Enhancing Financial Sentiment Classification via Fine Grained Emotion Modeling using ML, DL, and Attention Mechanism

Abstract

Financial sentiment analysis is essential for investment decisions and market prediction. However, traditional methods that classify text into positive, negative, or neutral often fail to capture nuanced emotions that drive investor behavior. To address these limitations, this thesis introduces an innovative hybrid framework that incorporates fine-grained emotion modeling to enhance financial sentiment classification. The core methodology is structured around three complementary approaches, beginning with Approach 1 (Traditional Financial Sentiment Representations), which utilizes FinBERT through both feature extraction and fine-tuning strategies, alongside general-domain models(RoBERTa, XLM-RoBERTa, and DeBERTa), to capture deep contextual financial semantics. This is complemented by Approach 2 (Emotion-Based Representations), which extracts fine-grained emotion vectors using a RoBERTa model pre-trained on the GoEmotions dataset to model underlying textual affect. Finally, Approach 3 (Hybrid Feature Fusion Strategies) integrates the deep contextual representations from the financial models with the fine-grained emotion vectors. The proposed framework was evaluated across three benchmark financial datasets: Financial PhraseBank (100% and 50%), Twitter Financial News, and the Gold dataset. Classification tasks were executed using traditional Machine Learning algorithms, Deep Learning architectures, and attention mechanisms, followed by a comprehensive comparative analysis between the three developed approaches. Results demonstrated that incorporating fine-grained emotional dimensions substantially enhanced classification accuracy and robustness, bridging the gap between financial textual semantics and investor sentiment dynamics, a benefit most clearly reflected within Approach 3, where Fine-Tuning (FT) consistently surpassed Feature Extraction (FE) across all backbone models. The fusion of RoBERTa with RoBERTa-GoEmotion embeddings delivered the strongest results on PhraseBank 100% (LightGBM, F1-W: 99.55%), PhraseBank 50% (XGBoost, F1-W: 92.85%), Gold (BiLSTM, F1-W: 96.59%), and Twitter (LightGBM, F1-W: 92.92%), establishing RoBERTa as the most consistently effective backbone for emotion-augmented financial sentiment classification