Fundamentals of Classification

Classification 分类

• Input:

  • A document d: Often represented as a vector of features 一个文档 d：通常表示为特征向量
  • A fixed output set of classes C = {c₁, c₂, …, c_k}: Categorical, not continuous or ordinal 一个固定的输出类别集合 C = {c₁, c₂, …, c_k}：类别是离散的，不是连续或有序的

• Output:

  • A predicted class c ∈ C 一个预测的类别 c ∈ C

Text Classification Tasks 文本分类任务

• Some common examples:

  • Topic classification 主题分类
  • Sentiment analysis 情感分析
  • Native-language identification 母语识别
  • Natural language interence 自然语言推理
  • Automatic fact-checking 自动事实检查
  • Paraphrase 释义（paraphrase）

• Input may not be a long document 输入可能不是一个长文档

Topic Classification 主题分类

• Motivation: Library science, information retrieval 动机：图书馆科学，信息检索

• Classes: Topic categories E.g. “jobs”, “international news” 类别：主题类别 例如"工作"，“国际新闻”

• Features:

  • Unigram bag-of-words, with stop-words removed 一元词袋模型，移除了停用词
  • Longer n-grams for phrases 长的n-grams表示短语

• Examples of corpora:

  • Reuters news corpus. E.g. RCV1, NLTK
  • Pubmed abstracts
  • Tweets with hashtags

Sentiment Analysis 情感分析

• Motivation: Opinion mining, business analytics 动机：观点挖掘，商业分析

• Classes: Positive/Negative/(Neutral) 类别：积极/消极/(中立)

• Features:

  • N-grams N-grams
  • Polarity lexicons(极性词词典)

• Examples of corpora:

  • Movie review dataset in NLTK
  • SEMEVAL Twitter polarity datasets

Native-Language Identification 母语识别

• Motivation: Forensic linguistics, educational applications 动机：法证语言学，教育应用

• Classes: First language of author 类别：作者的第一语言

• Features:

  • Word N-grams
  • Syntactic patterns (POS, parse trees) 句法模式
  • Phonological features 语音特征

• Example of corpora:

  • TOEFL/IELTS essay corpora

Natural Language Inference 自然语言推理

• Also called textual entailment 也称为文本蕴含

• Motivation: Language understanding 动机：语言理解

• Classes: entailment, contradiction, neutral 类别：蕴含，矛盾，中立

• Features:

  • Word overlap 词重叠
  • Length difference between the sentences 句子之间的长度差异
  • N-grams N-grams

• Examples of corpora:

  • SNLI, MNLI

Building a Text Classifier 构建文本分类器

1. Identify a task of interest 确定感兴趣的任务
2. Collect an appropriate corpus 收集适当的语料库
3. Carry out annotation 执行注释
4. Select features 选择特征
5. Choose a machine learning algorithm 选择机器学习算法
6. Train model and tune hyperparameters using hold-out development data 使用留出法开发数据训练模型并调整超参数
7. Repeat earlier steps as needed 根据需要重复前面的步骤
8. Train fianl model 训练最终模型
9. Evaluate model on hold-out test data 在留出测试数据上评估模型

Algorithms for Classification

Choosing a Classification Algorithm 选择分类算法

• Bias vs. Variance 偏差与方差

  • Bias: Assumptions made in the model 偏差：模型中做出的假设
  • Variance: Sensitivity to training set 方差：对训练集的敏感度

• Underlying assumptions E.g. Independence 潜在假设 例如 独立性

• Complexity 复杂度

• Speed 速度

Naive Bayes 朴素贝叶斯

• Find the class with the highest likelihood under Bayes Law: 根据贝叶斯定理找到最有可能的类别
  

  • Probability of the class times probability of features given the class 类别的概率乘以给定类别的特征概率

• Naively assumes features are independent: 天真地假设特征是独立的
  

• Pros:

  • Fast to train and classify 训练和分类速度快
  • Robust, low-variance -> good for low data situations 稳健，方差小 -> 对于数据量小的情况效果好
  • Optimal classifier if independence assumption is correct 如果独立性假设正确，那么它是最优分类器
  • Extremely simple to implement 实现极其简单

• Cons:

  • Independence assumption rarely holds 独立性假设很少成立

  • Low accuracy compared to similar methods in most situations 在大多数情况下，与类似方法相比，精度较低

  • Smoothing required for unseen class/feature combinations 对于未见过的类别/特征组合，需要进行平滑处理

Logistic Regression 逻辑回归

• A classifier, even called regression 一个分类器，甚至被称为回归

• A linear model, but users softmax function squashing to get valid probability 一个线性模型，但使用 softmax 函数将其压缩成有效的概率
  

• Training maximizes probability of training data subject to regularization which encourages low or sparse weights 训练过程最大化训练数据的概率，同时通过正则化来鼓励低权重或稀疏权重

• Pros:

  • Unlike Naive Bayes not confounded by diverse, correlated features gives better performance 与朴素贝叶斯不同，不会被多样性、相关特征所混淆，性能更好

• Cons:

  • Slow to train 训练速度慢
  • Feature scaling needed 需要特征缩放
  • Requires a lot of data to work well in practice 在实际应用中需要大量数据才能表现良好
  • Choosing regularization strategy is important since overfitting is a big problem 选择正则化策略很重要，因为过拟合是一个大问题

