princeton-nlp / NegotiationToM

Code release for Improving Dialog Systems for Negotiation with Personality Modeling.

NegotiationToM

Improving Dialog Systems for Negotiation with Personality Modeling. Implements our ToM algorithms.

Our code for internal using is available in another repo.

A more clear code with proper instruction will be available soon here.

---

Installation

Dependencies

0. Clone code from github:

git clone [e-mail]:princeton-nlp/NegotiationToM.git
cd NegotiationToM 

1. Create a conda environment:

conda create --name <env> --file tom_conda10.txt

2. install pytorch

# For Linux
# CUDA 9.0
conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch

# CUDA 10.0
conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=10.0 -c pytorch

# CPU Only
conda install pytorch-cpu==1.1.0 torchvision-cpu==0.3.0 cpuonly -c pytorch

• Check official website for more installation instructions .

3. pip install additional dependencies

# install other dependecies
pip install -r requirements_cuda10.txt
# install our code
pip install -e .

Dependence Reference

• Python 3.5, pytorch=1.1.0

  • For cuda9: pytorch=1.1.0=py3.5_cuda9.0.176_cudnn7.5.1_0
  • For cuda10: pytorch=1.1.0=py3.5_cuda10.0.130_cudnn7.5.1_0

• Additional dependencies:

  • tom_new.txt, requirements.txt
  • For cuda10: tom_cuda10.txt, requirements_cuda10.txt

Traning instruction

Preprocess

PYTHONPATH=. python core/price_tracker.py --train-examples-path data/train-luis-clean2.json --output data/price_tracker.pkl

Train a Supervised Learning Agent

bash craigslistbargain/exp_scripts/identifier/old/train_sl.sh

Train a Reinforcement Learning Agent

• Generate scenarios

PYTHONPATH=. python ../scripts/chat_to_scenarios.py --chats data/train-luis-post.json --scenarios data/train-scenarios.json
PYTHONPATH=. python ../scripts/chat_to_scenarios.py --chats data/dev-luis-post.json --scenarios data/dev-scenarios.json

• Train the RL model

  parameter <use_gpu>: "--gpuid 0" for gpu, "" for cpu.

bash exp_scripts/rl/train_a2c.sh <exp_name> <use_gpu> <seed> <learning_rate> 

Train a ToM model

• Sample data

bash exp_scripts/identifier/sample_data.sh <use_gpu> <seed> 

• Implicit Model

bash exp_scripts/identifier/train_uttr_history_tom.sh <exp_name> <use_gpu>

• Explicit Model

bash exp_scripts/identifier/train_uttr_id_history_tom.sh <exp_name> <use_gpu>

Evaluate result of different model.

For evaluation process, gpu acceleration is unnecessary.

Parameter <1/beta> equal to $\frac{1}{beta}$ of tom inference process in our paper.

• Rule Model

bash exp_scripts/rl/eval_rule.sh <exp_name> "" <seed> <1/beta> "${CHECK_POINT}"

• SL Model

bash exp_scripts/rl/eval_sl.sh <exp_name> "" <seed> <1/beta> "${CHECK_POINT}" 

• RL Model

bash exp_scripts/rl/eval_rl.sh <exp_name> "" <seed> <1/beta> "${CHECK_POINT}" 

• Implicit ToM Model

bash exp_scripts/rl/eval_tom_noid.sh <exp_name> "" <seed> <1/beta>" "${CHECK_POINT}" "${TOM_CHECK_POINT}"

• Explicit ToM Model

bash exp_scripts/rl/eval_tom.sh <exp_name> "" <seed> <1/beta> "${CHECK_POINT}" "${TOM_CHECK_POINT}"