Skip to content
/ Jiayan Public

甲言,专注于古代汉语(古汉语/古文/文言文/文言)处理的NLP工具包,支持文言词库构建、分词、词性标注、断句和标点。Jiayan, the 1st NLP toolkit designed for Classical Chinese, supports lexicon construction, tokenizing, POS tagging, sentence segmentation and punctuation.

License

Notifications You must be signed in to change notification settings

jiaeyan/Jiayan

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

50 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

甲言Jiayan

PyPI License

中文
English

简介

甲言,取「甲骨文言」之意,是一款专注于古汉语处理的NLP工具包。
目前通用的汉语NLP工具多以现代汉语为核心语料,对古代汉语的处理效果并不如人意(详见分词)。本项目的初衷,便是辅助古汉语信息处理,帮助有志于挖掘古文化矿藏的古汉语学者、爱好者等更好地分析和利用文言资料,从「文化遗产」中创造出「文化新产」。
当前版本支持词库构建自动分词词性标注文言句读标点五项功能,更多功能正在开发中。

功能

  • 词库构建
  • 分词
    • 利用无监督、无词典的N元语法隐马尔可夫模型进行古汉语自动分词。
    • 利用词库构建功能产生的文言词典,基于有向无环词图、句子最大概率路径和动态规划算法进行分词。
  • 词性标注
  • 断句
    • 基于字符的条件随机场的序列标注,引入点互信息及t-测试值为特征,对文言段落进行自动断句。
  • 标点
    • 基于字符的层叠式条件随机场的序列标注,在断句的基础上对文言段落进行自动标点。
  • 文白翻译
  • 注意:受语料影响,目前不支持繁体。如需处理繁体,可先用OpenCC将输入转换为简体,再将结果转化为相应繁体(如港澳台等)。

安装

$ pip install jiayan 
$ pip install https://github.com/kpu/kenlm/archive/master.zip

使用

以下各模块的使用方法均来自examples.py

  1. 下载模型并解压:百度网盘,提取码:p0sc

    • jiayan.klm:语言模型,主要用来分词,以及句读标点任务中的特征提取;
    • pos_model:CRF词性标注模型;
    • cut_model:CRF句读模型;
    • punc_model:CRF标点模型;
    • 庄子.txt:用来测试词库构建的庄子全文。
  2. 词库构建

    from jiayan import PMIEntropyLexiconConstructor
    
    constructor = PMIEntropyLexiconConstructor()
    lexicon = constructor.construct_lexicon('庄子.txt')
    constructor.save(lexicon, '庄子词库.csv')
    

    结果:

    Word,Frequency,PMI,R_Entropy,L_Entropy
    之,2999,80,7.944909328101839,8.279435615456894
    而,2089,80,7.354575005231323,8.615211168836439
    不,1941,80,7.244331150611089,6.362131306822925
    ...
    天下,280,195.23602384978196,5.158574399464853,5.24731990592901
    圣人,111,150.0620531154239,4.622606551534004,4.6853474419338585
    万物,94,377.59805590304126,4.5959107835319895,4.538837960294887
    天地,92,186.73504238078462,3.1492586603863617,4.894533538722486
    孔子,80,176.2550051738876,4.284638190120882,2.4056390622295662
    庄子,76,169.26227942514097,2.328252899085616,2.1920058354921066
    仁义,58,882.3468468468468,3.501609497059026,4.96900162987599
    老聃,45,2281.2228260869565,2.384853500510039,2.4331958387289765
    ...
    
  3. 分词

    1. 字符级隐马尔可夫模型分词,效果符合语感,建议使用,需加载语言模型 jiayan.klm

      from jiayan import load_lm
      from jiayan import CharHMMTokenizer
      
      text = '是故内圣外王之道,暗而不明,郁而不发,天下之人各为其所欲焉以自为方。'
      
      lm = load_lm('jiayan.klm')
      tokenizer = CharHMMTokenizer(lm)
      print(list(tokenizer.tokenize(text)))
      

      结果:
      ['是', '故', '内圣外王', '之', '道', ',', '暗', '而', '不', '明', ',', '郁', '而', '不', '发', ',', '天下', '之', '人', '各', '为', '其', '所', '欲', '焉', '以', '自', '为', '方', '。']

      由于古汉语没有公开分词数据,无法做效果评估,但我们可以通过不同NLP工具对相同句子的处理结果来直观感受本项目的优势:

      试比较 LTP (3.4.0) 模型分词结果:
      ['是', '故内', '圣外王', '之', '道', ',', '暗而不明', ',', '郁', '而', '不', '发', ',', '天下', '之', '人', '各', '为', '其', '所', '欲', '焉以自为方', '。']

