Urja Khurana

I am a fourth-year PhD student at the Computational Linguistics and Text Mining Lab (CLTL) at the Vrije Universiteit Amsterdam. I am advised by Antske Fokkens (Vrije Universiteit Amsterdam) and Eric Nalisnick (Johns Hopkins University).

I am passionate about understanding what knowledge a language model has captured and whether it can reliably apply this knowledge across diverse (unseen) contexts; what will its impact be in the real world? I am particularly interested in analyzing and developing tools to ensure responsible deployment, with a big focus on robustness for safety-critical applications, e.g. hate speech detection.

My work includes evaluating the generalization of model capabilities to unseen data, for instance through the lens of model averaging and consistency. I have also proposed a method to calibrate language models according to human subjectivity. Collaborating in an interdisciplinary environment with a social scientist and law expert, I characterized the subjective aspects of hate speech and its impact on real-world deployment. Building on these insights, I developed a framework to evaluate if a hate speech detection model’s behavior aligns with the type of hate speech it is intended to address.

I earned my BSc and MSc degrees in Artificial Intelligence from the University of Amsterdam.

If you want to (briefly) know more about my research interests, see the research page.

selected publications.

DefVerify: Do Hate Speech Models Reflect Their Dataset's Definition?

Khurana, Urja, Nalisnick, Eric, and Fokkens, Antske

To appear at COLING 2025

HTML
Crowd-Calibrator: Can Annotator Disagreement Inform Calibration in Subjective Tasks?

Khurana, Urja, Nalisnick, Eric, Fokkens, Antske, and Swayamdipta, Swabha

In First Conference on Language Modeling 2024

HTML
Hate Speech Criteria: A Modular Approach to Task-Specific Hate Speech Definitions

Khurana, Urja, Vermeulen, Ivar, Nalisnick, Eric, Van Noorloos, Marloes, and Fokkens, Antske

In Proceedings of the Sixth Workshop on Online Abuse and Harms (WOAH) 2022

Abs Bib HTML

The subjectivity of automatic hate speech detection makes it a complex task, reflected in different and incomplete definitions in NLP. We present hate speech criteria, developed with insights from a law and social science expert, that help researchers create more explicit definitions and annotation guidelines on five aspects: (1) target groups and (2) dominance, (3) perpetrator characteristics, (4) explicit presence of negative interactions, and the (5) type of consequences/effects. Definitions can be structured so that they cover a more broad or more narrow phenomenon and conscious choices can be made on specifying criteria or leaving them open. We argue that the goal and exact task developers have in mind should determine how the scope of hate speech is defined. We provide an overview of the properties of datasets from hatespeechdata.com that may help select the most suitable dataset for a specific scenario.
@inproceedings{khurana-etal-2022-hate, bibtex_show = {true}, title = {Hate Speech Criteria: A Modular Approach to Task-Specific Hate Speech Definitions}, author = {Khurana, Urja and Vermeulen, Ivar and Nalisnick, Eric and Van Noorloos, Marloes and Fokkens, Antske}, booktitle = {Proceedings of the Sixth Workshop on Online Abuse and Harms (WOAH)}, month = jul, year = {2022}, address = {Seattle, Washington (Hybrid)}, publisher = {Association for Computational Linguistics}, url = {https://aclanthology.org/2022.woah-1.17}, html = {https://aclanthology.org/2022.woah-1.17}, doi = {10.18653/v1/2022.woah-1.17}, pages = {176--191}, selected = {true} }
How Emotionally Stable is {ALBERT}? Testing Robustness with Stochastic Weight Averaging on a Sentiment Analysis Task

Khurana, Urja, Nalisnick, Eric, and Fokkens, Antske

In Proceedings of the 2nd Workshop on Evaluation and Comparison of NLP Systems 2021

Abs Bib HTML

Despite their success, modern language models are fragile. Even small changes in their training pipeline can lead to unexpected results. We study this phenomenon by examining the robustness of ALBERT (Lan et al., 2020) in combination with Stochastic Weight Averaging (SWA)—a cheap way of ensembling—on a sentiment analysis task (SST-2). In particular, we analyze SWA’s stability via CheckList criteria (Ribeiro et al., 2020), examining the agreement on errors made by models differing only in their random seed. We hypothesize that SWA is more stable because it ensembles model snapshots taken along the gradient descent trajectory. We quantify stability by comparing the models’ mistakes with Fleiss’ Kappa (Fleiss, 1971) and overlap ratio scores. We find that SWA reduces error rates in general; yet the models still suffer from their own distinct biases (according to CheckList).
@inproceedings{khurana-etal-2021-emotionally, bibtex_show = {true}, title = {How Emotionally Stable is {ALBERT}? Testing Robustness with Stochastic Weight Averaging on a Sentiment Analysis Task}, author = {Khurana, Urja and Nalisnick, Eric and Fokkens, Antske}, booktitle = {Proceedings of the 2nd Workshop on Evaluation and Comparison of NLP Systems}, month = nov, year = {2021}, address = {Punta Cana, Dominican Republic}, publisher = {Association for Computational Linguistics}, url = {https://aclanthology.org/2021.eval4nlp-1.3}, html = {https://aclanthology.org/2021.eval4nlp-1.3}, selected = {true}, pages = {16--31} }