llm-enhanced-emotion-dynamics-decoding

name	llm-enhanced-emotion-dynamics-decoding
description	LLM-enhanced multi-target regression framework for decoding continuous emotion trajectories from brain fMRI using dynamic functional connectivity
version	1.0.0
author	Neuroscience Cron Job
created	2026-06-09T00:00:00.000Z
arxiv_id	2606.07707v1
paper_title	Decoding Naturalistic Emotion Dynamics from the Brain: An LLM-Enhanced Regression Framework
paper_date	2026-06-05T00:00:00.000Z
activation_keywords	["emotion decoding","LLM annotation","dynamic functional connectivity","multi-target regression","continuous emotion trajectories","affective neuroscience","naturalistic stimuli","graph-theoretical XAI"]

LLM-Enhanced Emotion Dynamics Decoding

概述

该方法论重新构想情感解码任务，从传统的离散单标签分类转变为多目标连续回归框架，追踪多个重叠的情感维度随时间的连续轨迹。核心创新在于利用 LLM 的强大泛化能力，从自然叙事中自动提取细粒度、连续的情感轮廓作为主观情感的代理。

核心方法论

1. LLM 自动标注范式

问题背景：

• 传统情感解码基于离散分类任务，简化了情感的连续、流动、共现特性
• 自然叙事场景下的情感标注成本高昂且主观性强

解决方案：

• 使用 LLM（如 GPT 系列）从自然叙事文本中提取连续情感轮廓
• 情感维度包括：valence（愉悦度）、arousal（唤醒度）、dominance（主导度）等多个维度
• 标注是连续的时间序列而非离散标签

实施步骤：

1. 将叙事文本按时间窗口分段（如每秒或每句）
2. 对每个分段调用 LLM API 进行情感分析
3. 提取多维度情感评分（如 0-1 连续值）
4. 生成时间序列情感轮廓作为训练标签

代码示例：

import openai
import numpy as np

def extract_sentiment_trajectory(text_segments):
    """
    使用 LLM 提取连续情感轨迹
    
    Args:
        text_segments: 时间分段的文本列表
    
    Returns:
        trajectory: 多维度情感评分时间序列 (n_segments, n_dimensions)
    """
    trajectory = []
    dimensions = ['valence', 'arousal', 'dominance']
    
    for segment in text_segments:
        prompt = f"""
        Analyze the emotional content of this text segment.
        Rate each dimension on a scale from 0 to 1:
        - Valence (pleasantness)
        - Arousal (activation level)
        - Dominance (control/power)
        
        Text: "{segment}"
        
        Provide ratings as JSON: {"valence": X, "arousal": Y, "dominance": Z}
        """
        
        response = openai.ChatCompletion.create(
            model="gpt-4",
            messages=[{"role": "user", "content": prompt}]
        )
        
        scores = parse_llm_response(response)
        trajectory.append(scores)
    
    return np.array(trajectory)

2. 动态功能连接表示

：