Research

Publications & Research

Peer-reviewed contributions to multilingual NLP, toxic-content classification, and fairness in machine learning.

Peer-Reviewed Publications

Urdu Toxic Comment Classification with PURUTT Corpus Development

Saeed, Hafiz Hassaan, et al. — IEEE Access 13 (2025): 21635–21651

DOI: 10.1109/ACCESS.2025.3535862

Roman Urdu Toxic Comment Classification

Saeed, Hafiz Hassaan, et al. — Language Resources and Evaluation 55.4 (2021): 971–996

DOI: 10.1007/s10579-021-09530-y

OSACT4 Shared Tasks: Ensembled Stacked Classification for Offensive and Hate Speech in Arabic Tweets

Saeed, Hafiz Hassaan, et al. — Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools (2020)

aclanthology.org/2020.osact-1.11

Overlapping Toxic Sentiment Classification using Deep Neural Architectures

Saeed, Hafiz Hassaan, et al. — IEEE International Conference on Data Mining Workshops (ICDMW), 2018

DOI: 10.1109/ICDMW.2018.00193

Trained word embedding models (Word2Vec, GloVe, FastText) alongside classical ML (Naïve Bayes, Random Forest, Logistic Regression, SVM) and deep learning architectures (CNN, BLSTM, BGRU) using scikit-learn, Keras, and TensorFlow with stratified 5-fold cross-validation.
Fine-tuned five multilingual transformer models (multilingual BERT, multilingual DistilBERT, XLM-RoBERTa, MuRIL, multilingual T5) using Hugging Face with stratified 5-fold cross-validation.

Implemented deep learning architectures from scratch in Python, Keras, and TensorFlow for both Arabic and English toxic comment classification.
Engineered a dual-embedding representation combining two distinct word embeddings per word, with domain-specific preprocessing for Arabic.

Developed an Urdu NER system with LSTM/GRU architectures and training and inference pipelines.

Conducted exploratory data analysis on the prevalence of Ischemic Heart Disease in Pakistan using Python, Pandas, and Matplotlib, developing visualizations to surface key contributing factors.
Applied hypothesis testing and correlation analysis to identify statistically significant factors associated with the disease.

Implemented post-processing fairness techniques to reduce bias and discrimination in machine learning models using the Adult and German Credit datasets.