An Interpretability-Guided Framework for Responsible Synthetic Data Generation in Emotional Text

2025-11-20

AI Summary

High annotation costs and platform constraints hinder the acquisition of high-quality labeled data for emotion recognition in textual data. Method: This paper proposes an explainability-guided synthetic data generation framework that integrates SHAP (Shapley Additive Explanations) with large language models (LLMs). Leveraging a small set of seed samples, it directs LLMs to generate representative emotional texts that preserve lexical diversity, individual expression styles, and temporal dynamics. Contributions/Results: (1) First application of SHAP values to guide LLM-based synthetic data generation, improving coverage and discriminability of rare emotion classes; (2) Classifiers trained on synthetic data achieve performance on par with those trained on real data—significantly outperforming naive generative baselines—despite limited seed data; (3) Dual validation via linguistic feature analysis and classification evaluation reveals inherent limitations of synthetic data in capturing expressive complexity and partially bridges this gap.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Large Multimodal Models (LMMs)
• Humans and AI: Emotional Intelligence

Application Category

• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems

Abstract

Emotion recognition from social media is critical for understanding public sentiment, but accessing training data has become prohibitively expensive due to escalating API costs and platform restrictions. We introduce an interpretability-guided framework where Shapley Additive Explanations (SHAP) provide principled guidance for LLM-based synthetic data generation. With sufficient seed data, SHAP-guided approach matches real data performance, significantly outperforms naïve generation, and substantially improves classification for underrepresented emotion classes. However, our linguistic analysis reveals that synthetic text exhibits reduced vocabulary richness and fewer personal or temporally complex expressions than authentic posts. This work provides both a practical framework for responsible synthetic data generation and a critical perspective on its limitations, underscoring that the future of trustworthy AI depends on navigating the trade-offs between synthetic utility and real-world authenticity.

Problem

• Addresses expensive training data access for emotion recognition from social media
• Proposes interpretability-guided synthetic data generation using SHAP explanations
• Investigates trade-offs between synthetic utility and real-world authenticity

Innovation

• SHAP guides LLM-based synthetic data generation
• Framework matches real data performance with seed data
• Improves classification for underrepresented emotion classes