Emotion Recognition from Text

This repository contains an end-to-end solution for recognizing emotions from text. The project leverages a fine-tuned DistilBERT model from Hugging Face to extract features from text and uses an MLP classifier with label-specific threshold adjustments to improve classification performance.

Table of Contents

• Overview
• Dataset
• Installation
• Usage
• Model Training & Evaluation

Overview

The project implements a pipeline that consists of:

• Fine-Tuning DistilBERT: Starting with a pre-trained DistilBERT model available on Hugging Face, we fine-tune it on our emotion recognition dataset.
• MLP Classifier with Threshold Adjustments: The output vector from the DistilBERT model is passed to an MLP classifier. A hyperparameter search is conducted to optimize the classifier and adjust decision thresholds for each emotion label.

This two-step approach has shown measurable improvements in certain metrics, as detailed below.

Dataset

• Source: The datasets were provided by the teaching assistant team as part of an ML course.
• Availability: The dataset files are included in the repository.

Installation

This project is implemented in Python using the Hugging Face Transformers library and PyTorch.

1. Clone the Repository:

   git clone https://github.com/Arash-ra03/Emotion-Recognition-Using-DistilBERT.git
   cd Emotion-Recognition-Using-DistilBERT
   

Usage

The project consists of two main Jupyter Notebooks:

DistilBERT Fine-Tuning Notebook:
File: DistilBERT_Emotion_Recognition_FineTuning.ipynb
Purpose: Fine-tune the pre-trained DistilBERT model on the provided emotion dataset.

MLP Classification Notebook:
File: FineTuned_DistilBERT_MLP_Classification.ipynb
Purpose: Utilize the output vectors from the fine-tuned DistilBERT model as input to an MLP classifier. This notebook includes a hyperparameter search to optimize performance and adjust thresholds for each emotion label.

Inference

For inference, we have implemented a simple Streamlit app. This app allows users to input text and receive emotion predictions in real time.

Running the Streamlit App

1. Run the Streamlit App:

   streamlit run emotion_app.py
   

Model Training & Evaluation

Fine-Tuning DistilBERT

Initial Performance (Before MLP and Threshold Adjustments):

• F1-micro: 0.47
• Accuracy: 0.35
• AUC-ROC: 0.68
• Precision-micro: 0.57
• Recall-micro: 0.40

MLP Classifier with Threshold Adjustments

After Applying the MLP Classifier:

• F1-micro: 0.53
• Accuracy: 0.33
• AUC-ROC: 0.87
• Precision-micro: 0.48
• Recall-micro: 0.59

Details

• The MLP classifier receives the output vector from the fine-tuned DistilBERT model.
• A hyperparameter search is performed to find the optimal model configuration.