This repository contains the code for the paper "For more detailed results and analysis, please refer to: Fine-Grained Emotion Recognition via In-Context Learning" published at CIKM 2025.

• Environment Setup
• Data Download
• Configuring File
• Running Instructions
• Performance
• Citation

• Python 3.7+
• CUDA-capable GPU (recommended for local models)

1. Clone this repository:

git clone https://github.com/zhouzhouyang520/EICL.git
cd EICL

2. Install the required dependencies:

pip install -r requirements.txt

Download the required data from one of the following sources:

• Baidu Netdisk: EICL_data (Extraction code: i3tn)
• Google Drive: EICL_data

1. Extract data: After downloading, extract data.zip and place it in the EICL root directory.

2. EmpatheticIntents (Optional, only needed if you need to rebuild data):

   • Extract EmpatheticIntents.zip and place it in the EICL root directory.

3. Build models folder with the following structure:

   models/
   ├── LLMs/
   │   ├── Llama3.1_8b/
   │   ├── Mistral_Nemo/
   │   └── Phi3.5_mini/
   └── pre_trained_models/  (Optional, only needed if you need to rebuild data)
       ├── EmpatheticIntents/
       ├── all_mpnet_base_v2/
       └── roberta_large_goEmotions/
   

   • LLMs: Download three large language models and place them in models/LLMs/ with the exact folder names: Llama3.1_8b, Mistral_Nemo, and Phi3.5_mini.

   • pre_trained_models (Optional): If you need to rebuild data, download three pre-trained models:

     • EmpatheticIntents
     • all_mpnet_base_v2
     • roberta_large_goEmotions

Before sending requests to Claude or ChatGPT, please fill in the “url” and “api_key” fields in configs/run.json.

Run experiments using the main script:

python main.py \
  --auxiliary_model EI \
  --experiment_type EICL \
  --dataset ED \
  --models Phi-3.5

• --auxiliary_model: Auxiliary model type, choices: EI, GE (can specify multiple)
• --experiment_type: Experiment type, choices: baseline, ICL, EICL, zero-shot (can specify multiple)
• --dataset: Dataset name, choices: ED, EDOS, GE, EI (can specify multiple)
• --models: LLM model names, e.g., Phi-3.5, Llama3.1_8b, Mistral-Nemo, ChatGPT, Claude, gpt-4o-mini (can specify multiple)

Run EICL experiment with Phi-3.5 on ED dataset:

python main.py --auxiliary_model GE --experiment_type EICL --dataset ED --models Phi-3.5

Run multiple experiments:

python main.py \
  --auxiliary_model EI GE \
  --experiment_type ICL EICL \
  --dataset ED EDOS \
  --models Phi-3.5 Llama3.1_8b

Run with GPU:

CUDA_VISIBLE_DEVICES=0 python main.py \
  --auxiliary_model GE \
  --experiment_type EICL \
  --dataset EDOS \
  --models Phi-3.5

Using the provided shell script:

sh eicl.sh 0  # 0 is the GPU ID

Important Note: Due to accidental deletion of previous data and code, we reconstructed the codebase in a different environment. The current implementation differs from the original paper in the following aspects:

• Data Re-splitting: The datasets were re-partitioned, which may cause slight performance differences.

• Updated Models: We use more advanced large language models: Claude-Haiku-4.5 and GPT-4o-mini. Although the original EICL experiments used Claude-Haiku and ChatGPT-3.5-Turbo, those models are now outdated. To build more comparable and up-to-date baselines, we adopted the newer models.

Despite these differences, EICL continues to deliver excellent performance, demonstrating its robustness and effectiveness.

The following tables present the performance results (Accuracy and Macro F1) on different datasets using different auxiliary models. All results are reported as percentages.

If you find this project helpful, please cite our paper:

@inproceedings{Ren_2025,
  series={CIKM' 25},
  title={Fine-Grained Emotion Recognition via In-Context Learning},
  url={http://dx.doi.org/10.1145/3746252.3761319},
  DOI={10.1145/3746252.3761319},
  booktitle={Proceedings of the 34th ACM International Conference on Information and Knowledge Management},
  publisher={ACM},
  author={Ren, Zhaochun and Yang, Zhou and Ye, Chenglong and Sun, Haizhou and Chen, Chao and Zhu, Xiaofei and Liao, Xiangwen},
  year={2025},
  month=nov,
  pages={2503--2513},
  collection={CIKM' 25}
}