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

2026-05-19 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Advanced, extended

Summary

A new generative framework for emotion intensity evaluation in text has been introduced, shifting from discrete emotion classification to continuous scoring. Researchers constructed a dataset of 1,177 conversational phrases, human-annotated with 0-100 intensity scores for eight emotions, plus valence (-100 to 100) and arousal (0 to 100). Open-weight generative language models, specifically Mistral-7B and Mistral-24B, were fine-tuned using LoRA to output these continuous values in a JSON format. This approach significantly outperforms classification baselines and zero-shot pretrained LLMs, demonstrating strong generalization capabilities to unseen emotions and transfer effects to valence and arousal. The framework aims to better align with applied domains like finance, where the degree of emotional content is crucial for interpretation and decision-making.

Key takeaway

For NLP engineers and research scientists developing affective computing systems, consider adopting generative models for emotion intensity evaluation instead of traditional classification. This approach provides more granular, interpretable emotional profiles, crucial for applications in finance or behavioral science. Your models will likely generalize better to new emotions and abstract affective dimensions like valence and arousal, enhancing their utility and adaptability in dynamic real-world scenarios.

Key insights

Generative LLMs fine-tuned on continuous emotion intensity scores outperform classification and generalize to unseen affective dimensions.

Principles

• Emotion intensity evaluation is more expressive than discrete classification.
• Generative LLMs can infer latent affective structure.
• Model scale enhances generalization to new affective constructs.

Method

Fine-tune open-weight generative LLMs (Mistral-7B, Mistral-24B) via LoRA on a human-annotated dataset of 0-100 emotion intensity, valence, and arousal scores, outputting results in JSON.

In practice

• Use fine-tuned generative LLMs for nuanced emotion analysis.
• Apply 4-bit quantization for efficient fine-tuning on GPUs.
• Evaluate models using Concordance Correlation Coefficient (CCC).

Topics

• Generative Language Models
• Emotion Intensity Evaluation
• Affective Computing
• Valence-Arousal Model
• LoRA Fine-tuning

Best for: AI Engineer, Machine Learning Engineer, AI Product Manager, AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• From Affect Prediction to Affect Forecasting: Evidence for Distinct Information Sources in Longitudinal Text — Current affect prediction relies on text semantics, while future affective change forecasting benefits from numeric trajectory dynamics.
• Proposing Topic Models and Evaluation Frameworks for Analyzing Associations with External Outcomes: An Application to Leadership Analysis Using Large-Scale Corporate Review Data — LLM-enhanced topic modeling improves interpretability, specificity, and polarity consistency for outcome-oriented text analysis.
• A Scalable Tool for Measuring Manner and Result Verbs in Developmental Language Research — A RoBERTa classifier, trained on LLM-generated annotations, accurately distinguishes manner and result verbs at scale.
• Do We Still Need Humans in the Loop? Comparing Human and LLM Annotation in Active Learning for Hostility Detection — LLM labels can match human-level F1 scores for hostility detection but exhibit distinct error profiles.
• Does Topic Sentiment Cause Perceived Ideology? Comparing Human and LLM Annotations in Political News Articles — Fine-tuned LLMs can develop spurious sentiment-ideology couplings, diverging from human judgment in political news annotation.
• Emergence of Hierarchical Emotion Organization in Large Language Models — LLMs spontaneously organize emotions hierarchically, mirroring human psychology, but also replicate human biases in emotion recognition.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.