Leveraging Multi-Task Learning in Code-Switched setting

Multilingualism is the ability to converse fluently in more than one language. In multilingual societies, code-switching, the practise of switching languages during a conversation, occurs often, creating a need for multilingual dialogue and voice recognition systems in natural language processing (NLP). However, interpreting code-switching utterances is extremely difficult for these NLP systems, as the model must adjust to code-switching patterns. In recent years, deep learning strategies have enabled natural language systems with massive volumes of training data to attain human-level performance in languages. However, they cannot accommodate a large number of low-resource languages, mostly mixed languages. Also, code-switching, despite being a common occurrence, is unique to spoken language and lacks the transcriptions necessary for training deep learning models. Our work in this paper, explores the usage of multi-task learning in code-switched setting to see if it can leverage from the training signals of related tasks and improve the generalizability as well as boost the performance and metrics of our model.