Auditing the Space Between Languages: Character Representation in Bilingual Young Adult Literature
Tanay Nagar, Pragati Maheshwary, Shamya Karumbaiah
FAccT '26: Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency · June 25–28, 2026, Montréal, QC, Canada.
We audit 70 English–Spanish YA novels, turning 619 extracted characters into contextual embeddings and projecting them onto 25 theory-grounded sociocultural axes (warmth, literacy, status, and more). Separating casting (who gets associated with each language) from framing (how that language shifts portrayal), we find that Spanish-associated characters occupy a contradictory semantic space, and code-mixed text behaves as its own distinct register, not a midpoint between its two languages.
Exploring Bias in Letters of Recommendation using NLP Techniques
Tanay Nagar, Sarah Jung, Peter Wirth, Alyssa Schappe, Amorn N. Salyapongse
Oral presentation: ACEPS 2025 · Extended abstract published in PRS Global Open 2025.
Hina and Faisal Mushtaq Scholarship
We use natural language processing to analyze plastic surgery residency letters of recommendation, surfacing systematic patterns in wording that may reflect bias and influence selection decisions. The work highlights how computational text analysis can support fairer, more transparent evaluation of applicants.
Breaking Language Barriers: Equitable Performance in Multilingual Language Models
Tanay Nagar, Grigorii Khvatskii, Anna Sokol, Nitesh V. Chawla
Poster: NAACL 2025 Student Research Workshop (SRW); ND Summer Research Symposium 2024 · Full paper on arXiv 2025.
Bromley Conference Travel Award
CDIS Student Travel Award
We fine-tune a multilingual language model on synthetic code-switched text at controlled mixing ratios and study how those ratios affect performance on both low- and high-resource languages. We find that the ratio matters: a medium ratio optimally closes ~80% of the Hindi–English reasoning gap on CommonsenseQA without degrading English performance.
Decoding Bias in Letters of Recommendation: A Word Embeddings Approach
Tanay Nagar
Senior Honors Thesis, University of Wisconsin-Madison, 2025 · Oral presentation: UW Honors Research Symposium 2025.
Advisors:
Dr. Fred Sala, Dr. Sarah Jung, Dr. Shamya Karumbaiah
Hilldale Undergraduate/Faculty Research Fellowship
Richard Ralph Phoenix Rising Humanitarian Scholarship
Using 1,585 de-identified letters of recommendation (~919k tokens), my thesis combines Word2Vec and fine-tuned Mistral-7B embeddings to measure gender-directional framing via bias axes, WEAT-style tests, and analogy probes. It turns long-standing concerns about biased LoR language into quantitative metrics that can be used for reviewer calibration, dashboards, and faculty workshops to support fairer admissions and evaluations.
TRAIL Lab