PaperlayerView exact revision on arXiv

Source-linked AI summary

Do LLMs Feel? Teaching Emotion Recognition with Prompts, Retrieval, and Curriculum Learning

Xinran Li, Yu Liu, Jiaqi Qiao, Xiujuan Xu

arXiv:2511.07061v3cs.AI

TL;DR

ERC models have limited ability to connect explicit and implicit emotions, despite the importance of interpreting emotion in conversation. PRC-Emo combines emotion-sensitive prompts, demonstration retrieval, and curriculum learning, and reports new state-of-the-art performance on IEMOCAP and MELD.

  • Problem

    LLMs have limited ability to capture connections between explicit and implicit emotions, which are both relevant to understanding speakers’ emotional states in conversation.

    Source: Introduction · p. 2

  • Method

    PRC-Emo combines prompts for explicit and implicit emotion interpretations, an ERC-specific demonstration repository, and curriculum learning ordered by weighted emotional shifts.

    Source: Introduction · p. 2

  • Results

    PRC-Emo achieves new state-of-the-art performance on the IEMOCAP and MELD benchmark datasets.

    Source: Conclusion · p. 7

  • Takeaways & Limitations

    The paper highlights the potential of combining prompt-based learning with curriculum learning strategies to advance ERC tasks.

    Source: Conclusion · p. 7

Abstract

from arXiv · show

Emotion Recognition in Conversation (ERC) is a crucial task for understanding human emotions and enabling natural human-computer interaction. Although Large Language Models (LLMs) have recently shown great potential in this field, their ability to capture the intrinsic connections between explicit and implicit emotions remains limited. We propose a novel ERC training framework, PRC-Emo, which integrates Prompt engineering, demonstration Retrieval, and Curriculum learning, with the goal of exploring whether LLMs can effectively perceive emotions in conversational contexts. Specifically, we design emotion-sensitive prompt templates based on both explicit and implicit emotional cues to better guide the model in understanding the speaker's psychological states. We construct the first dedicated demonstration retrieval repository for ERC, which includes training samples from widely used datasets, as well as high-quality dialogue examples generated by LLMs and manually verified. Moreover, we introduce a curriculum learning strategy into the LoRA fine-tuning process, incorporating weighted emotional shifts between same-speaker and different-speaker utterances to assign difficulty levels to dialogue samples, which are then organized in an easy-to-hard training sequence. Experimental results on two benchmark datasets -- IEMOCAP and MELD -- show that our method achieves new state-of-the-art (SOTA) performance, demonstrating the effectiveness and generalizability of our approach in improving LLM-based emotional understanding.

Introduction

ERC systems must recognize both expressed and underlying emotions, while current demonstration repositories and training strategies leave gaps in diversity and learning from imbalanced data. PRC-Emo combines emotion-sensitive prompts, a diverse retrieval repository, and easy-to-hard curriculum learning; the paper reports consistent state-of-the-art results across multiple benchmarks.

  • Most existing ERC prompts inadequately address both explicit emotions and implicitly felt emotions, limiting comprehension of complex emotional states and recognition accuracy.Source: Introduction · pp. 1–2
  • Existing demonstration repositories rely mainly on standard training sets, limiting sample diversity and generalization to varied conversational scenarios.Source: Introduction · p. 2
  • ERC training research pays relatively little attention to training-process optimization, while class imbalance makes low-frequency emotion categories difficult to learn.Source: Introduction · p. 2
  • PRC-Emo generates explicit and implicit emotion interpretations as complementary prompt knowledge to capture speakers’ nuanced emotional states.Source: Introduction · p. 2
  • Its ERC-specific repository combines samples from IEMOCAP, MELD, and EmoryNLP with over 10,000 LLM-generated, human-refined utterances across six domains.Source: Introduction · p. 2
  • A weighted emotional-shift difficulty metric covers same-speaker and different-speaker transitions, organizing training buckets from easy to hard.Source: Introduction · p. 2
  • Experiments on multiple benchmarks show PRC-Emo achieves consistent state-of-the-art results.Source: Introduction · p. 2

Related Work

ERC research has progressed from recurrent, graph, and pretrained-language-model approaches toward generative LLM methods, while curriculum learning has also been applied to emotional transitions. The related work describes these directions and their approaches to modeling dialogue context and emotion.

