Towards Efficient FinBERT via Quantization and Coreset for Financial Sentiment Analysis

Résumé

Real-time financial sentiment classification from social media is critical for applications in algorithmic trading, risk assessment, and market surveillance.However, deploying largescale models like FinBERT on edge devices remains impractical due to their high memory and compute demands.Meanwhile, financial text poses unique challenges such as class imbalance, noisy syntax, and temporal drift.We propose a unified framework that jointly applies coreset selection and post-training quantization to achieve scalable and efficient financial NLP.Our method reduces training data by up to 90% through coreset selection and compresses model size by up to 4× via 8-bit quantization, while preserving over 90% of the original classification accuracy on benchmark financial sentiment datasets.This demonstrates the viability of deploying domain-specific NLP models in constrained environments, offering a principled solution for low-latency, resource-efficient financial text processing.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0