Knowledge-Infused Topic Model for Empathetic Dialogue Response

This repository contains the implementation of the Knowledge-Infused Topic Model (KITM) for generating empathetic dialogue responses using topic modeling enhanced with commonsense knowledge.

📄 Accepted to APSIPA ASC 2025 (Proceedings Paper)

---

🔧 Installation

Install all required packages via:

conda install --yes --file requirements.txt

---

🛠️ Pre-processing

1. Download GloVe embeddings Download the GloVe.6B embeddings and place them in:

   /vectors
   

2. Dataset preprocessing

   • A preprocessed dataset (dataset_preproc.p) is already included under:

     /data/ED/dataset_preproc.p
     

   • If you would like to regenerate the dataset or modify knowledge types from COMET:

     1. Delete dataset_preproc.p

     2. Download the COMET-ATOMIC-2020 checkpoint

     3. Place it under:

        /data/ED/Comet
        

   • The dataset will be automatically reprocessed during training.

   • Note: We use BART-COMET because the GPT-2 COMET checkpoint is incompatible with this workflow.

---

📚 Datasets

This work uses two dialogue datasets, with the default configuration set to EmpatheticDialogues.

• EmpatheticDialogues
• DailyDialog

🔗 Data Resources (Download)

• EmpatheticDialogues

• DailyDialog

---

📁 Directory Structure

data
|--- ED
|    |--- Comet/          # COMET model for commonsense inference
|    |--- emp.pkl         # Topic appearance probabilities
|    |--- train.csv
|    |--- valid.csv
|    |--- test.csv
|
|--- DD
     |--- Comet/
     |--- dd.pkl          # Topic appearance probabilities
     |--- train.csv
     |--- valid.csv
     |--- test.csv

Directory Descriptions

data/ED/

Files and COMET knowledge used for EmpatheticDialogues.

data/DD/

Files and COMET knowledge used for DailyDialog.

---

🚀 Training

Run training with:

python main.py --cuda --save_path save/your_dir

---

🧪 Testing

python main.py --cuda --test \
    --save_path save/your_dir \
    --model_path save/dir_save/KITM_XXXX.XXX

---

📬 Contact

If you have questions or are interested in collaboration, feel free to contact:

present90308.ee11@nycu.edu.tw

Just let me know!