先看一下语言模型的输出格式 \data\ ngram = ngram = ngram = \-grams: -5.24036 'cause -0.2084827 -4.675221 'em -0.221857 -4.989297 'n -0.05809768 -5.365303 'til -0.1855581 -2.111539 </s> 0.0 - <s> -0.7736475 -1.128404 <unk> -0.8049794 -2.271447 a -0.616

先放上一张Demo的测试图 测试的句子及每个分词的词性标注为:   目前/t 这/rzv 条/q 高速公路/n 之间/f 的/ude1 路段/n 已/d 紧急/a 封闭/v ./w 需要基础知识 HMM模型(隐马尔可夫模型) 模型的定义 隐马尔科夫模型(hidden Markov model)是关于时序的概率模型,是最简单的动态贝叶斯网络 模型的参数 HMM模型由Pi.A.B 唯一决定   Pi.A.B 成为HMM模型的三要素 HMM用在词性标注问题 对于下面这句话可以用HMM进行建模 目前/t

关于n-gram 语言模型,大部分在这篇博客里 记过了,   SRILM 语言模型格式解读 , 其实看完了,ngram的大概用法都比较清楚了, 但是关于平滑算法,一直很模糊,就晓得一个"劫富济贫" ,也不知 回退 ,插值,折扣,平滑,都说的什么东西,模模糊糊的,找了很多资料,还是看官方文档吧,看具体公式,就明白了.   看全部翻译 参考 :   Ngram 折扣平滑算法 ,本文里夹带着自己的一些理解.  本文档翻译自 srilm 手册ngram-discount.7.html NAME

本文转载自:http://www.17bigdata.com/97-5%E5%87%86%E7%A1%AE%E7%8E%87%E7%9A%84%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0%E4%B8%AD%E6%96%87%E5%88%86%E8%AF%8D%EF%BC%88%E5%AD%97%E5%B5%8C%E5%85%A5bi-lstmcrf%EF%BC%89.html 摘要 深度学习当前在NLP领域发展也相当快,翻译,问答,摘要等基本都被深度学习占领了. 本文

http://blog.csdn.net/marising/article/details/5769653 前段时间写了中文分词的一些记录里面提到了CRF的分词方法,近段时间又研究了一下,特把方法写下来,以备忘,另外,李沫南同学优化过CRF++,见:http://www.coreseek.cn/opensource/CRF/.我觉得CRF++还有更大的优化空间,以后有时间再搞. 人民日报语料是分好词的,我下面贴出的代码就是把语料整理为CRF需要的训练数据,直接修改模板训练即可.不过有下面的同学给

文章摘自github,本次测试选用 HanLP 1.6.0 , LTP 3.4.0 测试思路 使用同一份语料训练两个分词库,同一份测试数据测试两个分词库的性能. 语料库选取1998年01月的人民日报语料库.199801人民日报语料 该词库带有词性标注,为了遵循LTP的训练数据集格式,需要处理掉词性标注. 测试数据选择SIGHan2005提供的开放测试集. SIGHan2005的使用可以参见其附带的readme. HanLP java -cp libs/hanlp-1.6.0.jar com.ha

本章是接前两篇<分词工具Hanlp基于感知机的中文分词框架>和<基于结构化感知机的词性标注与命名实体识别框架>的.本系统将同时进行中文分词.词性标注与命名实体识别3个任务的子系统称为“词法分析器”. 加载 对应的类为PerceptronLexicalAnalyzer,其构造方法为递增的3个模型地址: l public PerceptronLexicalAnalyzer(String cwsModelFile) throws IOException l public Perceptr

 结构化感知机标注框架是一套利用感知机做序列标注任务,并且应用到中文分词.词性标注与命名实体识别这三个问题的完整在线学习框架,该框架利用1个算法解决3个问题,时自治同意的系统,同时三个任务顺序渐进,构成流水线式的系统.本文先介绍中文分词框架部分内容. 中文分词 训练 只需指定输入语料的路径(单文档时为文件路径,多文档时为文件夹路径,灵活处理),以及模型保存位置即可: 命令行 java -cp hanlp.jar com.hankcs.hanlp.model.perceptron.Main -ta

这是另一套基于CRF的词法分析系统,类似感知机词法分析器,提供了完善的训练与分析接口. CRF的效果比感知机稍好一些,然而训练速度较慢,也不支持在线学习. 默认模型训练自OpenCorpus/pku98/199801.txt,随hanlp 1.6.2以上版本发布. 语料格式等与感知机词法分析器相同,请先阅读<感知机词法分析器>. 中文分词 训练 CRFSegmenter segmenter = new CRFSegmenter(null); segmenter.train("data

语料来自1998年1月份人民日报语料 1 语料处理 1.1 原始语料数据格式 语料中,句子已经被分词好,并且在人名后以“/”标注了“nr”表示是人名,其他非人名的分词没有进行标注 1.2 CRF++要求语料的格式 训练语料至少应具有两列,列间由空格或制表位间隔,且所有行(空行除外)必须具有相同的列数,句子间使用空行间隔 1.3 对原始数据进行处理 CRF++可以有多个特征,举例如下图 本次实验为了熟悉采用CRF++及进行中文人名标注,故将语料中的每一个单字作为特征,并进行BIEO标注,举例如下图

http://x-algo.cn/index.php/2016/02/29/crf-name-entity-recognition/ 类似使用CRF实现分词和词性标注,地域识别也是需要生成相应的tag进行标注.这里使用的语料库是1998年1月人民日报语料集.最终学习出来的模型,对复杂的地名识别准确率(F值)非常低,推测是预料中对地名的标注多处是前后矛盾.例如  [华南/ns 地区/n]ns  标为地名实体,但是 东北/f 地区/n 确分开标注,类似错误还有很多.将来有时间可以考虑使用微软的词库 

