Comparison and Optimization of Social Media Text Sentiment Analysis Models Based on Machine Learning

In this paper, a hybrid sentiment analysis framework combining lexical rules and machine learning is proposed, aiming to improve the performance of sentiment classification for English social media texts. The study first preprocesses the raw data through a multi-stage text cleaning process (including denoising, stemming extraction, and deactivated word filtering) and generates preliminary sentiment labels by calculating text sentiment scores with the VADER sentiment lexicon; further, the text is vectorized using the TF-IDF method to construct the feature space, and the passive-aggressive classifiers, logistic regression, random forests, support vector machines, and 14 models such as Simple Bayes for the classification task. In order to optimize the model performance, the study tunes the key hyperparameters by random search cross-validation, and finally compares the accuracy, precision, recall and F1 value of each model on the test set.