      再试比较 HanLP 分词结果:
      ['是故', '内', '圣', '外', '王之道', ',', '暗', '而', '不明', ',', '郁', '而', '不', '发', ',', '天下', '之', '人', '各为其所欲焉', '以', '自为', '方', '。']

      可见本工具对古汉语的分词效果明显优于通用汉语NLP工具。

      *更新:感谢HanLP的作者hankc告知——从2021年初,HanLP发布了深度学习驱动的2.x。由于使用了大规模语料上预训练的语言模型,这些语料已经包括了互联网上几乎所有的古汉语和现代汉语,所以在古汉语上的效果已经得到了质的提升。不仅仅是分词,就连词性标注和语义分析也有一定zero-shot learning的效果。相应的具体分词效果请参见该Issue

    2. 词级最大概率路径分词,基本以字为单位,颗粒度较粗

      from jiayan import WordNgramTokenizer
      
      text = '是故内圣外王之道,暗而不明,郁而不发,天下之人各为其所欲焉以自为方。'
      tokenizer = WordNgramTokenizer()
      print(list(tokenizer.tokenize(text)))
      

      结果:
      ['是', '故', '内', '圣', '外', '王', '之', '道', ',', '暗', '而', '不', '明', ',', '郁', '而', '不', '发', ',', '天下', '之', '人', '各', '为', '其', '所', '欲', '焉', '以', '自', '为', '方', '。']

  4. 词性标注

    from jiayan import CRFPOSTagger
    
    words = ['天下', '大乱', ',', '贤圣', '不', '明', ',', '道德', '不', '一', ',', '天下', '多', '得', '一', '察', '焉', '以', '自', '好', '。']
    
    postagger = CRFPOSTagger()
    postagger.load('pos_model')
    print(postagger.postag(words))
    

    结果:
    ['n', 'a', 'wp', 'n', 'd', 'a', 'wp', 'n', 'd', 'm', 'wp', 'n', 'a', 'u', 'm', 'v', 'r', 'p', 'r', 'a', 'wp']

  5. 断句

    from jiayan import load_lm
    from jiayan import CRFSentencizer
    
    text = '天下大乱贤圣不明道德不一天下多得一察焉以自好譬如耳目皆有所明不能相通犹百家众技也皆有所长时有所用虽然不该不遍一之士也判天地之美析万物之理察古人之全寡能备于天地之美称神之容是故内圣外王之道暗而不明郁而不发天下之人各为其所欲焉以自为方悲夫百家往而不反必不合矣后世之学者不幸不见天地之纯古之大体道术将为天下裂'
    
    lm = load_lm('jiayan.klm')
    sentencizer = CRFSentencizer(lm)
    sentencizer.load('cut_model')
    print(sentencizer.sentencize(text))
    

    结果:
    ['天下大乱', '贤圣不明', '道德不一', '天下多得一察焉以自好', '譬如耳目', '皆有所明', '不能相通', '犹百家众技也', '皆有所长', '时有所用', '虽然', '不该不遍', '一之士也', '判天地之美', '析万物之理', '察古人之全', '寡能备于天地之美', '称神之容', '是故内圣外王之道', '暗而不明', '郁而不发', '天下之人各为其所欲焉以自为方', '悲夫', '百家往而不反', '必不合矣', '后世之学者', '不幸不见天地之纯', '古之大体', '道术将为天下裂']

  6. 标点

    from jiayan import load_lm
    from jiayan import CRFPunctuator
    
    text = '天下大乱贤圣不明道德不一天下多得一察焉以自好譬如耳目皆有所明不能相通犹百家众技也皆有所长时有所用虽然不该不遍一之士也判天地之美析万物之理察古人之全寡能备于天地之美称神之容是故内圣外王之道暗而不明郁而不发天下之人各为其所欲焉以自为方悲夫百家往而不反必不合矣后世之学者不幸不见天地之纯古之大体道术将为天下裂'
    
    lm = load_lm('jiayan.klm')
    punctuator = CRFPunctuator(lm, 'cut_model')
    punctuator.load('punc_model')
    print(punctuator.punctuate(text))
    

    结果:
    天下大乱,贤圣不明,道德不一,天下多得一察焉以自好,譬如耳目,皆有所明,不能相通,犹百家众技也,皆有所长,时有所用,虽然,不该不遍,一之士也,判天地之美,析万物之理,察古人之全,寡能备于天地之美,称神之容,是故内圣外王之道,暗而不明,郁而不发,天下之人各为其所欲焉以自为方,悲夫!百家往而不反,必不合矣,后世之学者,不幸不见天地之纯,古之大体,道术将为天下裂。

版本

  • v0.0.21
    • 将安装过程分为两步,确保得到最新的kenlm版本。
  • v0.0.2
    • 增加词性标注功能。
  • v0.0.1
    • 词库构建、自动分词、文言句读、标点功能开放。

Introduction

Jiayan, which means Chinese characters engraved on oracle bones, is a professional Python NLP tool for Classical Chinese.
Prevailing Chinese NLP tools are mainly trained on modern Chinese data, which leads to bad performance on Classical Chinese (See Tokenizing). The purpose of this project is to assist Classical Chinese information processing.
Current version supports lexicon construction, tokenizing, POS tagging, sentence segmentation and automatic punctuation, more features are in development.

