This is a sentiment analysis for Amazon food reviews using Bert, embeddings and other NLP approaches. This project was done in the context of a class on NLP during the master of AI at the university of Paris-Saclay.
The project is based on the Amazon food review from customers dataset. The dataset comes from a Kaggle competition. It can be downloaded from here (254MB).
This is a multi-class classification. Reviews should be classified into three classes Negative, Positive, Neutral.
- Simple vectorization approaches: CountVectorizer and Tf-idf with Scikit-learn.
- Different Pretrained Word Embeddings: Glove, Word2Vec and FastText with Gensim.
- Different visualization approaches: PCA, t-SNE
- Bert embeddings (DistilBERT) with Transformers - Hugging Face.
- Fine-tuned Bert with pyTorch.
| Method | F1 macro avg. |
|---|---|
| Glove embeddings + SGD | 0.46 |
| DistilBERT embeddings + SGD | 0.566 |
| DistilBERT embeddings + Logistic + stack | 0.69 |
| DistliBERT Fine-tuned | 0.75 |