Adapter-based fine-tuning of pre-trained multilingual language models for code-mixed and code-switched text classification

Code-mixing and code-switching are frequent features in online conversations. Classification of such text is challenging if one the languages low-resourced. Fine-tuning pre-trained multilingual language models a promising avenue for code-mixed classification. In this paper, we explore adapter-based fine-tuning PMLMs CMCS We introduce sequential parallel stacking adapters, continuous training adapters without freezing original model as novel techniques with respect to single-task also present newly annotated dataset classification Sinhala–English code-switched data, where Sinhala low-resourced language. Our 10000 user comments has been manually five tasks: sentiment analysis, humor detection, hate speech identification, aspect thus making it first publicly available largest number task annotation types. addition dataset, tested our proposed on Kannada–English Hindi–English datasets. These experiments confirm that PMLM outperform or par basic models.