CR-Walker

Code for paper "CR-Walker: Conversational Recommender System with Tree-structured Graph Reasoning and Dialog Acts" EMNLP 2021.

you can find our paper at arxiv.

Cite this paper:

@inproceedings{ma2021crwalker,
  title={CR-Walker: Tree-Structured Graph Reasoning and Dialog Acts for Conversational Recommendation},
  author={Ma, Wenchang and Takanobu, Ryuichi and Huang, Minlie},
  booktitle={Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
  pages={1839--1851},
  year={2021},
  organization={ACL}
}

Data

• google link to raw data and our model checkpoints. Table of content:

  CR-Walker
  ├─data
  │  ├─gorecdial
  │  │  └─raw
  │  ├─gorecdial_gpt
  │  ├─redial
  │  │  └─raw
  │  └─redial_gpt
  └─saved
  

• download to [your home directory]/CR-Walker/.

Train

• For GoRecdial:

  python train_gorecdial.py --option train --model_name <your_model_name> --pretrain
  

• For Redial:

  python train_redial.py --option train --model_name <your_model_name> --pretrain 
  

  We implemented an MIM pretraining stage similar to KGSF to accelerate training. Also, we provided option of adding wordnet features by adding "--word_net" as command line option.

Test Recommendation

• For GoRecdial

  python train_gorecdial.py --option test --model_name gorecdial_reason_128
  

• For Redial:

  python train_redial.py --option test --model_name redial_reason_128
  

  You can directly evaluate the best model checkpoints for the two datasets that we provided. The results may slightly differ from the paper since we re-trained the model. Note that the reasoning width ('sample' argument in conf.py) has been set to 1 for speed during training. You can tune it larger along with the selection threshold ('threshold' argument in conf.py) to yield better performance.

Test Generation

• For GoRecdial

  python train_gorecdial.py --option test_gen --model_name gorecdial_reason_128
  

• For Redial:

  python train_redial.py --option test_gen --model_name redial_reason_128
  

  Similarly, you can tune the selection threshold, reasoning width and max number of leaf nodes ('max_leaf' argument in conf.py) to control generation.

Requirements

python==3.6.10

pytorch==1.4.0

torch_geometric==1.6.0