Sentiment analysis on products reviews with lexicon model and transformers

• code from google collab
• The project is based on the Amazon product review from customers dataset.
• This is a multi-class classification. Reviews should be classified into five classes 1,2,3,4,5.(from very bad to very good)

Used approaches

• EDA and Different visualization approaches:WordCloud,Pandas,Seaborn,nltk
• detection of polarity and intensity of emotion:lexicon and rule-based sentiment analysis tool---Vader4
• Sentiment analysis and multi-labels classification:Transformers provided by Hugging Face, DistilBert,Pytorch,Python

fine-tuning our pretrained model in native PyTorch on our dataset.

• Creating the model and defining its loss and optimize
• The dataloader passes data to the model based on the batch size.
• Subsequent output from the model and the actual category are compared to calculate the loss(Loss value is used to optimize the weights of the neurons in the network).
• Do regularization and early stopping to prevent overfitting

Saving the Trained Model for inference

We use pytorch :torch_save With the saved final model we can make new prediction by entering new raw text for reviewing.

Credits

The Sentiment Analysis model was fine-tuned using the DistilBERT model from Hugging Face Transformers library.