This repo contains code and data for the paper: "Reinforced Target-driven Conversational Promotion".
Please install libraries/packages listed in the requirements.txt file. Make sure that you are using CUDA 11.6. Otherwise, some unexpected behaviors might happend.
To preprocess and repurpose the DuRecDial 2.0 dataset for our task, please run:
sh scripts/preprocess.sh
To train our short-term planning model, please run:
sh scripts/planning/train_planning.sh
To produce the plan with the trained short-term planning model, please run the following command:
sh scripts/planning/test_planning.sh
To evaluate the trained short-term planning model on the next goal and topic prediction tasks, please run:
python evaluate.py
To train our long-term planning model, you need to first train the reward model. The reward model take as inputs a sequence of dialogue actions and output whether the given sequence is smooth or not. To train the reward model, please run the following command:
sh scripts/rl/train_reward.sh
To train our short-term planning model, please following commands:
sh scripts/rl/pretrain_rl.sh
sh scripts/rl/train_rl_after.sh
The strategic balancing method computes a weighted combination of two probability distributions, one from short-term planning and the other from long-term planning. Then based on the computed distribution, we will sample the next dialogue action. To perform strategic balancing, please run the following command.
sh scripts/balancing.sh
Given a generated plan from the planning part, you could run the action-guided prefix tuning method with the following command.
sh scripts/train_gpt2_prompt_new.sh
To generate responses with the generation model, you could run the following command:
sh scripts/test_gpt2_prompt_new.sh
To evaluate the performance of the trained generation model, please run the following command:
python eval/eval_dialogue.py --eval_file ${eval_file} --gold_file caches/path/test.pkl