Capstone Project Showcase

CSE499B.16

December 2, 2024

Summer 2024

Department of ECE

North South University

Bangla Text Summarizer

Transforming lengthy Bengali articles into concise summaries using advanced NLP techniques

SummarizationTransformersPyTorchBengalimt5text2text-generation

Bytes of Bengal

Model Size

300M params

Tensor Type

F32

Base Model

google/mt5-small

Type

Seq2Seq

Training Length

210 Epochs

Training Data

20k News Articles

Latest Model

Publisher

Hugging Face

License

MIT

Supervisor

Dr. Nafisa Noor [NaNr]

Assistant Professor

Team Members

Md Tashfiqul Islam

161 1593 042

Tashin Mahmud Khan

201 1819 042

Amir Hamja Marjan

202 1171 642

Md Simul Hossain

171 1949 642

Problem Statement

The digital age has brought an unprecedented surge in Bangla news content, creating a paradox of information abundance and time scarcity.

Content Overload

daily articles

1000+

Time Constraint

reading time/day

4-6h

Information Discovery

irrelevant content

60%

Content Length

words/article

1500+

Impact on Readers

Reduced reader engagement

Information overload

Digital divide growth

Project Overview

The Bangla Text Summarizer, developed for CSE499B, leverages advanced NLP to create concise, accurate summaries of Bengali news articles, tackling digital information overload.

User Input

Bangla Long Essay

Raw Content

200-5000 Words

Text Summarizer

Model Output

Summarized Content

Concise Summary

~ 150 Words

CSE499B

Senior Design Project II

Key Features

Bengali Article Library

Extensive collection of diverse Bengali articles

Bengali Article Input

Seamless integration of new Bengali content

Concise Summaries

AI-powered extraction of key information

Responsive Design

Optimal viewing on all devices

Dark Mode Support

Enhanced reading experience in low light

Real-time Processing

Instant summarization and analysis

Revolutionizing Bengali content consumption through cutting-edge AI technology

Project Timeline

Web Crawler

Month 1

Data Collection

Month 2

Model Training

Month 3-4

Model Testing

Month 5

Web Interface

Month 6

Model + UI Connection

Month 7

Web Deploy

Month 8

System Architecture

Tech Stack

Model Tech Stack

Python
v3.13.0
NumPy
v2.1.3
Pandas
v2.2.3
PyTorch
v2.5.1
spaCy
v3.8.0
Matplotlib
v3.9.2

Web UI Tech Stack

React
v19.0.0-rc
Next.js
v15.0.3
TypeScript
v5.7.2
ESLint
v9.15.0
Tailwind CSS
v3.4.1
shadcn/ui
v2.1.6

ML/Dev Ops

Hugging Face
v4.46.3
Inference API
latest
Vercel
v39.1.2

Training Loss

This table presents a comprehensive overview of our model's training progress, showcasing the evolution of loss across multiple rounds with varying epochs and data sizes.

Round	Loss	Epochs	Model	Data Size
1	6.9987	3	v3	10k
2	2.2538	3	v3	10k
3	1.0143	5	v3	10k
4	2.5004	100	v4	20k

BLEU & METEOR Scores

Content Coverage

BERTScore

Light Mode

Dark Mode

Website QR Code

Scan to Visit

Github QR Code

Scan for GitHub