부산대학교 도서관

With a sharp rise in the number of users and skilled fluency in the use of “Hinglish”, a cross-lingual text of Hindi and English, in a linguistically diverse country, India, it has increasingly become important to analyze social content written in the aforementioned language over platforms such as Twitter, Reddit, Facebook. This project focuses on the classification of Hinglish text into offensive and non-offensive categories. We present a state-of-the-art novel use of “Longformer” and “Bigbird” transformer-based language modeling models on Hinglish texts for natural language processing (NLP) sentiment classification task, alongside mBERT, pre-trained for Hinglish, hence “HingMBert” model. Our dataset is a combination of pre-existing datasets and further addition of approximately 8,200 records, a total of 24,660 records. The tweepy library has been used to scrape novel tweets. Experimental results for Longformer, Bigbird, and HingMBert have achieved F1 scores of 0.76, 0.77, and 0.79, and accuracy of 76%, 77%, and 80% respectively. F1 score of 0.78 and an accuracy of 78% was achieved using the majority voting ensemble technique.