hankcs / sub-character-cws

Sub-Character Representation Learning

Sub-Character Representation Learning

Codes and corpora for paper "Dual Long Short-Term Memory Networks for Sub-Character Representation Learning" (accepted at ITNG 2018).

We proposed to learn character and sub-character level representations jointly for capturing deeper level of semantic meanings. When applied to Chinese Word Segmentation as a case example, our solution achieved state-of-the-art results on both Simplified and Traditional Chinese, without extra Traditional to Simplified Chinese conversion.

Dependencies

• Python >= 3
• DyNet==2.0.1

Quick Start

Simply run one command:

./script/run.sh pku 1

It does everything for you on the fly, including data preparation, training and test.

• You can replace pku with msr, cityu and as.
• The second parameter indicates model options from 1 to 6, details are listed in the next chapter.

Configuration Table

We have presented 6 models in our paper. Their configurations are shown in following table:

#. model	char	subchar	radical	tie	bigram
1. baseline	YES
2. +subchar		YES
3. +radical		YES	YES
4. +radical -char			YES
5. +radical +tie		YES	YES	YES
6. +radical +tie +bigram		YES	YES	YES	YES

Performance

Acknowledgments

• Thanks for those friends who helped us with the experiments.
• Corpora are from SIGHAN05, which should only be used for research purposes.
• Model implementation modified from a Dynet-1.x version by rguthrie3.