daricoemma/Voxy-Sentiment-Analysis

Cross-cultural sentiment analysis of Toyota Voxy reviews comparing Ghana and Japan, using VADER, Naive Bayes, and SVM classifiers # Comparative Sentiment Analysis of Toyota Voxy Reviews: Ghana vs Japan This project looks at how people in Ghana and Japan talk about the Toyota Voxy online, and whether their opinions actually differ in any measurable way. It was originally my final year thesis, and I later turned the analysis into a proper coding project with trained models instead of just a lexicon-based approach. I collected and hand-labelled 1,497 comments from reviews, forums, and social media, then trained a few different models to classify each comment as positive, negative, or neutral. Naive Bayes and a linear SVM both ended up performing far better than a standard sentiment lexicon (VADER), which made sense once I looked at how much slang and product-specific language was in the comments that a generic tool just wouldn't pick up on. ## Results VADER, which doesn't need any training, only got about 61% accuracy. Once I trained a Naive Bayes and an SVM model on the labelled data using TF-IDF features, accuracy jumped to around 95% for both, with the SVM coming out slightly ahead. | Model | Accuracy | Macro F1 | |---|---|---| | VADER (lexicon baseline) | 60.7% | 0.58 | | Multinomial Naive Bayes | 94.7% | 0.94 | | Support Vector Machine | 95.0% | 0.94 | ## What I found Japanese reviewers were generally more positive about the Voxy (about 53% positive) compared to Ghanaian reviewers (around 37% positive). Ghanaian comments had noticeably more complaints, especially around safety and issues that came up during import/conversion of the vehicle. Across both countries, comfort and fuel economy were the features people liked most, while safety and conversion-related topics got the most negative comments. ## How it works The script cleans the raw text (lowercasing, removing links and punctuation), splits the data into training and test sets, then runs three different classifiers on it: 1. VADER, a rule-based sentiment tool that doesn't need training 2. A Naive Bayes classifier trained on …