ht1221 / leap-lstm

tensorflow implementation for Leap-LSTM

Leap-LSTM

This repository contains source code (tensorflow-version) to reproduce the results presented in the paper Leap-LSTM: Enhancing Long Short-Term Memory for Text Categorization (IJCAI 2019).

@inproceedings{Leap-LSTM,
  title={Leap-LSTM: Enhancing Long Short-Term Memory for Text Categorization},
  author={Ting Huang, Gehui Shen, Zhi-Hong Deng},
  booktitle={IJCAI},
  year={2019}
}

model

The architecture of Leap-LSTM:

results

The accuracies on text classification tasks:

usage

To run the codes, you need to

1. download datasets from the github repository LEAM
2. to train the model, use a command like:

python train_classifier_yelp.py --rnn_model leap-lstm --yelp_set FULL \
  --gpu 0 --if_schedule 0 --decay_start 1  --target_skip_rate 0.6 \
  --keep_prob_word 1.0 --skip_reg_weight 1.0 --max_gradient_norm 1.0 \
  --nb_epoches 5 

The details of arguments can be found in skiplstm.py

other things

• tensorflow 1.11, maybe 1.12/1.13 is also okay
• Note that here we also provide the reproduction of skip-rnn and skim-rnn.