Exploring clustering, deep learning, and LLMs in text classification

Abstract

Natural language processing and sentiment analysis are important in the current era. Many people are working in this domain to understand the human language and try to classify it. Humans are more like to express their opinions using open text rather than pre-defined questions. This study uses different unsupervised, deep learning, and Large Language Models to classify text data. Two datasets of different topics have been chosen for training and evaluating all models. It included Apple product reviews and airline tweets. The study aims to evaluate the performance of different classification algorithms and models to see which one is more accurately classifying tweets as compared to others. This study is also important as the comparison is done with the latest model of openai which is GPT-4. The findings of this research demonstrate that among all the algorithm tested, BERT based model and GPT-4 exhibit superior performance. The Roberta-based bert model depicts 81% accuracy on the Apple dataset while the bert-based-uncased model outperformed others on the airline dataset with an impressive accuracy of 95%. GPT-4 also depicts strong results with an accuracy of 79% for the Apple dataset and 85% for airline sentiment. This is a strong indication that future model of openai or other LLM models might surpass the BERT model.

These results and analysis show that LLM models like BERT and GPT-4 are more effective for sentiment classification as compared to traditional machine learning and deep learning algorithms. It is also worth noting that LLM models require less cleaning and pre-processing of datasets as those are already pre-trained models. This feature enhances efficiency and usability. This research provides potential for LLM models in text classification which also offer valuable insights for future research. Overall, this study highlights the power of the LLM model over a conventional model for sentiment data classification. It provides a detail comparison of their performance and to discuss the implificaiotn of these method for the field of natural language processing.