Pywordseg

基於 BiLSTM 及 ELMo 的 State-of-the-art 開源中文斷詞系統。
An open source state-of-the-art Chinese word segmentation system with BiLSTM and ELMo.

• arXiv paper link: https://arxiv.org/abs/1901.05816
• PyPI page: https://pypi.org/project/pywordseg/

Performance

• 此專案提供圖中的 "character level ELMo" model 以及 "baseline" model，其中 "character level ELMo" model 是當前準確率最高。這兩個 model 都贏過目前常用的斷詞系統 Jieba (HMM-based) 及 CKIP (rule-based) 許多。
• This repo provides the "character level ELMo" model and "baseline" model in the figure. Our "character level ELMo" model outperforms the previous state-of-the-art Chinese word segmentation (Ma et al. 2018), and also largely outerform "Jieba" and "CKIP", which are most popular toolkits in processing simplified/traditional Chinese text.

• 當處理訓練時未見過的詞時，"character level ELMo" model 仍然保有不錯的正確率，相較於"baseline" model。
• When considering OOV accuracy, our "character level ELMo" model outperforms our "baseline" model about 5%.

Usage

Requirements

• python >= 3.6 (do not use 3.5)
• pytorch 0.4
• overrides

Install with Pip

• $ pip install pywordseg
• the module will automatically download the models while your first import within 1 minute.
• if you use MacOS and encounter the urllib.error.URLError problem when downloading your models,
  try $ sudo /Applications/Python\ 3.6/Install\ Certificates.command to bypass the certificate issue.

Install manually

• $ git clone https://github.com/voidism/pywordseg
• download ELMoForManyLangs.zip and unzip it to the pywordseg/pywordseg (the code of the ELMo model is from HIT-SCIR, training by myself in character-level)
• $ pip install . under the main directory

Segment!

# import the module
from pywordseg import *

# declare the segmentor.
seg = Wordseg(batch_size=64, device="cuda:0", embedding='elmo', elmo_use_cuda=True, mode="TW")

# input is a list of raw sentences.
seg.cut(["今天天氣真好啊!", "潮水退了就知道，誰沒穿褲子。"])

# will return a list of lists of the segmented sentences.
# [['今天', '天氣', '真', '好', '啊', '!'], ['潮水', '退', '了', '就', '知道', ',', '誰', '沒', '穿', '褲子', '。']]

Parameters:

• batch_size: batch size for the word segmentation model, default: 64.
• device: the CPU/GPU device to run you model, default: 'cpu'.
• embedding: (default: 'w2v')
  • 'elmo': the loaded model will be the "character level ELMo" model above, which runs slow.
  • 'w2v': the loaded model will be the "baseline model" above, which runs faster than 'elmo'.
• elmo_use_cuda: if you want your ELMo model be accelerated on GPU, use True, otherwise the ELMo model will be run on CPU. This param is no use when embedding='w2v'. default: True.
• mode: WordSeg will load different model according to the mode as listed below: (default: TW)
  • TW: trained on AS corpus, from CKIP, Academia Sinica, Taiwan.
  • HK: trained on CityU corpus, from City University of Hong Kong, Hong Kong SAR.
  • CN_MSR: trained on MSR corpus, from Microsoft Research, China.
  • CN_PKU or CN: trained on PKU corpus, from Peking University, China.

Include External Dictionary (Optional)

This feature was inspired by CKIPTagger.

# import the module
from pywordseg import *

# declare the segmentor.
seg = Wordseg(batch_size=64, device="cuda:0", embedding='elmo', elmo_use_cuda=True, mode="TW")

# create dictionary with their relative weights to prioritize.
word_to_weight = {
  "來辦": 2.0,
  "你本人": 1.0,
  "或者是": 1.0,
  "有興趣": 1.0,
  "有興趣的": "2.0",
}
dictionary = construct_dictionary(word_to_weight)
print(dictionary)
# [(2, {'來辦': 2.0}), (3, {'你本人': 1.0, '或者是': 1.0, '有興趣': 1.0}), (4, {'有興趣的': 2.0})]

# 1) segment without dictionary.
seg.cut(["你本人或者是親屬有興趣的話都可以來辦理"])
# [['你', '本人', '或者', '是', '親屬', '有', '興趣', '的話', '都', '可以', '來', '辦理']]

# 2) segment with dictionary to merge words (only merge words that will not break existing words).
seg.cut(["你本人或者是親屬有興趣的話都可以來辦理"], merge_dict=dictionary)
# [['你本人', '或者是', '親屬', '有興趣', '的話', '都', '可以', '來', '辦理']]
# merged: '你', '本人' --> '你本人'
# merged: '或者', '是' --> '或者是'
# merged: '有', '興趣' --> '有興趣'
# not merged: '來', '辦理' -x-> '來辦', '理' because it breaks existing words

# 3) segment with dictionary that force words to be segmented (ignore existing words).
seg.cut(["你本人或者是親屬有興趣的話都可以來辦理"], force_dict=dictionary)
# [['你本人', '或者是', '親屬', '有興趣的', '話', '都', '可以', '來辦', '理']]
# merged: '你', '本人' --> '你本人'
# merged: '或者', '是' --> '或者是'
# change: '有興趣', '的話' --> '有興趣的', '話'
# change: '來', '辦理' --> '來辦', '理'

TODO

• 目前只支援繁體中文(即使選擇CN mode，文字也要轉換成繁體才能運作，目前訓練資料都是經過 OpenCC 轉換的)，日後會加入簡體中文。

Citation

If you use the code in your paper, then please cite it as:

@article{Chuang2019,
  archivePrefix = {arXiv},
  arxivId       = {1901.05816},
  author        = {Chuang, Yung-Sung},
  eprint        = {1901.05816},
  title         = {Robust Chinese Word Segmentation with Contextualized Word Representations},
  url           = {http://arxiv.org/abs/1901.05816},
  year          = {2019}
}