Improving Sentiment Analysis of Social Media Captions Through Advancements in NLP

This research study focused on the development and implementation of advanced natural language processing techniques and machine learning models to perform a sentiment analysis on social media captions from Twitter and Facebook. An open-source repository was used to gather data, and the compiled dataset of 1,500 records covered a variety of human emotions. The data was preprocessed to remove noise, normalize, and reduce the dimensionality using PCA. The three models, Support Vector Machine, Naive Bayes, and K-Nearest Neighbors, were trained on 70% of the dataset and tested on the remaining 30% in order to evaluate their performance to classify content into specific emotions. The results demonstrated the efficiency of all three models, with the SVM achieving the highest accuracy of 96.57% and superior precision, recall, and F1 of 0.97, 0.97, and 0.98, respectively. This machine learning model appeared to be the most efficient and powerful for high-dimensional data and multidimensional feature space, showing a 2D representation of the dataset points in 99% information retention. The NB, with an accuracy of 91.20%, demonstrated and confirmed the appropriateness of the probability approach to make accurate predictions despite the assumption of independence of features. This model used such features as 0.5, 1.2, 1.7, 1.8 to classify specific emotions into neutral or other. The final accuracy of the KNN was 88.70%, and its special features allowed clustering and classification of the data in lower dimensions using the PCA information. Thus, the three models demonstrated competitive results, but the SVM was identified as the most efficient when dealing with the data and complex pattern text classification. The confusion matrix also proved the minimum misclassification rate of SVM.