Created by Jingkang Wang*, Jianing Zhou*，Jie Zhou and Gongshen Liu.
In this paper, we introduce multiple character embeddings including Pinyin Romanization and Wubi Input, both of which are easily accessible and effective in depicting semantics of characters. To fully leverage them, we propose a novel shared Bi-LSTM-CRF model, which fuses multiple features efficiently. Extensive experiments on five corpora demonstrate that extra embeddings help obtain a significant improvement. Specifically, we achieves the state-of-the-art performance in AS and CITYU corpora with F1 scores 96.9 and 97.3, respectively.
In this repository, we release code and data for reproducing the results given in the paper (ACL-SRW 2019).
- Python 2.7 or 3.5+
- Tensorflow 1.2+
- CUDA 8.0+ (For GPU)
- Python Libraries: numpy, cPickle, pypinyin
If you find our work is helpful and relevant, please consider citing
@inproceedings{multi-embed2019,
author = {Jianing Zhou and
Jingkang Wang and
Jie Zhou and
Gongshen Liu},
title = {Multiple Character Embeddings for Chinese Word Segmentation},
booktitle = {{ACL} {(2)}},
pages = {210--216},
publisher = {Association for Computational Linguistics},
year = {2019}
}
python preprocess.py --rootDir Corpora --corpusAll Corpora/all.txt --resultFile pre_chars_for_w2v.txt
python getpinyin.py
python getwubi.py
If you want to use your own data, please replace Corpora to the path of your corpus. Run python preprocess.py -h to see more details.
./third_party/word2vec -train pre_chars_for_w2v.txt -save-vocab pre_vocab.txt -min-count 3
./third_party/word2vec -train pre_pinyin_for_w2v.txt -save-vocab pre_vocab_pinyin.txt -min-count 3
./third_party/word2vec -train pre_wubi_for_w2v.txt -save-vocab pre_vocab_wubi.txt -min-count 3
python SentHandler/replace_unk.py pre_vocab.txt pre_chars_for_w2v.txt chars_for_w2v.txt
python SentHandler/replace_unk.py pre_vocab_pinyin.txt pre_pinyin_for_w2v.txt pinyin_for_w2v.txt
python SentHandler/replace_unk.py pre_vocab_wubi.txt pre_wubi_for_w2v.txt wubi_for_w2v.txt
./third_party/word2vec -train chars_for_w2v.txt -output char_vec.txt -size 256 -sample 1e-4 -negative 0 -hs 1 -binary 0 -iter 5
./third_party/word2vec -train pinyin_for_w2v.txt -output pinyin_vec.txt -size 256 -sample 1e-4 -negative 0 -hs 1 -binary 0 -iter 5
./third_party/word2vec -train wubi_for_w2v.txt -output wubi_vec.txt -size 256 -sample 1e-4 -negative 0 -hs 1 -binary 0 -iter 5
./third_party/word2vec -train chars_for_w2v.txt -output char_vec300.txt -size 300 -sample 1e-4 -negative 0 -hs 1 -binary 0 -iter 5
./third_party/word2vec -train pinyin_for_w2v.txt -output pinyin_vec300.txt -size 300 -sample 1e-4 -negative 0 -hs 1 -binary 0 -iter 5
./third_party/word2vec -train wubi_for_w2v.txt -output wubi_vec300.txt -size 300 -sample 1e-4 -negative 0 -hs 1 -binary 0 -iter 5
First, the file word2vec.c in third_party directory should be compiled (see third_party/compile_w2v.sh). Also, you could directly add executive permission to the third_party/word2vec using chmod +x third_party/word2vec. Second, the word2vec tool counts the characters which have a frequency more than 3 and saves them into file pre_vocab.txt. After that, the scripts replace the words that are not in pre_vocab.txt with "UNK". Finally, word2vec training process begins.
