Lecture 5 Part of Speech Tagging

Part of Speech(POS)

• Also called word classes, morphological classes, syntactic categories 也称为词类、形态类、句法类别

• E.g.: nouns, verbs, adjective 例如：名词、动词、形容词

• POS tells information about a word and its neighbors: 词性提供了关于单词及其相邻单词的信息

  • Nouns are often preceded by determiners 名词通常由限定词前置
  • Verbs preceded by nouns 动词通常由名词前置
  • content as a noun pronounced as /'kɑ:ntent/
  • content as an adjective pronounced as /kən’tent/

POS application: Information Extraction 词性应用：信息提取

• Given sentence: “Brasilia, the Brazilian capital, was founded in 1960”

• Extract information: 提取信息

  • capital(Brazil, Brasilia)
  • founded(Brasilia, 1960)

• First step of information extraction is finding all POS tags: 信息提取的第一步是找到所有的词性标签

  • nouns: Brasilia, capital
  • adjective: Brazilian
  • verbs: founded
  • numbers: 1960

POS Open Class 开放类词性

• Open vs. closed: How readily do POS categories take on new words? 开放类 vs. 封闭类：词性类别接受新词的频率如何？

• E.g. of open classes: 开放类的例子

  • Nouns:
    • Proper(专有名词) vs. common(普通名词): Australia, wombat
    • Mass(集合名词) vs. count(可数名词): rice, bowls
  • Verbs:
    • Rich inflection: go/goes/going/gone/went 富有变化
    • Auxiliary verbs(助动词): be, have, do 助动词
    • Transitivity: wait, hit, give 及物性
  • Adjectives:
    • Gradable(等级形容词) vs. non-gradable(非等级形容词): happy/happier/happiest, computational
  • Adverbs:
    • Manner(情状副词): slowly
    • Locative(处所副词): here
    • Degree(程度副词): really
    • Temporal(时间副词): today

• E.g. of closed classes: 封闭类的例子

  • Prepositions(介词):
    • in, on, with, for, of, over
  • Particles:
    • off
  • Determiners(限定词):
    • Articles(冠词): a, an, the
    • Demonstratives(指示词): this, that, these, those
    • Quantifiers(数量词): each, every, some, two
  • Pronouns(代词):
    • Personal(人称代词): I, me, she
    • Possessive(所有格代词): my, our
    • Interrogative(疑问代词): who, what
  • Conjunctions(连词):
    • Coordinating(并列连词): and, or, but
    • Subordinating(从属连词): if, although, that
  • Modal verbs(情态动词):
    • Ability: can, could
    • Permission: can, may
    • Possibility: may, might, could, will
    • Necessity: must

Problem of word classes: Ambiguity 词类问题：模糊性

• Many word types belong to multiple classes 许多单词类型属于多个类别

• POS depends on context 词性取决于上下文

• E.g.: flies

  

  • The word flies in the first sentence is an inflection of the verb “fly” 在第一句中，flies 是动词 “fly” 的变形
  • The word flies in the second sentence is the plural form of the noun “fly” 在第二句中，flies 是名词 “fly” 的复数形式

Tagsets

Tagsets 标记集

• A compact representation of POS information 词性信息的紧凑表示

  • Usually less than 4 capitalized characters. E.g. NN = noun 通常少于4个大写字符。例如 NN = noun
  • Often includes inflectional distinctions 经常包括形态变化的区别

• Major English tagsets: 主要的英语标记集

  • Brown: 87 tags
  • Penn Treebank: 45 tags
  • CLAWS/BNC: 61 tags
  • Universal: 12 tags

• At least one tagset for all major languages 所有主要语言至少有一个标记集

Penn Treebank Tags:

• Open classes: 开放类

  • NN: noun 名词
  • VB: verb 动词
  • JJ: adjective 形容词
  • RB: adverb 副词

• Closed classes: 封闭类

  • DT: determiner 限定词
  • CD: cardinal number 基数
  • IN: preposition 介词
  • PRP: personal pronoun 人称代词
  • MD: modal 情态动词
  • CC: coordinating conjunction 并列连词
  • RP: particle 助词
  • WH: wh-pronoun 疑问代词
  • TO: to

