Transforming Social Media Analysis: TweetEval Benchmarking with Advanced Transformer Models

X (formerly Twitter) is one of the most popular social media platforms globally, where users interact through short messages known as Tweets. The platform's vast user base and specific rules and limitations present unique opportunities and challenges for Natural Language Processing (NLP) tasks, particularly in Tweet classification. The rapid advancement of transformer-based deep learning models has introduced numerous sophisticated models. In this work, we apply seven of these models-BERT, RoBERTa, DistilBERT, Electra, GPT-2, Longformer, Luke, T5, and XLNet. We utilize the TweetEval benchmark, a standard developed by researchers for various Tweet classification tasks, and compare our results with other state-of-the-art approaches. We experiment with learning rates ranging from 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> to 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-6</sup> and two different loss functions, Negative Log-Likelihood (NLL) and Cross Entropy (CE), to determine the best-performing combinations. Our findings show that our top models for emoji and offensive language detection surpass all other works. Additionally, our best-performing models exceed the TweetEval results in all tasks except for hate speech detection, sentiment analysis, and stance detection, where our results remain very close to the TweetEval benchmark. The source code and results for this work is available at the following link: https://github.com/MdSaifulIslamSajol/benchamarking_Tweeteval_datasets/tree/main