Sentiment analysis on ChatGPT based on Twitter data: A comparison between different algorithms

Abstract

Advanced generative Artificial Intelligence (AI) models have been built as a result of the rapid growth of AI and natural language processing. ChatGPT, developed by OpenAI, is one of them. It is designed for generating human-like text responses and engaging in natural language discussions. Analyzing sentiment about ChatGPT is important for improving ChatGPT interactions, addressing real-world applications across industries, advancing natural language understanding, strengthening system robustness, and addressing ethical concerns. Therefore, the main objective of this research is to apply sentiment analysis on ChatGPT using twitter data. Today, Twitter is the main place to share sentiment publically. A Twitter dataset with 217,622 records was collected for the proposed approach. After data pre-processing, different feature vector generations such as TF-IDF, Word2Vec, Doc2Vec, and GloVe were applied to extract features. Then algorithms such as Support Vector Machine (SVM), Artificial Neural Network (ANN) and Long Short Term Memory (LSTM) were applied to classify twitter data with 84% training and 16% testing data set. Throughout our research, we examined algorithm performance using key measures such as accuracy, precision, recall, F-measure, and error values. These measures allowed us to evaluate the algorithm's efficacy in classifying sentiment as positive, negative and neutral. According to the results, the LSTM algorithm with the TF-IDF feature extraction method was the most effective solution for sentiment analysis, with 77.7% accuracy, higher recall, precision, f-measure values, and lowest error values than other algorithms. This research advances practical applications and contributes valuable insights specially on the AI-powered language model ChatGPT.