BLEU score. The first model had a higher BLEU score than the second model, supporting the hypothesis that loosely related datasets decrease the quality of translation. In further research, evaluation on this model on different language pairs and use datasets in other specific fields will be conducted.