Beyond Sentiment Classification: A Generative Framework for Emotion Intensity Evaluation in Text

Battle Count: 65

The paper explicitly motivates its work for financial applications, citing behavioral finance literature (Griffith et al., 2020; Shen et al., 2023; Taffler et al., 2024; Goetzmann et al., 2024). The authors argue that quantifying the degree of emotional content (e.g., fear magnitude) is more useful than binary emotion detection for financial decision-making. However, the paper does not directly test on trading signals, market data, or portfolio performance. The framework could serve as an input feature for sentiment-based trading strategies, risk models, or behavioral finance research, but empirical validation in trading contexts is left for future work.

The methodology uses standard LoRA fine-tuning with 4-bit quantization on open-weight Mistral models, which is well-supported by the Hugging Face ecosystem. Training requires 3x NVIDIA A100 80GB GPUs for the 24B model. The prompt format and JSON output parsing are straightforward. However, constructing the annotated dataset requires domain expertise in affective psychology, and the evaluation metrics (CCC, zero-match F1) require careful implementation. The leave-one-out and transfer experiments add complexity to the evaluation pipeline.

The paper provides detailed prompt templates (Appendix B), annotation rubric (Appendix A), model configurations (LoRA rank, alpha, dropout, quantization), training hyperparameters (learning rate, batch size, epochs, optimizer), and evaluation metrics. However, the dataset (1,177 samples) is self-constructed and its public availability is not explicitly stated. No GitHub repository is mentioned. The use of open-weight Mistral models aids reproducibility, but the small custom dataset and specific annotation process may limit exact replication.