Input-to-Output Gate for Language Models

This repository contains the Input-to-Output gate proposed in Input-to-Output Gate to Improve RNN Language Models (IOG) published in IJCNLP 2017. IOG, which is the method to improve existing trained RNN language models, has an extremely simple structure, and thus, can be easily combined with any RNN language models.

In addition, this repository also contains various RNN language models including LSTM with dropout, variational LSTM, and Recurrent Highway Network for the base RNN language models.

If you use this code or our results in your research, please cite:

@inproceedings{iog4lm,
  title={{Input-to-Output Gate to Improve RNN Language Models}},
  author={Takase, Sho and Suzuki, Jun and Nagata, Masaaki},
  booktitle={Proceedings of the 8th International Joint Conference on Natural Language Processing (IJCNLP 2017)},
  year={2017}
}

Performance on benchmark datasets

The following tables show the performance (word-level perplexity) of IOG with some baselines. The paper describes more detailed experiments.

Penn Treebak

Model	Param	Valid	Test
LSTM (medium)	20M	86.2	82.7
LSTM (medium) + IOG	26M	84.1	81.1
Variational LSTM (medium)	20M	81.9	79.7
Variational LSTM (medium) + IOG	26M	81.2	78.1
Variational Recurrent Highway Network (depth 8)	32M	71.2	68.5
Variational Recurrent Highway Network (depth 8) + IOG	38M	69.2	66.5
Variational Recurrent Highway Network (depth 8) + WT	23M	69.2	66.3
Variational Recurrent Highway Network (depth 8) + WT + IOG	29M	67.0	64.4

Wikitext-2

Model	Param	Valid	Test
LSTM (medium)	50M	102.2	96.2
LSTM (medium) + IOG	70M	99.2	93.8
Variational LSTM (medium)	50M	97.2	91.8
Variational LSTM (medium) + IOG	70M	95.9	91.0
Variational Recurrent Highway Network + WT	43M	80.1	76.1
Variational Recurrent Highway Network + WT + IOG	63M	76.9	73.9

Software Requirements

• Python 2
• Chainer

We note that it is impossible to run these codes on Python 3.

How to use

• Run preparedata.sh to obtain the preprocessed Penn Treebank dataset and construct .npz files.
• Train the base language model
  • E.g., to obtain LSTM language model with dropout, run

python learnLSTMLM.py -g 0 --train data/ptb/ptb.train.npz --valid data/ptb/ptb.valid.npz -v data/ptb/ptb.train.vocab
   --valid_with_batch --output mediumLSTMLM

• Train Input-to-Output gate for the above base language model
  • E.g., run the following command to train Input-to-Output gate for the lstm language model in the above example.

python learnIOG.py -g 0 --train data/ptb/ptb.train.npz --valid data/ptb/ptb.valid.npz -s mediumLSTMLM.setting
   -m mediumLSTMLM.epoch39_fin.bin --output iog4mediumLSTMLM

• Run testLM.py for computing perplexity of the base language model
  • E.g., run python testLM.py -g 0 -t data/ptb/ptb.test.npz -s mediumLSTMLM.setting -m mediumLSTMLM.epoch39_fin.bin for the lstm language model in the above example.
• Run testIOG.py for computing perplexity of the base language model with Input-to-Output gate.
  • E.g., run python testIOG.py -g 0 -t data/ptb/ptb.test.npz -s iog4mediumLSTMLM.setting -m iog4mediumLSTMLM.bin for the Input-to-Output gate in the above example.
• If you want to use the cache mecanism, specify --cachesize N in running testLM.py or testIOG.py