Skip to content


Folders and files

Last commit message
Last commit date

Latest commit



55 Commits

Repository files navigation

Deep Learning NLP Pipeline implemented on Tensorflow. Following the 'simplicity' rule, this project aims to use the deep learning library of Tensorflow to implement new NLP pipeline. You can extend the project to train models with your own corpus/languages. Pretrained models of Chinese corpus are distributed. Free RESTful NLP API are also provided. Visit for details.

Brief Introduction


  • NLP Pipeline Modules:

    • Word Segmentation/Tokenization
    • Part-of-speech (POS)
    • Named-entity-recognition(NER)
    • Dependency Parsing (Parse)
    • textsum: automatic summarization Seq2Seq-Attention models
    • textrank: extract the most important sentences
    • textcnn: document classification
    • Web API: Free Tensorflow empowered web API
    • Planed: Automatic Summarization
  • Algorithm(Closely following the state-of-Art)

    • Word Segmentation: Linear Chain CRF(conditional-random-field), based on python CRF++ module
    • POS: LSTM/BI-LSTM/LSTM-CRF network, based on Tensorflow
    • NER: LSTM/BI-LSTM/LSTM-CRF network, based on Tensorflow
    • Parse: Arc-Standard System with Feed Forward Neural Network
    • Textsum: Seq2Seq with attention mechanism
    • Texncnn: CNN
  • Pre-trained Model

    • Chinese: Segmentation, POS, NER, Parse (1998 china daily corpus)
    • Domain Specific NER Models are also provided: general, entertainment, o2o, etc... Contribution are welcome
    • English: POS (brown corpus)
    • For your Specific Language, you can easily use the script to train model with the corpus of your language choice.


  • Requirements

    • CRF++ (>=0.54)
    • Tensorflow(1.4)
    • Python (python2.7 and python3.6 are tested) This project is up to date with the latest tensorflow release. For tensorflow (<=0.12.0), use deepnlp <=0.1.5 version. tensorflow (1.0-1.3), use deepnlp = 0.1.6 version tensorflow (1.4), use deepnlp = 0.1.7 version See for more details
  • Pip

    # linux, run the script:
    pip install deepnlp

Due to pkg size restriction, english pos model, ner domain specific model files are not distributed on pypi You can download the pre-trained model files from github and put in your installation directory .../site-packages/.../deepnlp/... model files: ../pos/ckpt/en/pos.ckpt ; ../ner/ckpt/zh/ner.ckpt

    # linux, run the script:
    tar zxvf deepnlp-0.1.7.tar.gz
    cd deepnlp-0.1.7
    python install
  • Initial setup
    # install crf++0.58 package using the script
    sh ./deepnlp/segment/
    # Download all the pre-trained models
    python ./test/
    # Or Download pre-trained models from below command lines
    import deepnlp'segment')'pos')'ner')'parse')
  • Running Examples
    # ./deepnlp/test folder
    cd test
    python    # segmentation
    python       # POS tag
    python       # NER Zh
    python   # NER domain-specific models
    python # NER load user dict and UDF for disambiguation
    python    # dependency parsing


Set Coding

设置编码 For python2, the default coding is ascii not unicode, use future module to make it compatible with python3

from __future__ import unicode_literals # compatible with python3 unicode

Download pretrained models

下载预训练模型 If you install deepnlp via pip, the pre-trained models are not distributed due to size restriction. You can download full models for 'Segment', 'POS' en and zh, 'NER' zh, zh_entertainment, zh_o2o, 'Textsum' by calling the download function.

import deepnlp
# Download all the modules

# Download specific module'segment')'pos')'ner')'parse')

# Download module and domain-specific model = 'pos', name = 'en') = 'ner', name = 'zh_entertainment')



from __future__ import unicode_literals
from deepnlp import segmenter

tokenizer = segmenter.load_model(name = 'zh_entertainment')
text = "我刚刚在浙江卫视看了电视剧老九门,觉得陈伟霆很帅"
segList = tokenizer.seg(text)
text_seg = " ".join(segList)

# 我 刚刚 在 浙江卫视 看 了 电视剧 老九门 , 觉得 陈伟霆 很 帅



from __future__ import unicode_literals

import deepnlp'pos')

## English Model
from deepnlp import pos_tagger
tagger = pos_tagger.load_model(name = 'en')  # Loading English model, lang code 'en', English Model Brown Corpus

text = "I want to see a funny movie"
words = text.split(" ")     # unicode
print (" ".join(words))

tagging = tagger.predict(words)
for (w,t) in tagging:
    pair = w + "/" + t
    print (pair)
#I/nn want/vb to/to see/vb a/at funny/jj movie/nn

## Chinese Model
from deepnlp import segmenter
from deepnlp import pos_tagger
tagger = pos_tagger.load_model(name = 'zh') # Loading Chinese model, lang code 'zh', China Daily Corpus

text = "我爱吃北京烤鸭"
words = segmenter.seg(text) # words in unicode coding
print (" ".join(words))

tagging = tagger.predict(words)  # input: unicode coding
for (w,t) in tagging:
    pair = w + "/" + t
    print (pair)

#我/r 爱/v 吃/v 北京/ns 烤鸭/n



from __future__ import unicode_literals   # compatible with python3 unicode

import deepnlp'ner')  # download the NER pretrained models from github if installed from pip

from deepnlp import ner_tagger

# Example: Entertainment Model
tagger = ner_tagger.load_model(name = 'zh_entertainment')   # Base LSTM Based Model
#Load Entertainment Dict
text = "你 最近 在 看 胡歌 演的 猎场 吗 ?"
words = text.split(" ")
tagset_entertainment = ['actor', 'role_name', 'teleplay', 'teleplay_tag']
tagging = tagger.predict(words, tagset = tagset_entertainment)
for (w,t) in tagging:
    pair = w + "/" + t
    print (pair)




from __future__ import unicode_literals # compatible with python3 unicode coding

from deepnlp import nn_parser
parser = nn_parser.load_model(name = 'zh')

