Revanth-Reddy-Pingala / Abusive_Comment_Detector_BERT

Fine tuned BERT, mBERT and XLMRoBERTa for Abusive Comments Detection in Telugu, Code-Mixed Telugu and Telugu-English.

Abusive_Comment_Detector_BERT

Problem statement

• This project aims to identify and filter out content that is offensive, harmful, or disrespectful to provide safer online environment.

• With the rapid growth of online platforms and social media, there is an increasing need to address critical natural language processing task of Abusive Comment Detection in Low Resource Languages like Telugu.

Demo

Tech Stack Used

• Python, Google Colab for Training models using GPU.
• HuggingFace Transformers for model trainig and saving.
• Visual Studio (VS) Code with Conda Environment.
• Streamlit for Frontend, CSS for styles.

Live Demo Streamlit Abusive Comment Detector App

Video Demo of Local Run : Youtube Video

• Before Prediction

• After Prediction

Run Locally

Clone the project

  git clone https://github.com/Revanth-Reddy-Pingala/Abusive_Comment_Detector_BERT

Go to the project directory

  cd my-project

After setting up environment and installing packages using requirements.txt, Run the following command to Deploy app Locally

  streamlit run ./app.py

Sequence of steps followed until Deployment

• Github and Code Set Up.
• Text Preprocessing.
• Exploratory Data Analysis.
• Model Hyper Parameters tuning.
• Finetuning the BERT model using Google Colab's GPU.
• Saving the Finetuned BERT model locally.
• Pushing the saved model to Hugging Face.
• Development of web app using Streamlit.
• Getting the saved model using transformers from Hugging Face.
• Deploying the app to Streamlit Community Cloud.

Dataset Description

• This project provided in this repository is just a sample version of my Main Project at Central University of Tamil Nadu.

• The dataset is a real time dataset from Research paper using SnapChat data and another dataset from Codalab Competition. The final dataset used is a combination of both datasets.

• The Abusive Comment Dataset contains Comments in native Telugu script, Code-Mixed(combination of Telugu and English in the same comment, telugu comments written in English alphabet) Telugu and combination of Telugu-English Comments.

Dataset information

• Data : Comments present in native Telugu, code-mixed Telugu and Telugu-English.
• Sentiment : Abusive or Not Abusive

Data Checks Performed

• Missing values
• Duplicates
• data type
• Count of Abusive class and Not Abusive class

Exploratory Data Analysis (EDA)

• More Description can be found in the files of notebook folder.

Inference from Plots

• This is a balanced Dataset. So we don't have to perform dataset balancing tasks.

Models used for Training in this Repo

• BERT

Models used in my Main Project

• mBERT
• XLMRoBERTa

Evaluation metrics after Training

• loss
• accuracy
• precision
• recall
• f1 score

Accuracy for the Models

• BERT - 69% (sample dataset of 1k comments)
• mBERT - 86% (Entire dataset of 34.5k Comments)
• XLMRoBERTa - 89% (Entire dataset of 34.5k Comments)

Why sample dataset in this Repo for Deployment

• Due to limited access of cloud RAM after Deployment.

• The finetuned mBERT and XLMRoBERTa use lot of RAM while fetching and Predicting.

🔗 Links