This repository will provide the codes and data used in the AAAI'24 paper, COOPER: Coordinating Specialized Agents towards a Complex Dialogue Goal.

The data/ directory provides the annotated data we used in our experiments. Please refer to the data/README.md file for details about the downloading address and how to use the preprocessing codes.

The codes for implementing the topic ranking module in provided in the TopicRanking directory. Please run the following commands for obtaining the typical target states on the ESConv dataset (please refer to the Local Analysis with Specialized Agents for detailed explanation), and prepare the data for training the topic ranking module.

cd TopicRanking
python subgoal_state_clustering.py --dataset_type esconv
python generate_ranking_data.py --dataset_type esconv

Run the following command to train the topic ranking module.

python train.py --data_dir data/esconv/ --output_dir output/esconv/

Run the following command to conduct inference on the test set.

python inference.py \
--model_path output/esconv/ \
--data_dir data/esconv/ \
--output_dir inference_results/esconv/

To conduct the experiment on the P4G dataset, you can substitute esconv with P4G in the above commands.

To implement the prompt-based generator, you need to first set up the OpenAI API key. Specifically, you can simply do this by input the key in the utils/call_LLM_API/api_key.txt file.

To calculate the NLG metrics for evaluation, please download the evaluation codes nlgeval/ from this link. Decompress the zip file and put it in the utils/ directory.

Please run the following command to use the prompt-based generator to generate the utterances on the ESConv dataset.

cd PT_Generator
python inference.py --inference_data_path ../TopicRanking/inference_results/esconv/test.json --output_path results/esconv.json --dataset_type esconv

To conduct the experiment on the P4G dataset, you can substitute all esconv with P4G in the above command.

To train the finetuned generator, please run the following command.

cd FT_Generator
CUDA_VISIBLE_DEVICES=0,1 python train.py \
--model_path facebook/bart-base \
--output_dir ./model/esconv/  \
--data_dir ../TopicRanking/data/esconv/

To conduct inference on the test set, please run the following command.

CUDA_VISIBLE_DEVICES=0,1 python train.py \
--do_train False \
--model_path ./model/esconv/ \
--data_dir ../TopicRanking/data/esconv/

To conduct the experiment on the P4G dataset, you can substitute all esconv with P4G in the above command.