Deep-EOS PyTorch

Introduction

This is a pytorch implementation of "Deep-EOS: General-Purpose Neural Networks for Sentence Boundary Detection" (Schweter et al, 2019). Go here for the original Keras implementation: https://github.com/stefan-it/deep-eos

Currently only the LSTM and Bi-LSTM models are implemented.

The dataset download script and dataset creation methods are largely copied from Stefan Schweters repository.

Results

Test set

Language	PyTorch LSTM	PyTorch Bi-LSTM	Original LSTM	Original Bi-LSTM	Original CNN
German	0.9763	0.9763	0.9750	0.9760	0.9751
English	0.9862	0.9862	0.9861	0.9860	0.9858
Bulgarian	0.9893	0.9891	0.9922	0.9923	0.9919
Bosnian	0.9917	0.9919	0.9957	0.9959	0.9953
Greek	0.9958	0.9957	0.9967	0.9969	0.9963
Croatian	0.9921	0.9918	0.9946	0.9948	0.9943
Macedonian	0.9772	0.9786	0.9810	0.9811	0.9794
Romanian	0.9875	0.9874	0.9907	0.9906	0.9904
Albanian	0.9922	0.9920	0.9953	0.9949	0.9940
Serbian	0.9843	0.9838	0.9877	0.9877	0.9870
Turkish	0.9824	0.9829	0.9858	0.9854	0.9854

Download

The trained models can be downloaded in the releases section. The each model 7zip archive contains the best scoring model (on development data) out of 5 epochs, alongside with its corresponding vocabulary.

The datasets can be downloaded with the download_data.sh script from the original implementation.

To-Do

• Implementation:
  • LSTM
  • Bi-LSTM
  • CNN

References

• S. Schweter and S. Ahmed, "Deep-EOS: General-Purpose Neural Networks for Sentence Boundary Detection” in Proceedings of the 15th Conference on Natural Language Processing (KONVENS), 2019.