Paraphrase Detection with Label Noise Analysis
A controlled study of how injected label noise changes the accuracy and robustness of a paraphrase-detection model.
This research project introduced controlled amounts of incorrect labels into a paraphrase-detection dataset and measured how model quality changed as the noise increased. The goal was to distinguish headline accuracy from actual robustness when training data is imperfect.
The repository contains the experiment pipeline, noise-generation steps, and comparative evaluation used to study that sensitivity.