THUDM / KBRD

KBRD: Towards Knowledge-Based Recommender Dialog System

Paper accepted at EMNLP-IJCNLP 2019. Latest version at arXiv.

• New: code and README are improved.
• We curated a paper list for NLP + Recommender System at THUDM/NLP4Rec-Papers. Contributions are welcome.

Prerequisites

• Linux
• Python 3.6
• PyTorch >= 1.2.0

Getting Started

Installation

Clone this repo.

git clone https://github.com/THUDM/KBRD
cd KBRD

Please install dependencies by

pip install -r requirements.txt

Dataset

• We use the ReDial dataset, which will be automatically downloaded by the script.
• Download the refined knowledge base (dbpedia) used in this paper [Google Drive]. Decompress it and get the <path/to/KBRD/dbpedia/> folder, which should contain two files mappingbased_objects_en.ttl and short_abstracts_en.ttl.
• Download the proprocessed extracted entities set [Google Drive] and put it under <path/to/KBRD/data/redial/.

Training

1. To train the recommender part, run:

bash scripts/both.sh <num_exps> <gpu_id>
(optional) bash scripts/baseline.sh <num_exps> <gpu_id>

2. To train the dialog part, run:

bash scripts/t2t_rec_rgcn.sh <num_exps> <gpu_id>

The test results are displayed at the end of training and can also be found at saved/<model_name>.test.

Logging

Training outputs, TensorBoard logs and models files are be saved in saved/ folder.

Evaluation

1. scripts/score.py is used to hypothesis testing the significance of improvement between different models. To use, first run multiple experiments with num_exps > 1, for example:

bash scripts/both.sh 2 <gpu_id>
bash scripts/baseline.sh 2 <gpu_id>

Then,

python scripts/score.py --name-1 saved/release_baseline --name-2 saved/both_rgcn --num 2 --metric recall@50

where you should remove the trailing _0, _1 automatically added to the model names, nums should be set the same as num_exps above, and recall@50 can be replaced with other evaluation metrics in the paper.

Sample output:

[0.298, 0.2918]
0.2949
0.0031
[0.3417, 0.3369]
0.3393
0.0024
Ttest_indResult(statistic=-11.325204070341204, pvalue=0.007706635327863829)

2. scripts/display_model.py is used to generate responses.

python scripts/display_model.py -t redial -mf saved/transformer_rec_both_rgcn_0 -dt test

Example output ([TorchAgent] is our model output):

~~
[eval_labels_choice]: Oh, you like scary movies?
I recently watched __unk__
[movies]:
  37993
[redial]: 
Hello!
Hello!
What kind of movies do you like?
I am looking for a movie recommendation.   When I was younger I really enjoyed the __unk__
[label_candidates: 3|37993|50395||Oh, you like scary movies?
I recently watched __unk__]
[eval_labels: Oh, you like scary movies?
I recently watched __unk__]
   [TorchAgent]: have you seen "The Shining  (1980)" ?
~~

3. scripts/show_bias.py is used to show the vocabulary bias of a specific movie (like the qualitative analysis in Table 4)

python scripts/show_bias.py -mf saved/transformer_rec_both_rgcn_0

❗ Common Q&A

1. Understanding model outputs. Please see #15 (comment).

2. Adapting this code to other datasets. It is not straightforward for this code to be run on other datasets currently. The main reason is that we cached the entity linking process in KBRD for ReDial. Please see #10 (comment) for details.

3. Why the recommender and the dialog part are trained separatedly? Please refer to #9 (comment) for detailed explanation.

If you have additional questions, please let us know.

Cite

Please cite our paper if you use this code in your own work:

@article{chen2019towards,
  title={Towards Knowledge-Based Recommender Dialog System},
  author={Chen, Qibin and Lin, Junyang and Zhang, Yichang and Ding, Ming and Cen, Yukuo and Yang, Hongxia and Tang, Jie},
  journal={arXiv preprint arXiv:1908.05391},
  year={2019}
}