KOBE

xxxx

Project | arXiv

Towards KnOwledge-Based pErsonalized Product Description Generation in E-commerce.
Qibin Chen^*, Junyang Lin^*, Yichang Zhang, Hongxia Yang, Jingren Zhou, Jie Tang.
^*Equal contribution.
In KDD 2019 (Applied Data Science Track)

Prerequisites

• Linux or macOS
• Python 3
• PyTorch >= 1.0.1
• NVIDIA GPU + CUDA cuDNN

Getting Started

Installation

Clone this repo.

git clone https://github.com/Lumonk/KOBE
cd KOBE

Please install dependencies by

pip install -r requirements.txt

Dataset

• We use the TaoDescribe dataset, which contains 2,129,187 product titles and descriptions in Chinese.
• (optional) You can download the un-preprocessed dataset from here or here (for users in China).

Training

Download preprocessed data

• First, download the preprocessed TaoDescribe dataset by running python scripts/download_preprocessed_tao.py.
  • If you're in regions where Dropbox are blocked (e.g. Mainland China), try python scripts/download_preprocessed_tao.py --cn.
• (optional) You can peek into the data/aspect-user/preprocessed/test.src.str and data/aspect-user/preprocessed/test.tgt.str, which include product titles and descriptions in the test set, respectively. In src files, <x> <y> means this product is intended to show with aspect <x> and user category <y>. Note: this slightly differs from the <A-1>, <U-1> format descripted in the paper but basically they are the same thing. You can also peek into data/aspect-user/preprocessed/test.supporting_facts_str to see the knowledge we extracted from dbpedia for the corresponding product.

Start training

• Different configurations for models in the paper are stored under the configs/ directory. Launch a specific experiment with --config to specify the path to your desired model config and --expname to specify the name/number of this experiment which will be used in logging.

• We include three config files here: the baseline, KOBE without adding external knowledge, and full KOBE model.

• Baseline

python core/train.py --config configs/baseline.yaml --expname baseline

• KOBE without adding knowledge

python core/train.py --config configs/aspect_user.yaml --expname aspect-user

• KOBE

python core/train.py --config configs/aspect_user_knowledge.yaml --expname aspect-user-knowledge

The default batch size is set to 64. If you are having OOM problems, try to decrease it with the flag --batch-size.

Track training progress

• You can use TensorBoard. It can take (roughly) 12 hours for the training to stop. To get comparable results in paper, you need to train for even longer (by editing epoch in the config files). However, the current setting is enough to demonstrate the effectiveness of our model.

tensorboard --logdir experiments --port 6006

Generation

• During training, the generated descriptions on the test set is saved at experiments/<expname>/candidate.txt and the ground truth is at reference.txt. This is generated by greedy search to save time in training and doesn't block repetitive terms.
• To do beam search with beam width = 10, run the following command.

python core/train.py --config configs/baseline.yaml --mode eval --restore experiments/finals-baseline/checkpoint.pt --expname eval-baseline --beam-size 10

Evaluation

• BLEU
• DIVERSITY

Experiment results

• baseline
• aspect-user:
• aspect-User_2: aspect, no user, encoder layers = 4
• sapect-user-know :

TODO

If you have ANY difficulties to get things working in the above steps, feel free to open an issue. You can expect a reply within 24 hours.

Cite

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

@article{chen2019towards,
  title={Towards Knowledge-Based Personalized Product Description Generation in E-commerce},
  author={Chen, Qibin and Lin, Junyang and Zhang, Yichang and Yang, Hongxia and Zhou, Jingren and Tang, Jie},
  journal={arXiv preprint arXiv:1903.12457},
  year={2019}
}