Portfolio item number 1
Short description of portfolio item number 1
Posts by Collection
portfolio
Portfolio item number 2
Short description of portfolio item number 2
publications
Penta ML at EXIST 2024: Tagging Sexism in Online Multimodal Content With Attention-enhanced Modal Context
Published in EXIST, 2024
Authors:Deeparghya Dutta Barua, Md Sakib Ul Rahman Sourove, Fabiha Haider, Fariha Tanjim Shifat, Md Farhan Ishmam, Md Fahim and Farhad Alam Bhuiyan
Effective content moderation requires automated systems that understand nuanced, multimodal, and multilingual online content. This study explores an attention-based approach to better balance textual and visual cues, improving performance in identifying sexism and other complex labels.
Penta-nlp at EXIST 2024 Task 1-3: Sexism Identification, Source Intention, Sexism Categorization In Tweets
Published in EXIST, 2024
Authors:Fariha Tanjim Shifat, Fabiha Haider, Md Sakib Ul Rahman Sourove, Deeparghya Dutta Barua, Md Farhan Ishmam, Md Fahim and Farhad Alam Bhuiyan
Automated detection of sexist content is vital to curb harmful stereotypes on social media. This study explores various NLP techniques and attention mechanisms to identify and flag sexist tweets, using a large multilingual dataset to improve moderation effectiveness at scale.
BanglaTLit: A Benchmark Dataset for Back-Transliteration of Romanized Bangla
Published in EMNLP Findings, 2024
Authors:Md Fahim, Fariha Tanjim Shifat, Fabiha Haider, Deeparghya Dutta Barua, Md Sakib Ul Rahman Sourove, Md Farhan Ishmam, Md Farhad Alam
To tackle Romanized Bangla (commonly known as Banglish), the most widely used text form on social media, we introduce BanglaTLit and BanglaTLit-PT — two large-scale datasets that fuel state-of-the-art performance in Banglish NLP. Our novel back-transliteration techniques and further-pretrained encoders unlock the full potential of this noisy yet rich linguistic resource.
BanTH: A Multi-label Hate Speech Detection Dataset for Transliterated Bangla
Published in NAACL Findings, 2025
Authors:Fabiha Haider, Fariha Tanjim Shifat, Md Farhan Ishmam, Deeparghya Dutta Barua, Md Sakib Ul Rahman Sourove, Md Fahim, Md Farhad Alam
As transliterated Bangla (Banglish) is one of the most common forms of online communication, detecting hate content in this form is crucial. We present BANTH, the first multi-label transliterated Bangla hate speech dataset. We also propose a novel translation-based prompting strategy with LLMs achieving superior performance over existing baselines and paving the way for more effective hate speech detection in underrepresented languages.
ChitroJera: A Regionally Relevant Visual Question Answering Dataset for Bangla
Published in ECML PKDD, 2025
Authors:Deeparghya Dutta Barua, Md Sakib Ul Rahman Sourove, Md Fahim, Fabiha Haider, Fariha Tanjim Shifat, Md Tasmim Rahman Adib, Anam Borhan Uddin, Md Farhan Ishmam, Md Farhad Alam
In this research project, we explored the performance of Large Vision-Language Models on Bangla Visual Question Answering (VQA) tasks. Due to the lack of existing datasets in this domain, we developed a novel dataset specifically tailored for Bangla VQA. We conducted a series of experiments from multiple perspectives to effectively address our research objectives.
Robustness of LLMs to Transliteration Perturbations in Bangla
Published in BLP-Workshop, 2025
Authors:Fabiha Haider, Md Farhan Ishmam, Fariha Tanjim Shifat, Md Tasmim Rahman Adib, Md Fahim, Md Farhad Alam Bhuiyan
Bangla text online often mixes Bangla script with Romanized Bangla. To handle this, language models need to be able to understand this script mixing. In this work, we test how well current LLMs and Bangla models handle different types of transliteration-based changes. We create mixed-script versions of Bangla data by replacing some words, sentences, or important keywords with their transliterated forms. Our experiments show weaknesses and interesting behaviors in these models, highlighting the need to improve robustness for real-world use.
BANGLAPROTHA: Evaluating Vision Language Models in Underrepresented Long-tail Cultural Contexts
Published in WACV 2026 Conference, 2025
Authors:Md Fahim, Md Sakib Ul Rahman Sourove, Akm Moshiur Rahman Mazumder, Md Farhan Ishmam, Md Tasmim Rahman Adib, Fariha Tanjim Shifat, Fabiha Haider, Md Farhad Alam Bhuiyan
We introduce BanglaProtha, a Bengali culture–focused VQA dataset designed to expose cultural biases in Vision-Language Models. By evaluating multiple VLM classes across prompting, fine-tuning, model sizes, and augmentation strategies, we uncover systematic Western-centric tendencies. Our work offers diagnostic insights for developing fairer multicultural and multilingual AI systems.
talks
From Prompt to Product: A Workshop On The Working Principles of LLM Permalink
Published:
A workshop on the working principles of modern LLMs held at the University of Dhaka, organized by IEEE SBC DU Branch. In it, we talked at length how LLMs work, from explaining about transformers to the process of pre-training and finetuning. We also had a hands-on session where participants were able to implement these same ideas into code.
Static APIBOM & AuthZ Analysis: Beyond Regex-Based SAST Permalink
Published:
Talked about Deep Scan engine and API Call Graph generation, showcasing how multi-agent system is used to uncover zero-day Broken Access Control and Business Logic flaws that traditional scanning often misses.
teaching
Teaching experience 1
Undergraduate course, University 1, Department, 2014
This is a description of a teaching experience. You can use markdown like any other post.
Teaching experience 2
Workshop, University 1, Department, 2015
This is a description of a teaching experience. You can use markdown like any other post.