Derived Tags: 衍生标签

• Open classes: 开放类

  • NN (noun singular): 单数名词
    • NNS (plural) 复数
    • NNP (proper) 专有名词
    • NNPS (proper plural) 复数专有名词
  • VB (verb infinitive): 不定式动词
    • VBP (1st/2nd person present) 第一/第二人称现在时
    • VBZ (3rd person singular) 第三人称单数
    • VBD (past tense) 过去时
    • VBG (gerund) 现在分词
    • VBN (past participle) 过去分词
  • JJ (adjective): 形容词
    • JJR (comparative) 比较级
    • JJS (superlative) 最高级
  • RB (adverb): 副词
    • RBR (comparative) 比较级
    • RBS (superlative) 最高级

• Closed classes: 封闭类

  • PRP (pronoun personal): 人称代词
    • PRP$ (possessive) 所有格
  • WP (wh-pronoun): 疑问代词
    • WP$ (possessive) 所有格
    • WDT (wh-determiner) 疑问限定词)
    • WRB (wh-adverb) 疑问副词

Tagged Text Example 标记文本示例

Automatic Tagging

Reasons for automatic POS tagging 自动词性标注的原因

• Important for morphological analysis. E.g. lemmatization 对形态分析很重要。例如：词形还原

• For some applications, we want to focus on certain POS 对于某些应用，我们希望关注某些词性

  • E.g. nouns are important for information retreieval, adjectives for sentiment analysis 例如：名词对于信息检索很重要，形容词对于情感分析很重要

• Very useful features for certain classification tasks. 对于某些分类任务，这是非常有用的特性

  • E.g. genre attribution 体裁属性

• POS tags can offer word sense disambiguation 词性标签可以提供词义消歧

  • E.g. cross/NN, cross/VB, cross/JJ all have different means

• Can use them to create larger structures 可以用它们来创建更大的结构

Automatic Taggers 自动标注器

• Rule-based taggers 基于规则的标注器
• Statistical taggers 统计标注器
  • Unigram tagger 一元标注器
  • Classifier-based tagger 基于分类器的标注器
  • Hidden Markov Model tagger 隐马尔科夫模型标注器

Rule-Based Tagging

• Typically starts with a list of possible tags for each word. Source from a lexical resource or a corpus 通常从词典或语料库中为每个单词列出可能的标签开始
• Often includes other lexcial information. E.g. verb subcategorization 经常包括其他词汇信息。例如：动词下类化
• Apply rules to narrow down to a single tag 应用规则以缩小到一个标签
• Large systems have thousands of constraints 大型系统有数千个约束

Unigram Tagger

• Assign most common tag to each word type 为每个单词类型分配最常见的标签
• Requires a corpus of tagged words 需要一个标记过的词语的语料库
• Just a look-up table 只是一个查找表
• Approximately 90% accuracy 精度约为90%
• Often considered the baseline for more complex approaches 通常被认为是更复杂方法的基线

Classifier-Based Tagging

• Use a standard discriminative classifier like logistic regression or neural network with features: 使用如逻辑回归或神经网络这样的标准判别式分类器，其特征包括

  • Target word 目标词
  • Lexical context around the word 词周围的词汇上下文
  • Already classified tags in the sentence 句子中已分类的标签

• Can suffer from error propagation: wrong predictions from previous steps affect the next ones 可能受到错误传播的影响：前一步的错误预测影响下一步

Hidden Markov Models

• A basic sequential model 一个基本的序列模型
• Like sequential classifiers, use both previous tag and lexical evident 与序列分类器一样，使用前一个标签和词汇证据
• Unlike classifiers, considers all possibilities of previous tag and treat previous tag evidence and lexical evidence as independent from each other 与分类器不同的是，它考虑了前一个标签的所有可能性，并将前一个标签的证据和词汇证据视为相互独立的
  • Less sparsity 稀疏度较小
  • Fast algorithms for sequential prediction 针对序列预测的快速算法

Unknown Words

• Huge problem in morphologically rich languages 在形态丰富的语言中是一个巨大的问题

• Can use things already seen only once to best guess for things never seen before 可以使用已经看到一次的事物来对从未见过的事物进行最佳猜测

  • Tend to be nouns, followed by verbs 倾向于是名词，然后是动词
  • Unlikely to be determiners 不太可能是限定词

• Can use sub-word representations to capture morphology 可以使用子词表示来捕获形态