Emoji Context-Aware Tweets Classification using LSTM and Bi-LSTM

Project Overview

This project was developed as part of my M.Tech Final Year (Phase-I) project. The objective is to classify tweets into Hate and Non-Hate categories using context-aware deep learning techniques and emoji-enhanced textual representations.

The project uses the Olympics Twitter dataset and incorporates emoji information to improve the understanding of tweet context and sentiment.

Methodology

1. Data Preprocessing

• Dataset cleaning and preprocessing
• Removal of unwanted characters and noise
• Text normalization
• Manual emoji integration and emoji-aware tweet representation

2. Feature Extraction

Two feature extraction techniques were implemented:

Word Embedding

• GloVe (Global Vectors for Word Representation)

N-Grams

• Unigrams
• Bigrams
• Trigrams

Although multiple N-Gram approaches were explored, Bigrams were primarily used in the classification models.

3. Deep Learning Models

The following models were developed and evaluated:

• LSTM with GloVe Embeddings
• Bi-LSTM with GloVe Embeddings
• LSTM with Bigrams (Top 10 Features)
• LSTM with Bigrams (Top 5000 Features)
• Bi-LSTM with Bigrams (Top 10 Features)
• Bi-LSTM with Bigrams (Top 5000 Features)

4. Clustering and Attention Mechanism

To improve contextual understanding, DBSCAN clustering was integrated with the extracted features and combined with:

• LSTM + Attention Mechanism
• Bi-LSTM + Attention Mechanism

5. Evaluation

Model performance was evaluated using:

• Accuracy
• Precision
• Recall
• F1-Score
• Confusion Matrix

Additional Information

• All implementations were developed using Jupyter Notebook.
• Some ZIP files included in the repository contain datasets and outputs generated during program execution.
• Transfer Learning experiments were also attempted; however, training and execution required significant computational resources and time.

Technologies Used

• Python
• Jupyter Notebook
• TensorFlow
• Keras
• Scikit-Learn
• GloVe Embeddings
• DBSCAN Clustering
• LSTM
• Bi-LSTM
• Attention Mechanism
• N-Grams

Dataset

Olympics Twitter Dataset

Classification Task

Hate Speech Detection

Output Classes:

• Hate
• Non-Hate

Transfer Learning (Work in Progress)

As an extension of the project, Transfer Learning was explored to improve hate speech classification performance.

Transfer Learning experiments were conducted using a pre-trained hate speech classification dataset. Preliminary implementations were developed to transfer learned knowledge to the Olympics Twitter dataset. Due to training time and hardware limitations, the experiments remain a work in progress and are included as a foundation for future research.