Improving Text Classification Resilience and Efficiency with RETVec

Google introduces RETVec, a compact, multilingual text vectorizer designed to harden and accelerate text classification against adversarial manipulations; when integrated into Gmail’s spam classifier, RETVec improved spam detection by 38%, reduced false positives by 19.4%, and cut TPU usage by 83%. The open-source model works across all UTF-8 languages without preprocessing and supports deployments via TensorFlow, TFLite (mobile/edge), and TensorFlow.js (web), with research details provided in a NeurIPS 2023 paper and accompanying resources.