python pre_train.py --corpusAll Corpora/msr/train-all.txt --char_vecpath char_vec.txt --pinyin_vecpath pinyin_vec.txt --wubi_vecpath wubi_vec.txt --train_file Corpora/msr/ --test_file Corpora/msr/ --test_file_raw Corpora/msr/test_raw.txt --test_file_gold Corpora/msr/test_gold.txt
python pre_train.py --corpusAll Corpora/msr300/train-all.txt --char_vecpath char_vec300.txt --pinyin_vecpath pinyin_vec300.txt --wubi_vecpath wubi_vec300.txt --train_file Corpora/msr300/ --test_file Corpora/msr300/ --test_file_raw Corpora/msr300/test_raw.txt --test_file_gold Corpora/msr300/test_gold.txt
To see HELP for the training script:
python pre_train.py -h
python ./CWSTrain/fc_lstm3_crf_train.py --train_data_path Corpora/msr --test_data_path Corpora/msr --word2vec_path char_vec.txt --pinyin2vec_path pinyin_vec.txt --wubi2vec_path wubi_vec.txt --log_dir Logs_fc_lstm3/msr --embedding_size 256 --batch_size 256
python ./CWSTrain/pw_lstm3_crf_train.py --train_data_path Corpora/msr --test_data_path Corpora/msr --word2vec_path char_vec.txt --pinyin2vec_path pinyin_vec.txt --wubi2vec_path wubi_vec.txt --log_dir Logs_pw_lstm3/msr --embedding_size 256 --batch_size 256
python ./CWSTrain/share_lstm3_crf_train.py --train_data_path Corpora/msr --test_data_path Corpora/msr --word2vec_path char_vec.txt --pinyin2vec_path pinyin_vec.txt --wubi2vec_path wubi_vec.txt --log_dir Logs_share_lstm3/msr --embedding_size 256 --batch_size 256
python ./CWSTrain/nopy_fc_lstm3_crf_train.py --train_data_path Corpora/msr --test_data_path Corpora/msr --word2vec_path char_vec.txt --wubi2vec_path wubi_vec.txt --log_dir Logs_nopy/msr --embedding_size 256 --batch_size 256*
python ./CWSTrain/nowubi_fc_lstm3_crf_train.py --train_data_path Corpora/msr --test_data_path Corpora/msr --word2vec_path char_vec.txt --pinyin2vec_path pinyin_vec.txt --log_dir Logs_nowubi/msr --embedding_size 256 --batch_size 256
If you want to train on other corpora, please change the train_data_path, test_data_path and make a new log directory. Arguments of *_lstm*_crf_train.py are set by tf.app.flags.
python tools/freeze_graph.py --input_graph Logs_fc_lstm3/msr/graph.pbtxt --input_checkpoint Logs_fc_lstm3/msr/model.ckpt --output_node_names "input_placeholder_char,input_placeholder_pinyin,input_placeholder_wubi,transitions,Reshape_11" --output_graph Models/fc_lstm3_crf_model_msr.pbtxt
python tools/freeze_graph.py --input_graph Logs_nopy/msr/graph.pbtxt --input_checkpoint Logs_nopy/msr/model.ckpt --output_node_names "input_placeholder_char,input_placeholder_wubi,transitions,Reshape_11" --output_graph Models/nopy_fc_lstm3_crf_model_msr.pbtxt
python tools/freeze_graph.py --input_graph Logs_nowubi/msr/graph.pbtxt --input_checkpoint Logs_nowubi/msr/model.ckpt --output_node_names "input_placeholder_char,input_placeholder_pinyin,transitions,Reshape_11" --output_graph Models/nowubi_fc_lstm3_crf_model_msr.pbtxt
Build model for segmentation.
python tools/vob_dump.py --char_vecpath char_vec.txt --pinyin_vecpath pinyin_vec.txt --wubi_vecpath wubi_vec.txt --char_dump_path Models/char_dump.pk --pinyin_dump_path Models/pinyin_dump.pk --wubi_dump_path Models/wubi_dump.pk
python tools/vob_dump.py --char_vecpath char_vec300.txt --pinyin_vecpath pinyin_vec300.txt --wubi_vecpath wubi_vec300.txt --char_dump_path Models/char_dump300.pk --pinyin_dump_path Models/pinyin_dump300.pk --wubi_dump_path Models/wubi_dump300.pk
Note that this step is neccessary for the seg model.
Use file tools/crf_seg.py to segment words utilizing pre-trained models. You could refer to this file for detailed parameter configurations.
python tools/crf_seg.py --test_data Corpora/msr/test_raw.txt --model_path Models/fc_lstm3_crf_model_msr.pbtxt --result_path Results/crf_result_msr.txt
python tools/fc_lstm3_crf_seg_nopy.py --test_data Corpora/msr/test_raw.txt --model_path Models/nopy_fc_lstm3_crf_model_msr.pbtxt --result_path Results/nopy_crf_result_msr.txt
python tools/fc_lstm3_crf_seg_nowubi.py --test_data Corpora/msr/test_raw.txt --model_path Models/nowubi_fc_lstm3_crf_model_msr.pbtxt --result_path Results/nowubi_crf_result_msr.txt
python PRF_Score.py Results/crf_result_msr.txt Corpora/msr/test_gold.txt
Result files are put in directory Results/.
This code is based on the this repo (LSTM-CNN-CWS). Many thanks to the author.
Our code is released under MIT License.