Related Work

  • ERC work is broadly divided into traditional methods and LLM-based methods, with curriculum learning also considered in the field.Source: Related Work · p. 2

Traditional Methods

  • Traditional ERC methods use RNNs for sequential and speaker-state modeling, GNNs for utterance dependencies, and PLMs for language representations.Source: Traditional Methods · p. 2

LLM-based Methods

  • LLM-based ERC includes generative InstructERC with emotion alignment and retrieval prompting, and BiosERC with speaker background prompts and LoRA fine-tuning.Source: LLM-based Methods · pp. 2–3

Curriculum Learning

  • Prior curriculum methods model emotion shifts within a speaker, while weighted emotion shifts are described as omitting transitions between different speakers.Source: Curriculum Learning · p. 3

Methodology

PRC-Emo defines emotion recognition in conversation, adds emotion-related supplementary prompts and demonstration retrieval, and uses curriculum learning during training.

Problem Definition

  • ERC represents a conversation as speaker-attributed utterances and assigns each utterance an emotion category from the available label set.Source: Problem Definition · pp. 3–4

Extraction of External Supplementary Knowledge

  • Prompt templates use dialogue history to generate explicit and implicit emotion interpretations, plus speaker characteristics from the full dialogue, as auxiliary input knowledge.These sources are incorporated into the retrieval template for emotion recognition.Source: Extraction of External Supplementary… · p. 4

Retrieval Template Module

  • The retrieval template combines instructions, history, external knowledge, three SBERT-matched demonstrations, and dataset-specific label constraints with a 36,712-utterance ERC repository.The repository combines training utterances from IEMOCAP, MELD, and EmoryNLP with a self-constructed dataset; generated samples are retained only when two researchers agree with the original labels.Source: Components of Retrieval Template Module · p. 4, Demonstration Retrieval Repository · p. 4

Curriculum Learning

  • Difficulty scores use similarity-weighted shifts across same-speaker sequences and consecutive speaker changes; sorted difficulty buckets enter training from easiest upward.After all buckets are introduced, additional epochs train on the entire dataset.Source: Difficulty Measure Function · p. 5, Training Scheduler · p. 6

Experimental Settings

Experiments use two conversation datasets and compare conventional and LLM-based ERC methods, with weighted F1 as the default metric for most comparisons.

Experimental Settings

Datasets

  • IEMOCAP provides dyadic conversations, while MELD is a multiparty dataset derived from Friends.The datasets’ composition is shown in Table 1.Source: Datasets · p. 6

Baselines

  • Comparisons cover conventional ERC models and LLM-based methods, using weighted F1 as the default metric for most experiments.The LLM baselines include InstructERC and BiosERC.Source: Baselines · p. 6

Implementation Details

  • IEMOCAP experiments use Qwen2.5-7B-Instruct, MELD experiments use Qwen3-8B, and Qwen3-14B generates emotion interpretations and speaker characteristics.Experiments use one NVIDIA 4090D GPU and average results across five random seeds.Source: Implementation Details · p. 6

Results and Analysis

PRC-Emo achieves state-of-the-art results on IEMOCAP and MELD. Ablations favor the full prompt, while comparisons examine its components, curriculum design, and base-model variation across datasets.

