Work

Robust Sentiment Analysis with Minimal Preprocessing: An Adversarial-Contrastive Approach

Year: 2025
Type: dissertation
Abstract: Traditional sentiment analysis systems strongly rely upon detailed preprocessing of texts to handle noise, error, and informality in language, but such preprocessing removes potentially useful linguistic signals while increasing extra computation overhead. This work investigates whether one could merge adversarial training with contrastive learning to create robust systems that process minimally preprocessed texts. This introduced a unifying framework that unites Fast Gradient Sign Method (FGSM) adversarial training with InfoNCE contrastive learning and evaluated with three architectures (ELECTRA, BERT, LSTM) with both formal movie review texts and informal Twitter data. The cohesive framework features an enhanced noise injection system that evaluates nine distinct types of perturbations at intensities ranging from 0% to 80%. The findings indicate that performance correlates with both architecture and domain. The LSTM architecture achieved significant improvements of 10.6% on movie reviews and 7.4% on Twitter data. The transformer architecture achieved modest improvements (BERT: 1.2%, ELECTRA: 2.2%) in formal text, but had inconsistent results on social media. The robust analysis indicates that adversarially trained models perform at 96.3% under harsh noise conditions, whereas baseline models achieve 94.2% performance. The results challenge notions of architectural obsolescence, illustrating that appropriate training methods can rejuvenate simpler approaches. The study provides practical recommendations for developing sentiment analysis systems that can handle spontaneously occurring text with minimal preprocessing, especially advantageous for resource-constrained situations requiring dependable performance with noisy user-generated information.
Source: NORMA
Language: English
Open Access status: green
OpenAlex ID: W7204615171
Cites: 14
Cited by:
Related to: 10