Enhancing Sentiment Analysis of Corruption-Related Comments on Social Networks in Ecuador Using Techniques of Machine Learning: LSTM, Linear Regression, and Support Vector Machine

This research focuses on obtaining effective models that can predict a person’s sentiments from written messages. While the initially proposed models were within acceptable ranges as classifiers, applying quality metrics revealed a new perspective. For this research a dataset from comments from Facebook were collected and these comments were processed in two manners. the text was first processed as a normal bag of words and the second process was a TF-IDF. Both methods were used to train 3 different models the first model is a support vector machine, the second model is a linear regression, and the third model was long-short term memory model. Upon processing our dataset with the best model, the LSTM, which achieved an accuracy of 92% and obtained the highest scores in quality metrics. Although the model was not perfect, an interesting phenomenon emerged regarding sentiment classification. A considerable percentage of comments were classified as positive concerning the search hashtag, suggesting possible support for political actors. However, the vast majority of classified comments expressed negative sentiments. This leads to a significant conclusion that evaluating sentiment in social media comments is a complex challenge due to the presence of emojis and informal writing with spelling errors and inconsistencies, which impact model results.