Ablation Experiments

  • Removing Prompt, Retrieval, and Curriculum modules shows each contributes; Prompt has the largest impact, while Retrieval and Curriculum further enhance performance.Without all three modules, the model is essentially equivalent to directly fine-tuning an LLM using LoRA.Source: Ablation Experiments · p. 6

Comparative Experiments on Prompt Design

  • The full PRC-Emo prompt performs best; removing Retrieval, emotion interpretations, or speaker characteristics consistently lowers performance, with interpretations having the greatest impact.Source: Comparative Experiments on Prompt Design · p. 6

Comparative Experiments on Curriculum Learning Methods for ERC

  • Unlike ES and WES curricula that consider only same-speaker shifts, the proposed curriculum includes cross-speaker emotional changes to assess sentence difficulty.The comparison uses the same base model and prompt across the three curriculum strategies.Source: Comparative Experiments on Curriculum… · p. 6

Comparison with the State of the Art

Comparative Experiments of Different LLMs

  • With only Qwen2.5-7B and Qwen3-8B compared, Qwen2.5-7B significantly outperforms on IEMOCAP, while Qwen3-8B performs significantly better on MELD.The authors suggest Qwen3’s complex-dialogue optimization and possible data leakage from MELD’s widely available Friends script as possible explanations.Source: Comparative Experiments of Different… · p. 7

Conclusion

PRC-Emo combines prompt engineering, demonstration retrieval, and curriculum learning for ERC. It reports state-of-the-art performance on IEMOCAP and MELD and identifies several directions for future work.

Conclusion

  • PRC-Emo integrates prompt engineering, demonstration retrieval, and curriculum learning for emotion recognition in conversation.Its curriculum orders training from easy to hard using weighted emotional shifts between same-speaker and different-speaker utterances.Source: Conclusion · p. 7
  • The conclusion includes a table comparing curriculum learning methods.The caption identifies the table's scope but does not provide comparison outcomes.Source: Table 5 · p. 7
  • The conclusion includes a table comparing different large language models.The caption identifies the table's scope but does not provide comparison outcomes.Source: Table 6 · p. 7
  • PRC-Emo reports new state-of-the-art performance on the IEMOCAP and MELD benchmark datasets.The paper highlights combining prompt-based learning with curriculum learning as a potential direction for ERC.Source: Conclusion · p. 7
  • Future work will explore stronger LLMs, more efficient prompting paradigms, and advanced curriculum designs.The authors aim to enhance emotional reasoning and ERC-system robustness and generalization.Source: Conclusion · p. 7

Appendix A

Appendix A describes a self-constructed dataset spanning multiple domains and emotions, with tables showing category distributions and representative dialogues.

Appendix A

  • 14,009 sentences span seven domains, with comprehensive scenarios largest at 3,314 and healthcare smallest at 1,625.Family interactions contain 2,015 sentences and workplace scenarios 1,862.Source: Appendix A · p. 9
  • Neutral is the most prevalent emotion category at 3,309 sentences, while anger is least frequent at 2,566.Table 7 presents sample and emotion-label distributions across categories.Source: Appendix A · p. 9
  • Tables 8 and 9 provide representative dialogue examples from different dataset domains.The paper describes these examples as illustrating dialogue diversity and contextual richness.Source: Appendix A · p. 9

Appendix B

Appendix B reports LoRA fine-tuning settings and points to tables on dataset distributions and representative dialogues.

Appendix B

  • LoRA fine-tuning uses rank 32, a learning rate of 3e-4, four epochs, and a context window of five.Curriculum learning is applied in the first two epochs, followed by two epochs of overall training; difficulty has two levels updated each epoch.Source: Appendix B · p. 9
  • Training uses batch size 1 with four gradient-accumulation steps, maximum gradient norm 0.3, and warm-up ratio 0.03.Evaluation, model saving, and logging occur every 50 steps, and weighted F1 selects the best model.Source: Appendix B · p. 9
  • Table 7 presents the distribution of emotions across domains in the self-constructed dataset.Its caption specifies the table's scope but provides no distribution values.Source: Table 7 · p. 10
  • Tables 8 and 9 contain representative dialogue examples from different dataset domains.The captions label the tables as Parts I and II, respectively.Source: Table 8 · p. 10, Table 9 · p. 11

Related papers

Loading 2511.07061v3…