LLMs are Vulnerable to Malicious Prompts Disguised as Scientific Language
Published in Under Review, 2025
Neeraja Kirtane*, Yubin Ge, Hao Peng, Dilek Hakkani-Tür
Download here
Publications
FactCheckmate: Preemptively Detecting and Mitigating Hallucinations in LMs
Published in Under Review, 2024
Neeraja Kirtane, Deema Alnuhait, Muhammad Khalifa, Hao Peng
Download here
Efficient Gender Debiasing of Pre-trained Indic Language Models
Published in Deployable AI workshop, AAAI, 2023
Neeraja Kirtane, Aditya Kane, V Manushree
Quantified bias in Hindi Language model- Muril.Efficiently finetuned by unfreezing less than 1 percent of the parameters. Results showed that debiasing reduced the bias.
Download here
Hidden Voices: Reducing gender data gap, one Wikipedia article at a time
Published in Wiki Workshop, 2023
Neeraja Kirtane, Anuraag Shankar, Chelsi Jain, Ganesh Katrapati, Senthamizhan V, Raji Baskaran, Balaraman Ravindran
Wikipedia is the most widely available structured repository of information on the Internet. However, gender disparity has been observed in wiki articles, and it is a major issue. We aim to tackle this problem using Machine Learning methods to generate wiki-like biographies for notable women on Wikipedia. We present Hidden Voices, a project which will assist wiki editors and enthusiasts in writing more biographies about women, thereby increasing their representation on Wikipedia.
Download here
ReGrAt: Regularization in graphs using attention mechanism to handle class imbalance
Published in GCLR workshop, AAAI, 2023
- Neeraja Kirtane, Jeshuren Chelladurai, Balaraman Ravindran, Ashish Tendulkar
Used attention mechanism to tackle imbalance. Used a custom loss function by adding a regularizer that handles imbalance. Got better results than already existing methods.
Download here
Mitigating gender stereotypes in Hindi and Marathi
Published in GeBNLP, NAACL, 2022
- Neeraja Kirtane and Tanvi Anand
Created a dataset of occupations and emotion in Hindi and Marathi. Proposed methods to quantify the bias in the word embeddings. Used existing methods to debias the embeddings.
Download here
Transformer based ensemble for emotion detection
Published in WASSA, ACL, 2022
Aditya Kane, Shantanu Patankar, Neeraja Kirtane, Sahil Khose
Developed ensemble based solution consisting of multiple ELECTRA and BERT models. Proposed methods for synthetically generating datasets to mitigate class imbalance. Studied the behaviour of our models on various raw and synthetically generated datasets.
Download here
Occupational Gender Stereotypes in Indian Languages
Published in WiNLP, EMNLP, 2021
- Neeraja Kirtane and Tanvi Anand
Devised a metric to calculate bias in gendered languages like Hindi and Marathi. Used this metric on ULMFiT language model and quantified the bias present.
Download here