Reproduce “Exploring Persona Sentiment Sensitivity in Personalized Dialogue Generation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yonghyun Jun, Hwanhee Lee

Researchers from Chung-Ang University revealed that Large Language Models exhibit sensitivity to the sentimental polarity of assigned user personas, finding that negative and neutral sentiments often lead to reduced dialogue coherence and consistency. A proposed approach combining turn-based generation, specific persona ordering, and sentiment-aware prompting was shown to enhance dialogue quality for challenging persona types.

Key insights

  • LLMs demonstrate a clear sensitivity to user persona sentiment, with highly positive personas yielding more consistent and coherent dialogues, while negative or neutral personas often lead to overemphasized attributes, contradictions, or reduced integration.
  • Dialogue quality, in terms of both consistency and coherence, often follows a U-shaped trend with persona polarity level; dialogues from highly polarized (positive or negative) personas tend to outperform neutral ones, with positive sentiment consistently superior.
  • A combined strategy of turn-based generation, 'center-out score ascending' persona ordering, and sentiment-aware prompting effectively improves dialogue quality, particularly enhancing consistency and coherence for challenging negative and neutral persona sentiments.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.