用CRF做命名实体识别(二) 用CRF做命名实体识别(三) 用BILSTM-CRF做命名实体识别 博客园的markdown格式可能不太方便看,也欢迎大家去我的简书里看 摘要 本文主要讲述了关于人民日报标注语料的预处理,利用CRF++工具包对模型进行训练以及测试 目录 明确我们的标注任务 语料和工具 数据预处理 1.数据说明 2.数据预处理 模型训练及测试 1.流程 2.标注集 3.特征模板 4.CRF++包的使用说明 总结与展望 正文 1.明确我们的标注任务 这篇文章主要是介绍用CRF模型去提取

在学习NLP过程中,遇到了Good-turning的介绍,网上找了相关的资料查看,总结如下. 思想: 其主要思想是从概率的总量中分配少量的比例给零概率项. 思路: 假定给定的语料库中出现 \(r\)次数的词有\(N_r\)个,则有 \[N = \sum_{i=1}^{\infty }rN_r \] 当\(r\)较小时,我们使用\(d_r\)代替\(r\),这里\(d_r < r\),\(d_r\)的值可有下式表示 \[d_r = ( r+1 )N_{r+1}/N_r \] 其期望可以表示为 \[

SRILM是一个建立和使用统计语言模型的开源工具包,从1995年开始由SRI 口语技术与研究实验室(SRI Speech Technology and Research Laboratory)开发,现在仍然不断推出新版本,被广泛应用于语音识别.机器翻译等领域.这个工具包包含一组C++类库.一组进行语言模型训练和应用的可执行程序等.利用它可以非常方便地训练和应用语言模型.给定一组连续的词,调用SRILM提供的接口,可以得到这组词出现的概率. http://www.jianshu.com/p/5b1

目录: 一基本训练 二语言模型打分 三语言模型剪枝 四语言模型合并 五语言模型使用词典限制 一.基本训练 #功能 读取分词后的text文件或者count文件,然后用来输出最后汇总的count文件或者语言模型 #参数 输入文本: -read 读取count文件 -text 读取分词后的文本文件 词典文件: -vocab 限制text和count文件的单词,没有出现在词典的单词替换为<unk>:如果没有,所有的单词将会被自动加入词典 -limit-vocab 只限制count文件的单词(对text

原创作者 | 杨健 论文标题: K-BERT: Enabling Language Representation with Knowledge Graph 收录会议: AAAI 论文链接: https://ojs.aaai.org/index.php/AAAI/article/view/5681 项目地址: https://github.com/autoliuweijie/K-BERT 01 背景论述 笔者在前面的论文解读中提到过ERNIE使用基于自注意力机制来克服异构向量的融合,而KEPLER

1.简介 SRILM是通过统计方法构建语言模型,主要应用于语音识别,文本标注和切分,以及机器翻译等. SRILM支持语言模型的训练和评测,通过训练数据得到语言模型,其中包括最大似然估计及相应的平滑算法:评测是计算测试集的困惑度.其最基础和最核心的模块是n-gram模块,包括两个工 具:ngram-count和ngram,被用来估计语言模型和计算语言模型的困惑度. 2.使用方法 (1).语料初始化 a.数据清洗        b.分词(以空格划分)        c.将数据分为训练集和测试集 (2

http://52opencourse.com/111/斯坦福大学自然语言处理第四课-语言模型(language-modeling) 一.课程介绍 斯坦福大学于2012年3月在Coursera启动了在线自然语言处理课程,由NLP领域大牛Dan Jurafsky 和 Chirs Manning教授授课:https://class.coursera.org/nlp/ 以下是本课程的学习笔记,以课程PPT/PDF为主,其他参考资料为辅,融入个人拓展.注解,抛砖引玉,欢迎大家在“我爱公开课”上一起探讨学

最近试了一下Word2Vec, GloVe 以及对应的python版本 gensim word2vec 和 python-glove,就有心在一个更大规模的语料上测试一下,自然而然维基百科的语料进入了视线.维基百科官方提供了一个很好的维基百科数据源:https://dumps.wikimedia.org,可以方便的下载多种语言多种格式的维基百科数据.此前通过gensim的玩过英文的维基百科语料并训练LSI,LDA模型来计算两个文档的相似度,所以想看看gensim有没有提供一种简便的方式来处理维基

原文转载:http://licstar.net/archives/328 Deep Learning 算法已经在图像和音频领域取得了惊人的成果,但是在 NLP 领域中尚未见到如此激动人心的结果.关于这个原因,引一条我比较赞同的微博. @王威廉:Steve Renals算了一下icassp录取文章题目中包含deep learning的数量,发现有44篇,而naacl则有0篇.有一种说法是,语言(词.句子.篇章等)属于人类认知过程中产生的高层认知抽象实体,而语音和图像属于较为底层的原始输入信号,所以