#Example 1, Input Words and Tags Both
words = ['它', '熟悉', '一个', '民族', '的', '历史']
tags = ['r', 'v', 'm', 'n', 'u', 'n']

dep_tree = parser.predict(words, tags)

#Fetch result from Transition Namedtuple
num_token = dep_tree.count()
print ("id\tword\tpos\thead\tlabel")
for i in range(num_token):
    cur_id = int(dep_tree.tree[i+1].id)
    cur_form = str(dep_tree.tree[i+1].form)
    cur_pos = str(dep_tree.tree[i+1].pos)
    cur_head = str(dep_tree.tree[i+1].head)
    cur_label = str(dep_tree.tree[i+1].deprel)
    print ("%d\t%s\t%s\t%s\t%s" % (cur_id, cur_form, cur_pos, cur_head, cur_label))

# Result
id	word	pos	head	label
1		r	2	SBV
2	熟悉	v	0	HED
3	一个	m	4	QUN
4	民族	n	5	DE
5		u	6	ATT
6	历史	n	2	VOB


from __future__ import unicode_literals

from deepnlp import pipeline
p = pipeline.load_model('zh')

text = "我爱吃北京烤鸭"
res = p.analyze(text)

print (res[0].encode('utf-8'))
print (res[1].encode('utf-8'))
print (res[2].encode('utf-8'))

words = p.segment(text)
pos_tagging = p.tag_pos(words)
ner_tagging = p.tag_ner(words)

print (pos_tagging.encode('utf-8'))
print (ner_tagging.encode('utf-8'))



See details: README



See details: README



Train your model

自己训练模型 ###Segment model See instructions: README

###POS model See instructions: README

###NER model See instructions: README

###Parsing model See instructions: README

###Textsum model See instructions: README

Web API Service provides free web API service for common NLP modules of sentences and paragraphs. The APIs are RESTful and based on pre-trained tensorflow models. Chinese language is now supported.

Testing API from Browser, Need to log in first


Calling API from python

See ./deepnlp/test/ for more details.

from __future__ import unicode_literals

import json, requests, sys, os
if (sys.version_info>(3,0)): from urllib.parse import quote 
else : from urllib import quote

from deepnlp import api_service
login = api_service.init()          # registration, if failed, load default empty login {} with limited access
conn = api_service.connect(login)   # save the connection with login cookies

# Sample URL

# Define text and language
text = ("我爱吃上海小笼包").encode("utf-8")  # convert text from unicode to utf-8 bytes

# Set up URL for POS tagging
url_pos = '"+ "lang=" + quote('zh') + "&text=" + quote(text)
web = requests.get(url_pos, cookies = conn)
tuples = json.loads(web.text)
print (tuples['pos_str'].encode('utf-8'))    # POS json {'pos_str', 'w1/t1 w2/t2'} return string


deepnlp项目是基于Tensorflow平台的一个python版本的NLP套装, 目的在于将Tensorflow深度学习平台上的模块,结合 最新的一些算法,提供NLP基础模块的支持,并支持其他更加复杂的任务的拓展,如生成式文摘等等。

  • NLP 套装模块

    • 分词 Word Segmentation/Tokenization
    • 词性标注 Part-of-speech (POS)
    • 命名实体识别 Named-entity-recognition(NER)
    • 依存句法分析 Dependency Parsing (Parse)
    • 自动生成式文摘 Textsum (Seq2Seq-Attention)
    • 关键句子抽取 Textrank
    • 文本分类 Textcnn (WIP)
    • 可调用 Web Restful API
    • 计划中: 句法分析 Parsing
  • 算法实现

    • 分词: 线性链条件随机场 Linear Chain CRF, 基于CRF++包来实现
    • 词性标注: 单向LSTM/ 双向BI-LSTM, 基于Tensorflow实现
    • 命名实体识别: 单向LSTM/ 双向BI-LSTM/ LSTM-CRF 结合网络, 基于Tensorflow实现
    • 依存句法分析: 基于arc-standard system的神经网络的parser
  • 预训练模型

    • 中文: 基于人民日报语料和微博混合语料: 分词, 词性标注, 实体识别

API 服务 出于技术交流的目的, 提供免费API接口供文本和篇章进行深度学习NLP的分析, 简单注册后就可以使用。 API符合RESTful风格, 内部是基于tensorflow预先训练好的深度学习模型。具体使用方法请参考博客:



  • 需要

    • CRF++ (>=0.54) 可以从 下载安装
    • Tensorflow(1.0) 这个项目的Tensorflow函数会根据最新Release更新,目前支持Tensorflow 1.0版本,对于老版本的Tensorflow(<=0.12.0), 请使用 deepnlp <=0.1.5版本, 更多信息请查看
  • Pip 安装

    pip install deepnlp
    # linux, run the script:
    tar zxvf deepnlp-0.1.7.tar.gz
    cd deepnlp-0.1.7
    python install
  • 初始设置
    # 运行脚本安装 crf++0.58 包
    sh ./deepnlp/segment/
    # 运行脚本下载预训练模型测试
    python ./test/



Deep Learning NLP Pipeline implemented on Tensorflow







No packages published