Redundant product titles compression

Introduction

Code for redundant product titles compression, it includes GRU_SATT model and some baselines.

Requirements

Python 3.5+
Tensorflow 1.10+

Train

GRU_SATT means self_attention based GRU model for redundant product titles compression. GRU_SATT_SMD is an optimization on GRU_SATT. The simple training instructions are shown as follows. You can modify other parameters to train some baselines. "Compress" tasks represents the human label compression titles and "Phrase" means the labeled tokens are less than three.

Train on GRU_SATT

The default includs represents the "Phrase" tasks and SELF_ATT model

python -m main

For "Compress" tasks

python -m main -task=Compress

Train on GRU_SATT_SMD

For Phrash tasks

python -m main -model=GRU_SATT_SMD

For "Compress" tasks

python -m main -model=GRU_SATT_SMD -task=Compress

Load pretrained train model

Several pretrained models have been provided in the model folder. You can use the following command to load them.

Load default pretrained model GRU_SATT for "Phrase" tasks.

python -m load

Load pretrained model GRU_SATT for "Compress" tasks.

python -m load -task=Compress

Load pretrained model GRU_SATT_SMD for "Phrase" tasks.

python -m load -model=GRU_SATT_SMD

Load pretrained model GRU_SATT_SMD for "Compress" tasks.

python -m load -model=GRU_SATT_SMD -task=Compress

Citation

If you use the code in your paper, please cite it as:

@article{fu2019ersnet,
  title={Self-attention based neural networks for product titles compression},
  author={FU Yu and LI You and LIN Yu-ming and ZHOU Ya and others},
  journal={Journal of East China Normal University (Natural Sciences)},
  volume={5},
  pages={113--122},
  year={2019}
}

flycsuu / enn