Najrin Sultana

Bangladesh University of Engineering and Technology

CGPA: 3.90 / 4.00

BanglaParaphrase: A High-Quality Bangla Paraphrase Dataset

• In this work, we present BanglaParaphrase, a high-quality synthetic Bangla Paraphrase dataset curatedby a novel filtering pipeline.
• We aim to take a step towards alleviating the low resource status of the Bangla language in the NLP domain through the introduction of BanglaParaphrase, which ensures quality by preserving both semantics and diversity, making it particularly useful to enhance other Bangla datasets.
• We show a detailed comparative analysis between our dataset and models trained on it with other existing works to establish the viability of our synthetic paraphrase data generation pipeline.

Unsupervised Bangla Text Simplification

• Using a diverse paraprase generating model, we attempted to simplify complex Bangla texts in an unsupervised way by leveraging style transfer mechanisms.
• The available non-parallel labelled data was insufficient to fine-tune a text generation model. So we fine-tuned BanglaBERT’s token classifier with the data and labelled more data to use them in the simplification task.
• The final outcome looks promising. The output looks simplified by manual observation. But the accuracy is lower than expectation which is determined by the result of the classifier. The main reason we identified is the noise in the initial labelled data which was used to train the classifier.
• We plan to work on this further by gathering more cleaner labelled data and hope to achieve much better results.

Butterfly species classification & similar image recommender

• Trained different models on Butterfly Image Classification 75 species dataset of Kaggle using tensorflow and pytorch.
• Also designed a recommendation system.

CNN from scratch

• Implemented a convolutional neural network from scratch with python for an image classification task with support for variable number of layers, dimensions as well as batching.

Logistic Regression and AdaBoost for Classification

• Implemented Logistics Regression and adaboost algorithms from scratch with python and tested on 3 datasets from kaggle.

Hidden Markov Models

• Hidden Markov Model (HMM) implemented with the Viterbi algorithm.
• The parameters of the model was estimated with Baulm-Welch Algorithm.

Interactive Code Learning Platform

• Built a platform where students can learn in an interactive manner.
• In this platform, there are games, quizes, videos and other possible ways of interaction to keep the attention of the user.

Classical AI Projects

• AI-game-LinesOfAction: Used minimax algorithm with alpha-beta pruning with suitable heuristics to create and agent which is able to play the board game. Can also be played in multiplayer mode.
• Ghost Hunting: Modeled Ghost Hunting problem using HMM and Particle Filtering.
• Latin Square CSP Solver: Used several CSP solving algorithms with heuristics to solve latin square.
• Optimal Exam Scheduling with local search: Exam scheduling problem was solved by local search with appropriate heuristics.
• Solving n-puzzle by A-star search: Solved n-puzzle problem with A-star search by using two heuristics (the misplaced tiles an the Manhattan distance).

Security Tools

• AES standard implementation: An implementation of AES by python from scratch. This tool can encrypt as well as decrypt texts and files of any kind.
• Cross-Site Scripting (XSS) Attack: Used a web application named Elgg provided in a pre-built Ubuntu VM image and exploited XSS vulnerability to launch an XSS attack on the modified Elgg.
• DHCP Starvation Attack: Implemented and demonstrated DHCP Starvation Attack. The script was developed with C and a video demonstration of the attack is also available.