Support Vector Machines 支持向量机

• Finds hyperplane which separates the training data with maximum margin 找到一个超平面，该超平面最大限度地分离训练数据

• Pros:

  • Fast and accurate linear classifier 快速且准确的线性分类器
  • Can do non-linearity with kernel trick 可以使用核技巧进行非线性处理
  • Works well with huge feature sets 对于大规模特征集合工作良好

• Cons:

  • Multiclass classification awkward 多类别分类不方便
  • Feature scaling needed 需要特征缩放
  • Deals poorly with class imbalances 对类别不平衡处理不佳
  • Interpretability 可解释性差

K-Nearest Neighbor K-最近邻

• Classify based on majority class of k-nearest training examples in feature space 根据特征空间中最近的 k 个训练样本的多数类别进行分类

• Definition of the nearest can vary: "最近"的定义可以变化

  • Euclidean distance 欧几里得距离
  • Cosine distance 余弦距离

• Pros:

  • Simple but surprisingly effective 简单但效果惊人
  • No training required 无需训练
  • Inherently multiclass 内置多类别
  • Optimal classifier with infinite data 在无穷大的数据下是最优分类器

• Cons:

  • Have to select k 必须选择 k
  • Issues with imbalanced classes 类别不平衡问题
  • Often slow for finding the neighbors 在寻找最近邻时往往很慢
  • Features must be selected carefully 特征必须仔细选择

Decision Tree 决策树

• Construct a tree where nodes correspond to tests on individual features 构建一个树，其中节点对应于对单个特征的测试

• Leaves are final class decisions 叶子节点是最终的类别决策

• Based on greedy maximization of mutual information 基于贪心最大化互信息

• Pros:

  • Fast to build and test 建立和测试速度快
  • Feature scaling irrelevant 特征缩放无关
  • Good for small feature sets 对于小规模特征集合表现良好
  • Handles non-linearly-separable problems 能处理非线性可分问题

• Cons:

  • In practice, not very interpretable 在实际中，解释性不强
  • Highly redundant sub-trees 高度冗余的子树
  • Not competitive for large feature sets 对于大规模特征集合，表现不强

Random Forests 随机森林

• An ensemble classifier 一种集成分类器

• Consists of decision trees trained on different subsets of the training and feature space 由在训练集和特征空间的不同子集上训练的决策树组成

• Final class decision is majority voting of sub-classifiers 最终的类别决策是子分类器的多数投票

• Pros:

  • Usually more accurate and more robust than decision trees 通常比决策树更准确、更稳健
  • Great classifier for medium feature sets 对于中等规模的特征集合是很好的分类器
  • Training easily parallelized 训练过程可以轻易并行化

• Cons:

  • Interpretability 可解释性
  • Slow with large feature sets 在大规模特征集合中训练慢

Neural Networks 神经网络

• An interconnected set of nodes typically arranged in layers 一个由多个节点相互连接的集合，通常按层排列

• Input layer(features), output layer(class probabilities), and one or more hidden layers 输入层（特征），输出层（类别概率），以及一个或多个隐藏层

• Each node performs a linear weighting of its inputs from previous layer, passes result through activation function to nodes in next layer 每个节点对其来自上一层的输入进行线性加权，然后通过激活函数将结果传递给下一层的节点

• Pros:

  • Extremely powerful, dominant method in NLP and computer vision 极其强大，是 NLP 和计算机视觉中的主导方法
  • Little feature engineering 很少需要特征工程

• Cons:

  • Not an off-the-shelf classifier 不是开箱即用的分类器
  • Many hyperparameters, difficult to optimize 多个超参数，难以优化
  • Slow to train 训练慢
  • Prone to overfitting 易于过拟合

Hyperparameter Tuning 超参数调优

• Dataset for tuning: 调优的数据集

  • Development set 开发集
  • Not the training set or the test set 不是训练集或测试集
  • k-fold cross-validation k-折交叉验证

• Specific hyper-parameters are classifier specific. But many hyper-parameters relate to regularization 具体的超参数是分类器特定的。但许多超参数与正则化有关

  • Regularization hyperparameters penalize model complexity 正则化超参数惩罚模型复杂性
  • Used to prevent overfitting 用于防止过拟合

• For multiple hyperparameters, use grid search 对于多个超参数，使用网格搜索

Evaluation

Confusion Matrix 混淆矩阵

	Classified As
Class	A	B
A	True Positive	False Positive
B	False Negative	True Negative

Evaluation Metrics 评估指标

• Accuracy = True Positive / (True Positive + False Positive + True Negative + False Negative) 准确率 = 真正例 / (真正例 + 假正例 + 真负例 + 假负例)
• Precision = True Positive / (True Positive + False Positive) 精确度 = 真正例 / (真正例 + 假正例)
• Recall = True Positive / (True Positive + False Negative) 召回率 = 真正例 / (真正例 + 假负例)
• F1-score = (2 * precision * recall) / (precision + recall) F1-分数 = (2 * 精确度 * 召回率) / (精确度 + 召回率)
  • Macroaverage: Average F-score across classes 宏平均：所有类别的 F-分数的平均值
  • Micreaverage: Calculate F-score using sum of counts 微平均：使用计数和来计算 F-分数