Selva Sarmila Kasiviswanathan

Title: Selva Sarmila Kasiviswanathan: Advancing Gender Fairness in Large Language Models

Introduction

Selva Sarmila Kasiviswanathan is an inventor based in Kochi, India, whose work addresses how gender fairness can be assessed in large language models (LLMs). She is associated with Tata Consultancy Services Limited and is named on one patent.

Latest Patents

Her patent, “Counterfactual Based Gender Fairness Algorithm for Large Language Models,” presents a method and system for evaluating gender bias in LLMs. The disclosure notes that existing assessment techniques often focus on male and female genders and may not adequately represent LGBTQ+ identities.

The system receives a context-aware prompt, anonymizes names in its context, and transforms pronouns to generate counterfactual contexts for different gender-pronoun groups. It then queries an LLM with counterfactual prompts, creates an embedding data frame from the responses, and trains a Gaussian Mixture Model. Cluster distances are used with a predefined threshold to assess gender bias.

Career Highlights

Kasiviswanathan’s patent focuses on a structured, counterfactual approach to evaluating gender fairness in language-model outputs. The disclosed method combines prompt transformation, response embeddings, and statistical clustering to support bias assessment.

Collaborations

Kasiviswanathan is associated with Tata Consultancy Services Limited. Bivek Benoy Paul and Ramesh Balaji are identified as her coworkers.

Conclusion

Selva Sarmila Kasiviswanathan’s patent describes a counterfactual-based framework for assessing gender fairness in large language models, with the aim of supporting more inclusive evaluation.