Sk Mainul Islam
PhD Fellow · NLP & Interpretability · Copenhagen

I am a second-year PhD fellow at the CopeNLU group, University of Copenhagen, supervised by Prof. Isabelle Augenstein and Asst. Prof. Pepa Atanasova. My research sits at the intersection of explainability, fact-checking, and robustness in large language models — with a particular focus on interpretable reasoning for misinformation detection and trustworthy NLP systems.

Before Copenhagen, I was a Research Engineer at the Center for Cyber Security, NYU Abu Dhabi (2022–2024). I hold an MS by Research from IIT Kharagpur, where my thesis developed scalable, explainable methods for Aspect-Based Sentiment Analysis.

I am open to academic visiting positions and research scientist internships in industry for 2026.

Portrait of Sk Mainul Islam
Research

My work focuses on making NLP models more interpretable, robust, and trustworthy — especially in high-stakes scenarios like automated fact-checking and misinformation detection. Current threads include understanding in-context learning mechanisms in LLMs, compositional proof reasoning on out-of-distribution data, knowledge probing, and constrained/controlled text generation for counter-narrative and emotional response tasks.

Interpretability & Explainability Fact-checking Robustness of LLMs In-context learning Knowledge probing Misinformation & propaganda Constrained text generation Multimodal NLP Aspect-based sentiment analysis
News & Highlights
Publications
AR-BERT architecture diagram
AR-BERT: Aspect-relation enhanced Aspect-level Sentiment Classification with Multi-modal Explanations
Sk Mainul Islam, Sourangshu Bhattacharya
The Web Conference (WWW), 2022

Proposes a two-level global-local entity embedding scheme for joint training of knowledge-graph-based aspect embeddings and ALSC models, along with a novel incorrect disambiguation detection technique. Achieves a consistent 2.5–4.1 point improvement over BERT-based baselines on benchmark datasets, alongside multi-modal explanation generation.

PDF
Fair data generation diagram
Fair Data Generation using Language Models with Hard Constraints
Sk Mainul Islam, Abhinav Nagpal, Balaji Ganesan, Pranay Kumar Lohia
CtrlGen Workshop @ NeurIPS, 2021

Introduces new methods to inject hard constraints and domain knowledge into pre-trained language models for generating fair and privacy-respecting synthetic text data, addressing ethical concerns in masked entity prediction tasks.

PDF