Twitter Sentiment Analysis with Linear SVM Classifier with TF-IDF and word2vec features.
This project describes how I developed my understanding of text sentiment classification by implementing an SVM classifier and a Logistic Regressor and performing some further experiments to seek approaches that also work. The main system that I tried to implement for the final project is the approach used by NILC-USP that participated in SemEval 2017. NILC-USP uses three classifiers which is trained in a different feature space using bag-of-words model and word embeddings to represent sentences and uses a Linear SVM and a Logistic Regressor as base classifiers. Instead of implementing the voting ensemble, I focused more on how each of three classifiers was built, strengths and drawbacks of the system, and how well it performs.
- Scikit Learn
- Gensim (Word2Vec)
- NLTK
| Classifier | AvgRec | Accuracy |
|---|---|---|
| SVM /w Tf-idf | 0.579 | 0.581 |
| SVM /w Tf-idf and Curse Word | 0.579 | 0.582 |
| SVM /w Word Embeddings | 0.548 | 0.600 |
| LogisticReg /w Weighted Word2Vec | 0.577 | 0.528 |
| SVM /w Weighted Word2Ve | 0.576 | 0.520 |