NLP Research Intern | LLM Evaluation for Patient Narratives

KaLC Lab - University of Tsukuba, Ibaraki, Japan

Studied emotion-cause extraction from patient narratives, with potential applications in understanding and monitoring patients’ experiences during treatment. The work also aimed to establish a standardized dataset and evaluation framework for research.

• Built a privacy-aware collection and preprocessing pipeline for 18 breast cancer patient blogs and 2,500+ posts across two platforms, following scraping policies.
• Iteratively refined emotion-cause annotation guidelines across three versions and curated a 147-pair reference set.
• Designed prompting strategies and evaluated zero-shot emotion-cause extraction with GPT-OSS, Qwen 3.6, and Gemma 4, using a proprietary model as a performance reference.
• Developed an automated evaluation framework combining ROUGE-L, sentence-embedding cosine similarity, and Hungarian matching.
• Conducted systematic error analysis to compare model performance and identify extraction limitations.

Data Science Intern | ML for Carbon-Footprint Estimation

LIP6 - Sorbonne University, Paris, France

In collaboration with Association pour la transition Bas Carbone (ABC), explored explainable machine learning to estimate corporate carbon footprints and help companies better understand their emissions.

• Cleaned, preprocessed, and analyzed 2,029 corporate carbon-footprint records across 137 features, including missing-value handling, statistical analysis, and visualization.
• Developed and compared classification, regression, and Bayesian-network models across feature-encoding strategies, evaluating performance using task-appropriate metrics.
• Applied Bayesian networks and introduced SHALL analysis to detect outliers and generate exploratory data, contributing to the open-source library PyAgrum.
• Delivered structured technical reports and presented findings to 10+ domain experts.