Deromanization of Hindi-English Code-Mixed Text and its Influence on Toxic Comment Classification and Sentiment Analysis

Code-mixing has become increasingly prevalent in multilingual societies, primarily due to the widespread use of social media platforms. However, processing code-mixed text presents significant challenges, mainly stemming from non-standardized spellings and poor grammatical structure. To effectively leverage pre-trained multilingual models like mBERT, we propose a strategy to deromanize code-mixed text, by which non-English words are transliterated to their native script. In this article, we introduce two categories of deromanization models. The first category comprises a language identification module, a back transliteration module, and a sequence prediction module. It is meant to be used when parallel corpus is not available. These components work together to accomplish the deromanization task by leveraging context. In the second category, we propose Neural Machine Translation (NMT) models that operate at both the character level and the sub-word level. These models offer an alternative approach to deromanization when a parallel corpus is available to train the NMT models. To train these models, we constructed a human-annotated dataset along with a synthetic dataset. This article provides a detailed description of the deromanization models employed and presents comparative results against existing models. The proposed NMT character level model demonstrates a 3% improvement in deromanization performance over the FIRE 2013 dataset, outperforming the current state-of-the-art results. Moreover, the results indicate that the process of deromanization significantly boosts the effectiveness of downstream tasks, such as toxic comment classification and sentiment analysis, with a minimum improvement of 3% in F1 score upon fine-tuning XLMR for classification.