Features

  • Lexicon Construction
    • With an unsupervised approach, construct lexicon with Trie -tree, PMI (point-wise mutual information) and neighboring entropy of left and right characters.
  • Tokenizing
    • With an unsupervised, no dictionary approach to tokenize a Classical Chinese sentence with N-gram language model and HMM (Hidden Markov Model).
    • With the dictionary produced from lexicon construction, tokenize a Classical Chinese sentence with Directed Acyclic Word Graph, Max Probability Path and Dynamic Programming.
  • POS Tagging
    • Word level sequence tagging with CRF (Conditional Random Field). See POS tag categories here.
  • Sentence Segmentation
    • Character level sequence tagging with CRF, introduces PMI and T-test values as features.
  • Punctuation
    • Character level sequence tagging with layered CRFs, punctuate given Classical Chinese texts based on results of sentence segmentation.
  • Note: Due to data we used, we don't support traditional Chinese for now. If you have to process traditional one, please use OpenCC to convert traditional input to simplified, then you could convert the results back.

Installation

$ pip install jiayan 
$ pip install https://github.com/kpu/kenlm/archive/master.zip

Usages

The usage codes below are all from examples.py.

  1. Download the models and unzip them:Google Drive

    • jiayan.klm:the language model used for tokenizing and feature extraction for sentence segmentation and punctuation;
    • pos_model:the CRF model for POS tagging;
    • cut_model:the CRF model for sentence segmentation;
    • punc_model:the CRF model for punctuation;
    • 庄子.txt:the full text of 《Zhuangzi》 used for testing lexicon construction.
  2. Lexicon Construction

    from jiayan import PMIEntropyLexiconConstructor
    
    constructor = PMIEntropyLexiconConstructor()
    lexicon = constructor.construct_lexicon('庄子.txt')
    constructor.save(lexicon, 'Zhuangzi_Lexicon.csv')
    

    Result:

    Word,Frequency,PMI,R_Entropy,L_Entropy
    之,2999,80,7.944909328101839,8.279435615456894
    而,2089,80,7.354575005231323,8.615211168836439
    不,1941,80,7.244331150611089,6.362131306822925
    ...
    天下,280,195.23602384978196,5.158574399464853,5.24731990592901
    圣人,111,150.0620531154239,4.622606551534004,4.6853474419338585
    万物,94,377.59805590304126,4.5959107835319895,4.538837960294887
    天地,92,186.73504238078462,3.1492586603863617,4.894533538722486
    孔子,80,176.2550051738876,4.284638190120882,2.4056390622295662
    庄子,76,169.26227942514097,2.328252899085616,2.1920058354921066
    仁义,58,882.3468468468468,3.501609497059026,4.96900162987599
    老聃,45,2281.2228260869565,2.384853500510039,2.4331958387289765
    ...
    
  3. Tokenizing

    1. The character based HMM, recommended, needs language model: jiayan.klm

      from jiayan import load_lm
      from jiayan import CharHMMTokenizer
      
      text = '是故内圣外王之道,暗而不明,郁而不发,天下之人各为其所欲焉以自为方。'
      
      lm = load_lm('jiayan.klm')
      tokenizer = CharHMMTokenizer(lm)
      print(list(tokenizer.tokenize(text)))
      

      Result:
      ['是', '故', '内圣外王', '之', '道', ',', '暗', '而', '不', '明', ',', '郁', '而', '不', '发', ',', '天下', '之', '人', '各', '为', '其', '所', '欲', '焉', '以', '自', '为', '方', '。']

      Since there is no public tokenizing data for Classical Chinese, it's hard to do performance evaluation directly; However, we can compare the results with other popular modern Chinese NLP tools to check the performance:

      Compare the tokenizing result of LTP (3.4.0):
      ['是', '故内', '圣外王', '之', '道', ',', '暗而不明', ',', '郁', '而', '不', '发', ',', '天下', '之', '人', '各', '为', '其', '所', '欲', '焉以自为方', '。']

      Also, compare the tokenizing result of HanLP:
      ['是故', '内', '圣', '外', '王之道', ',', '暗', '而', '不明', ',', '郁', '而', '不', '发', ',', '天下', '之', '人', '各为其所欲焉', '以', '自为', '方', '。']

      It's apparent that Jiayan has much better tokenizing performance than general Chinese NLP tools.

