Research

Explainable NLP & interpretability

I build NLP models whose decisions domain experts can inspect and question. This covers post-hoc interpretability methods, token-level explanations, and transparent pipelines for sensitive classification tasks in the social sciences. I also bridge computational methods and social-science questions: social media analysis, online discourse, and tools for social scientists. All of this work follows open-science practice: code, data, and paper materials are released so results can be reproduced and extended.

Cultural fairness & moral reasoning in LLMs

Large language models encode moral judgments, but not everyone’s. I study how LLMs represent cross-cultural variation in moral judgments and societal norms, and how to evaluate their alignment across cultures, including EvalMORAAL, an interpretable chain-of-thought and LLM-as-judge framework for moral alignment.

Human-AI collaboration & annotation reliability

When can we trust an LLM’s annotations and explanations, and when do we still need a human? I compare model outputs against human rationalizations, measure demographic bias in annotation, and study how explanations should be written to match human expectations. A related line uses reinforcement learning to extract decision rules from human evaluations, and I’m also interested in using it for dynamic adaptation from user feedback. I ask the same question of LLM-as-judge evaluation: my open-source audits probe a judge for position, verbosity and self-preference bias, and estimate its accuracy when there are no gold labels.

I am open to research collaborations in explainable NLP, cultural fairness in LLMs, annotation reliability, and applied machine learning for the social sciences. If our interests overlap, email me at hadi.mohammadi@outlook.com.