CSC722 Persona and Knowledge Empowered Conversational Model

Source codes for CSC722 course project "Persona and Knowledge Empowered Conversational Model".

This implementation is based on the paper Call for Customized Conversation: Customized Conversation Grounding Persona and Knowledge and its source code.

Environment Setup

We trained the models under the setting of python==3.8 and torch==2.0.1.

1.Make a virtual environment

conda create -n ENV_NAME python=3.8

2.Activate virtual environment

conda activate ENV_NAME

3.Install cuda-toolkit

conda install --channel "nvidia/label/cuda-11.8.0" cuda-toolkit

4.Install torch

pip3 install torch torchvision torchaudio

5.Install the required libraries.

pip install -r requirements.txt

Downloading Dataset

Please refer to the Dataset section below from the original FoCus Dataset paper to download the dataset and put into correct file path.

Run Program

The script csc722_runscripts.sh includes the commands that we run to train/validate the model. Below are some examples.

Training

Train with GPT-2 model

python train_focus.py --n_epochs 2 --lm_coef 10 --flag E2_L10_FOCUS --kp_method focus --model_name GPT2 --incontext > train_log_focus_e2_L10_gpt.log 2>&1

Train with BART model

python train_focus.py --n_epochs 2 --lm_coef 10 --flag E2_L10_FOCUS --kp_method focus --model_name BART --incontext > train_log_focus_e2_L10_bart.log 2>&1

Testing

Test with GPT-2 model

# Generation metrics
python evaluate_test.py --kp_method focus --model_name GPT2 --model_checkpoint models/train_focus_GPT2_E2_L10_FOCUS > test_log_focus_txt_test_GPT_MTL.log 2>&1
# Perplexity
python evaluate_test_ppl.py --kp_method focus --model_name GPT2 --model_checkpoint models/train_focus_GPT2_E2_L10_FOCUS > test_log_focus_txt_test_GPT_MTL_ppl.log 2>&1

Test with BART model

# Generation metrics
python evaluate_test.py --kp_method focus --model_name BART --model_checkpoint models/train_focus_BART_E2_L10_FOCUS > test_log_focus_txt_test_BART_MTL.log 2>&1
# Perplexity
python evaluate_test_ppl.py --kp_method focus --model_name BART --model_checkpoint models/train_focus_BART_E2_L10_FOCUS > test_log_focus_txt_test_BART_MTL_ppl.log 2>&1

Additional Information from original FoCus Dataset paper

Dataset [FoCus dataset v2]

This data is the modified version of the original data (which is reported in the paper) after ethical inspection.

FoCus v2 STATISTICS	Train	Valid
# dialogues	12,484	1,000
# avg rounds	5.63	5.64
# knowledge-only answers	37,488	3,007
# persona-knowledge answers	32,855	2,630
# landmarks	5,152	923
avg len of Human's utterances	40.70	40.21
avg len of Machine's utterances	138.16	138.60

You should create directories named infer_log_focus, train_log_focus, test_log_focus, models, data under FoCus folder.

We put train, valid, test files of the dataset in the data folder. (The test set will be available after March 2022.)

The project directory should follow this directory structure:

📦FoCus
┣ 📂data
┃ ┗ 📜train.json
┃ ┗ 📜valid.json
┣ 📂ignite
┣ 📂infer_log_focus
┣ 📂models
┣ 📂python_tf_idf
┣ 📂test_log_focus
┣ 📂train_log_focus
┣ 📜classification_modules.py
┣ 📜data_utils.py
┣ 📜evaluate_test.py
┣ 📜evaluate_test_ppl.py
┣ 📜inference.sh
┣ 📜inference_test.py
┣ 📜LICENSE
┣ 📜README.md
┣ 📜requirements.txt
┣ 📜test.sh
┣ 📜train.sh
┣ 📜train_focus.py
┗ 📜utils_focus