Rating Predictor App using Sentiment Analysis

This repository contains code and resources for a Rating Predictor App that utilizes the HuggingFace's Gradio library and Transformers library with BERT for sentiment analysis. The app takes input in the form of reviews and predicts the corresponding rating (1 to 5 stars) based on the sentiment of the review.

Table of Contents

• Dataset and Model
• Prerequisites
• Process Overview
• Challenges
• Updates

Dataset and Model

I have used the Yelp Reviews Dataset for this model.

The base model that is used is the bert-base-cased model. It is a pre-trained model on English language using a masked language modeling (MLM) objective.

The model after fine-tuning has been saved here.

Prerequisites

Before you begin, ensure you have the following prerequisites:

• Python (>=3.6)
• Pip package manager

Process Overview

1. Install the required libraries (datasets, transformers, gradio, accelerate, evaluate)

2. Next, load the dataset yelp_reviews_full from datasets

3. Load tokenizer from the pre-trained BERT Model

4. Next, fine-tune the model using Sequence Classification and then train the model using the labels created.

5. Finally, after training the model, save the model locally and then load it for your specific application.

6. In the specific task, create a tokenizer function to process the user inputs and return the output (In my application, I have given the confidences corresponding to each label as the output.)

7. Next, simply launch your model using Gradio GUI.

   And, we're done 🥳

Now, enter any review text relating to a topic of your choice and click the "Submit" button. The app will process the input using BERT-based sentiment analysis and provide a predicted rating.

Challenges

Any project cannot be completed without facing challenges. Some of them which might be helpful in future are

• Converting a tensor to a numpy array if using argmax for prediction
• Installing accelerate after importing transformers causes issues. It is better to install all libraries together at the start.
• If you are using GPU acceleration, then you need to add .to('cuda') in predictions to be able to predict successfully.

Enjoy! 🚀

Updates

• 13^th December 2023 - The application has now been deployed on HF Spaces for accessibility to a broader interested audience. You can view it at rating_predictor