Mitigating Subjectivity and Annotation Inconsistencies in Sentiment Analysis via an SVM-RoBERTa Ensemble

Title

Mitigating Subjectivity and Annotation Inconsistencies in Sentiment Analysis via an SVM-RoBERTa Ensemble

Creator

Roy, Alphin; Roy, Apash; Shrivallabha, S.

Description

This research addresses a main limitation in the Natural Language Processing that is the impact of subjectivity and annotation inconsistencies on the accuracy of the sentiment classification. We did a systematic comparison of two fundamentally different architectures. A traditional feature based Support Vector Machine and a deep contextual fine tuned RoBERTa transformer using a challenging, noisy, real-world Twitter dataset. This corpus retains ambiguity and sarcasm on purpose and serve the crucible for testing model robustness. We developed a soft voting ensemble method that combines the probability scores from both models to obtain the best predictive capabilities. The results showed a clear technological hierarchy. The RoBERTa model with its deep semantic grasp outperformed the SVM by a substantial margin achieving 90% accuracy against 83.5% accuracy. But the hybrid ensemble model attained the highest overall accuracy of 91.35% and showed better reliability across all the sentiment classes. These findings shows that a hybrid approach fusing a transformer's nuanced understanding with the stabilization provided by ensemble learning is the most effective and robust method for mitigating data imperfections in modern sentiment analysis. 2025 IEEE.

Source

IC-DECON 2025 - 2025 International Conference on Data, Energy and Communication Network, Proceedings;

Date

01-01-2025

Publisher

Institute of Electrical and Electronics Engineers Inc.

Subject

Annotation Inconsistency; Ensemble Learning; RoBERTa; Sentiment Analysis; Support Vector Machine

Coverage

Roy A., Christ (Deemed to be University), Dept. of Statistics and Data Science, Bangalore, India; Roy A., Christ (Deemed to be University), Dept. of Statistics and Data Science, Bangalore, India; Shrivallabha S., Christ (Deemed to be University), Dept. of Statistics and Data Science, Bangalore, India

Rights

Restricted Access; Hardcopy may be available in the library

Relation

ISBN: 979-833159442-8;

Format

online

Language

English

Type

Conference paper

Identifier

https://doi.org/10.1109/DECoN67170.2025.11447642

https://www.scopus.com/pages/publications/105037371064?origin=resultslist