PromptECL: Leveraging Prompt Engineering to Unlock Emotion Classiffcation Capabilities in LLMs

ACL ARR 2025 May Submission496 Authors

Abstract:

Emotion Recognition in Conversation (ERC)
is a crucial task in natural language processing
that aims to identify emotional states
within dialogic interactions. While existing
approaches typically employ conventional models
like BERT, GRU, and GNN for contextual
feature extraction and emotion classiffcation,
they often face limitations in interpretability,
architectural complexity, and performance
constraints. To address these challenges, we
present PromptECL, a novel framework that
harnesses few-shot prompt engineering to unlock
the intrinsic emotion classiffcation capabilities
of Large Language Models (LLMs)
through strategic prompt templates that activate
LLMs’ generative potential for latent emotion
cue extraction and semantic data augmentation,
followed by ffne-tuning to enhance emotional
analysis proffciency. Comprehensive evaluations
across four benchmark datasets demonstrate
PromptECL’s superior effectiveness and
generalization capacity, establishing new stateof-the-art
performance with improved average
weighted F1 scores on IEMOCAP, MELD,
and EmoryNLP, while achieving a remarkable
7.67% micro F1 score improvement over previous
best results on DailyDialog.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: Sentiment Analysis, Stylistic Analysis, Argument Mining, Question Answering,

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Keywords: Emotion Recognition in Conversation, ​​Large Language Models, Natural Language Processing

Submission Number: 496