Publications

You can find the updated list of my articles on my Google Scholar profile.

Extracting Structured Labor Market Information from Job Postings with Generative AI

This work, in collaboration with NASWA, leverages generative AI to extract real-time labor market data from online job postings, helping policymakers better understand trends and characteristics like education, remote work, and benefits across occupations. The findings aim to fill data gaps and inform labor, workforce, and economic policy.

Recommended citation: Mark Howison, William O. Ensor, Suraj Maharjan, Rahil Parikh, Srinivasan H. Sengamedu, Paul Daniels, Amber Gaither, Carrie Yeats, Chandan K. Reddy, and Justine S. Hastings. 2024. Extracting Structured Labor Market Information from Job Postings with Generative AI. Digit. Gov.: Res. Pract. Just Accepted (July 2024). https://dl.acm.org/doi/abs/10.1145/3674847

Controlling the Extraction of Memorized Data from Large Language Models via Prompt-Tuning

This work introduces a prompt-tuning method to control memorized content extraction in LLMs, demonstrating both attack and defense strategies. Using GPT-Neo models, they show their attack increases extraction rates by 9.3 percentage points while their defense reduces extraction by up to 97.7% with minimal impact on model utility.

Recommended citation: Mustafa Ozdayi, Charith Peris, Jack FitzGerald, Christophe Dupuy, Jimit Majmudar, Haidar Khan, Rahil Parikh, and Rahul Gupta. 2023. Controlling the Extraction of Memorized Data from Large Language Models via Prompt-Tuning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1512–1521, Toronto, Canada. Association for Computational Linguistics. https://aclanthology.org/2023.acl-short.129/

Impact of Acoustic Event Tagging on Scene Classification in a Multi-task Learning Framework

This work explores the relationship between Acoustic Event Tagging (AET) and Acoustic Scene Classification (ASC) in a multi-task learning framework. Through extensive empirical analysis, we demonstrate that using AET as an auxiliary task improves ASC performance through regularization, regardless of the event types or dataset size.

Recommended citation: Parikh, R., Sundar, H., Sun, M., Wang, C., Matsoukas, S. (2022) Impact of Acoustic Event Tagging on Scene Classification in a Multi-Task Learning Framework. Proc. Interspeech 2022, 4192-4196, doi: 10.21437/Interspeech.2022-10905 https://www.isca-archive.org/interspeech_2022/parikh22_interspeech.pdf

Canary Extraction in Natural Language Understanding Models

In this work we demonstrate a white-box model inversion attack on Natural Language Understanding models. We show that an adversary can obtain sensitive information from the training data if given access to the model parameters.

Recommended citation: Rahil Parikh, Christophe Dupuy, and Rahul Gupta. 2022. Canary Extraction in Natural Language Understanding Models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pages 552–560, Dublin, Ireland. Association for Computational Linguistics. https://aclanthology.org/2022.acl-short.61.pdf

Harmonicity Plays a Critical Role in DNN Based Versus in Biologically-Inspired Monaural Speech Segregation Systems

In this work we demonstrate that deep neural network based end-to-end speech segregation models cue on to the harmonic structure of speech for grouping and segregating sources. We demonstrate that these networks completely fail to separate inharmonic sources, and that they are unable to learn how to segregate speech when trained on mixtures of inharmonic speech.

Recommended citation: R. Parikh, I. Kavalerov, C. Espy-Wilson and S. Shamma, "Harmonicity Plays a Critical Role in DNN Based Versus in Biologically-Inspired Monaural Speech Segregation Systems," ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Singapore, Singapore, 2022, pp. 536-540, doi: 10.1109/ICASSP43922.2022.9747314. https://ieeexplore.ieee.org/abstract/document/9747314

Acoustic To Articulatory Speech Inversion Using Multi-Resolution Spectro-Temporal Representations Of Speech Signals

In this work we develop a speech inversion system to predict vocal tract parameters using the cortical features of acoustic speech. We demonstrate that the cortical features are correlated to the vocal tract parameters highlighting that the audiotry theory of speech perception is linked to the motor theory of speech production.

Recommended citation: Parikh, R., Seneviratne, N., Sivaraman, G., Shamma, S., Espy-Wilson, C. (2022) Acoustic To Articulatory Speech Inversion Using Multi-Resolution Spectro-Temporal Representations Of Speech Signals. Proc. Interspeech 2022, 4681-4685, doi: 10.21437/Interspeech.2022-10926 https://www.isca-archive.org/interspeech_2022/parikh22b_interspeech.pdf