    2. Max probability path approach tokenizing based on words

      from jiayan import WordNgramTokenizer
      
      text = '是故内圣外王之道,暗而不明,郁而不发,天下之人各为其所欲焉以自为方。'
      tokenizer = WordNgramTokenizer()
      print(list(tokenizer.tokenize(text)))
      

      Result:
      ['是', '故', '内', '圣', '外', '王', '之', '道', ',', '暗', '而', '不', '明', ',', '郁', '而', '不', '发', ',', '天下', '之', '人', '各', '为', '其', '所', '欲', '焉', '以', '自', '为', '方', '。']

  4. POS Tagging

    from jiayan import CRFPOSTagger
    
    words = ['天下', '大乱', ',', '贤圣', '不', '明', ',', '道德', '不', '一', ',', '天下', '多', '得', '一', '察', '焉', '以', '自', '好', '。']
    
    postagger = CRFPOSTagger()
    postagger.load('pos_model')
    print(postagger.postag(words))
    

    Result:
    ['n', 'a', 'wp', 'n', 'd', 'a', 'wp', 'n', 'd', 'm', 'wp', 'n', 'a', 'u', 'm', 'v', 'r', 'p', 'r', 'a', 'wp']

  5. Sentence Segmentation

    from jiayan import load_lm
    from jiayan import CRFSentencizer
    
    text = '天下大乱贤圣不明道德不一天下多得一察焉以自好譬如耳目皆有所明不能相通犹百家众技也皆有所长时有所用虽然不该不遍一之士也判天地之美析万物之理察古人之全寡能备于天地之美称神之容是故内圣外王之道暗而不明郁而不发天下之人各为其所欲焉以自为方悲夫百家往而不反必不合矣后世之学者不幸不见天地之纯古之大体道术将为天下裂'
    
    lm = load_lm('jiayan.klm')
    sentencizer = CRFSentencizer(lm)
    sentencizer.load('cut_model')
    print(sentencizer.sentencize(text))
    

    Result:
    ['天下大乱', '贤圣不明', '道德不一', '天下多得一察焉以自好', '譬如耳目', '皆有所明', '不能相通', '犹百家众技也', '皆有所长', '时有所用', '虽然', '不该不遍', '一之士也', '判天地之美', '析万物之理', '察古人之全', '寡能备于天地之美', '称神之容', '是故内圣外王之道', '暗而不明', '郁而不发', '天下之人各为其所欲焉以自为方', '悲夫', '百家往而不反', '必不合矣', '后世之学者', '不幸不见天地之纯', '古之大体', '道术将为天下裂']

  6. Punctuation

    from jiayan import load_lm
    from jiayan import CRFPunctuator
    
    text = '天下大乱贤圣不明道德不一天下多得一察焉以自好譬如耳目皆有所明不能相通犹百家众技也皆有所长时有所用虽然不该不遍一之士也判天地之美析万物之理察古人之全寡能备于天地之美称神之容是故内圣外王之道暗而不明郁而不发天下之人各为其所欲焉以自为方悲夫百家往而不反必不合矣后世之学者不幸不见天地之纯古之大体道术将为天下裂'
    
    lm = load_lm('jiayan.klm')
    punctuator = CRFPunctuator(lm, 'cut_model')
    punctuator.load('punc_model')
    print(punctuator.punctuate(text))
    

    Result:
    天下大乱,贤圣不明,道德不一,天下多得一察焉以自好,譬如耳目,皆有所明,不能相通,犹百家众技也,皆有所长,时有所用,虽然,不该不遍,一之士也,判天地之美,析万物之理,察古人之全,寡能备于天地之美,称神之容,是故内圣外王之道,暗而不明,郁而不发,天下之人各为其所欲焉以自为方,悲夫!百家往而不反,必不合矣,后世之学者,不幸不见天地之纯,古之大体,道术将为天下裂。

Versions

  • v0.0.21
    • Divide the installation into two steps to ensure to get the latest version of kenlm.
  • v0.0.2
    • POS tagging feature is open.
  • v0.0.1
    • Add features of lexicon construction, tokenizing, sentence segmentation and automatic punctuation.

About

甲言,专注于古代汉语(古汉语/古文/文言文/文言)处理的NLP工具包,支持文言词库构建、分词、词性标注、断句和标点。Jiayan, the 1st NLP toolkit designed for Classical Chinese, supports lexicon construction, tokenizing, POS tagging, sentence segmentation and punctuation.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages