跳转至

DiaRelay：利用恒定大小内存进行对话情感识别的对话上下文传递机制

文章背景与核心概要

对话情感识别（ERC）任务的核心挑战在于捕捉分布在长距离对话轮次中的细微情感线索。传统方法通常依赖于固定的上下文窗口，但这种方式存在两难困境：较短的窗口会丢失关键的长距离信息，而较大的窗口则会引入冗余噪声，并导致计算和内存成本呈指数级增长。此外，现有的参数高效微调方法（如 LoRA）难以显式维护对话级别的状态，也无法有效适应动态演变的对话过程。

为了解决这些挑战，本文提出了 DiaRelay，这是一种轻量级适配器，使大语言模型（LLM）能够显式维护一个恒定大小的对话级内存。DiaRelay 通过“选择性中继内存转换”和“双轴中继内存读取”两个协作组件，在无需重新编码历史对话或扩展上下文长度的情况下，成功传递历史情感证据。实验结果表明，DiaRelay 仅需 7.1M 个额外可训练参数，便在 MELD 数据集上达到了最先进（SOTA）的性能，并在 IEMOCAP 数据集上表现出极强的竞争力。

📋 摘要

Emotion Recognition in Conversation (ERC) often requires detecting subtle emotional cues distributed across distant dialogue turns. Traditional methods rely on fixed context windows: short windows discard valuable long-range information, while larger windows introduce redundant/irrelevant text and exponentially increase computational and memory costs. Furthermore, parameter-efficient adaptation methods like LoRA fail to maintain an explicit dialogue-level state or adapt to evolving conversations.

To resolve these challenges, the authors introduce DiaRelay, a lightweight adapter that enables Large Language Models (LLMs) to maintain a constant-size dialogue-level memory explicitly. Powered by two collaborative components (Selective Relay Memory Transition and Dual-axis Relay Memory Read), DiaRelay successfully carries forward historical evidence without re-encoding past dialogues or expanding context lengths. Experiments show that DiaRelay achieves state-of-the-art (SOTA) performance on MELD (weighted F1 and accuracy) and competitive results on IEMOCAP using only 7.1M additional trainable parameters.

🛠️ DiaRelay 的核心组件

  1. 选择性中继内存转换 (Selective Relay Memory Transition)
  2. 将关键的历史线索逐步聚合到一个有界的、恒定大小的中继内存中。
  3. 在连续的话语预测中传播该内存，使得早期的情感信号即使在离开局部上下文窗口后，仍能对后续预测产生影响。
  1. Selective Relay Memory Transition
  2. Progressively aggregates critical historical cues into a bounded, constant-size relay memory.
  3. Propagates the memory across successive utterance predictions, allowing earlier emotional signals to influence later predictions even after exiting the local context window.
  1. 双轴中继内存读取 (Dual-axis Relay Memory Read)
  2. 利用传播的内存动态调节低秩特征变换。
  3. 在无需测试时梯度更新的情况下，无缝促进上下文相关的表征适配。
  1. Dual-axis Relay Memory Read
  2. Dynamically modulates low-rank feature transformations utilizing the propagated memory.
  3. Facilitates context-dependent representation adaptation seamlessly without requiring test-time gradient updates.

📊 实验结果

  • MELD 数据集： 在加权 F1 分数和准确率上均达到了最先进（SOTA）的性能。
  • IEMOCAP 数据集： 取得了极具竞争力的结果。
  • 效率： 仅引入了 7.1M 个可训练参数，展现了卓越的参数效率以及在大模型情感理解任务中的强大泛化能力。
  • MELD Dataset: Achieved State-of-the-Art (SOTA) performance in both weighted F1 score and accuracy.
  • IEMOCAP Dataset: Achieved highly competitive results.
  • Efficiency: Introduced only 7.1M trainable parameters, demonstrating exceptional parameter efficiency and strong generalizability for LLM-based emotional understanding.

🔗 链接与资源

license icon