Beyond Prompt-Sensitive Emotion Words: Stable Embeddings for Tang Poetry Analysis

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Science & Research — Social Sciences & Behavioral Studies, Research Methodology & Innovation · Depth: Expert, medium

Summary

The paper "Beyond Prompt-Sensitive Emotion Words: Stable Embeddings for Tang Poetry Analysis" by Linyue Zhang and Feiyue Li proposes a fine-grained sentence-level workflow for analyzing emotions in Tang poetry, specifically around the An Lushan Rebellion. It addresses the issue of prompt-sensitive one-word emotion outputs from LLMs, which show only 50.3% A/B exact agreement on 3,198 emotional sentences. The proposed method uses continuous hidden-state vectors, automatic clustering, and label consolidation, achieving stable and well-distributed results ($H_{norm}=0.989$; 20/20 active clusters). This embedding-based approach supports historically grounded findings, like an emotional turning point around 762, and reveals layered emotional patterns. A char-based baseline on 7,195 labeled sentences achieved 0.446 micro-F1 and 0.395 macro-F1. The authors argue that stable representation is crucial for credible computational humanities interpretation.

Key takeaway

For digital humanists or AI scientists analyzing historical texts for nuanced emotional patterns, relying solely on prompt-sensitive LLM outputs for emotion labeling risks irreproducible and untrustworthy interpretations. Instead, consider implementing embedding-based clustering workflows using continuous hidden-state vectors to achieve stable, fine-grained emotional representations. This approach provides more credible evidence for historical findings, such as identifying specific emotional turning points, and uncovers deeper, layered patterns.

Key insights

Stable emotion embeddings, not prompt-sensitive LLM outputs, are essential for credible, reproducible fine-grained historical text analysis.

Principles

• Prompt-sensitive LLM outputs hinder reproducible interpretation.
• Stable representations are prerequisite for credible computational humanities.
• Fine-grained analysis reveals layered patterns missed by coarse labels.

Method

A fine-grained sentence-level workflow uses continuous hidden-state vectors, automatic clustering, and label consolidation for emotion interpretation.

In practice

• Analyze Tang poetry emotions around An Lushan Rebellion.
• Identify emotional turning points in historical texts.
• Reveal layered emotional patterns in literary works.

Topics

• Tang Poetry Analysis
• Emotion Embeddings
• Natural Language Processing
• Digital Humanities
• LLM Prompt Sensitivity
• Historical Text Analysis
• An Lushan Rebellion

Best for: Research Scientist, AI Scientist

Related on AIssential

• Emotion-Aware Image Generation from Korean Diary Text via LLM-based Prompt Translation and LoRA Fine-Tuning — A pipeline generates emotion-aware, children's drawing style images from Korean diary text using LLM sentiment analysis and LoRA fine-tuned Stable Diffusion.
• Predicting Poets' Origins from Verse: A Computational Analysis of Regional Linguistic Fingerprints in the Complete Tang Poems — Tang poets' geographic origins leave detectable linguistic traces in their verse, revealing historical regional divergence.
• SubtleMemory: A Benchmark for Fine-Grained Relational Memory Discrimination in Long-Horizon AI Agents — AI agents struggle to discriminate subtle, context-dependent relations (complementary, nuanced, contradictory) among long-term memories.
• Training Data Size Sensitivity in Unsupervised Rhyme Recognition — RhymeTagger excels at unsupervised rhyme recognition, outperforming humans and LLMs when adequately trained.
• Statistical Embeddings for Similarity, Retrieval, and Interpretable Alignment of Numeric Tabular Datasets — Statistical embeddings enable interpretable similarity and alignment of heterogeneous numeric tabular datasets for LLMs, preserving statistical context.
• Audience Engagement with Arabic Women's Social Empowerment and Wellbeing: A Decadal Corpus — The corpus provides a decadal view of Arabic women's social empowerment discourse on Facebook.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.