com.hankcs.hanlp.model.hmm

Class HiddenMarkovModel

java.lang.Object

com.hankcs.hanlp.model.hmm.HiddenMarkovModel

Direct Known Subclasses:

FirstOrderHiddenMarkovModel, SecondOrderHiddenMarkovModel

public abstract class HiddenMarkovModel
extends Object

Author:

hankcs

Field Summary

Fields

Modifier and Type	Field and Description
float[][]	emission_probability 观测概率矩阵
float[]	start_probability 初始状态概率向量
float[][]	transition_probability 状态转移概率矩阵

Constructor Summary

Constructors

Constructor and Description
HiddenMarkovModel(float[] start_probability, float[][] transition_probability, float[][] emission_probability) 构造隐马模型

Method Summary

Modifier and Type	Method and Description
protected static Object	deepCopy(Object object)
protected static int	drawFrom(double[] cdf) 采样
protected void	estimateEmissionProbability(Collection<int[][]> samples, int max_state, int max_obser) 估计状态发射概率
protected void	estimateStartProbability(Collection<int[][]> samples, int max_state) 估计初始状态概率向量
protected void	estimateTransitionProbability(Collection<int[][]> samples, int max_state) 利用极大似然估计转移概率
abstract int[][]	generate(int length) 生成样本序列
List<int[][]>	generate(int minLength, int maxLength, int size) 生成样本序列
protected static double[]	logToCdf(float[] log) 对数概率转为累积分布函数
protected static double[][]	logToCdf(float[][] log) 对数概率转化为累积分布函数
protected void	normalize(float[] freq) 频次向量归一化为概率分布
abstract float	predict(int[] o, int[] s) 预测（维特比算法）
float	predict(int[] o, Integer[] s) 预测（维特比算法）
protected static boolean	similar(float[] A, float[] B)
boolean	similar(HiddenMarkovModel model)
protected void	toLog()
void	train(Collection<int[][]> samples) 训练
void	unLog()

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

start_probability

public float[] start_probability

初始状态概率向量

emission_probability

public float[][] emission_probability

观测概率矩阵

transition_probability

public float[][] transition_probability

状态转移概率矩阵

Constructor Detail

HiddenMarkovModel

public HiddenMarkovModel(float[] start_probability,
                         float[][] transition_probability,
                         float[][] emission_probability)

构造隐马模型

Parameters:

start_probability - 初始状态概率向量

transition_probability - 状态转移概率矩阵

emission_probability - 观测概率矩阵

Method Detail

logToCdf

protected static double[] logToCdf(float[] log)

对数概率转为累积分布函数

Parameters:

log -

Returns:

logToCdf

protected static double[][] logToCdf(float[][] log)

对数概率转化为累积分布函数

Parameters:

log -

Returns:

drawFrom

protected static int drawFrom(double[] cdf)

采样

Parameters:

cdf - 累积分布函数

Returns:

normalize

protected void normalize(float[] freq)

频次向量归一化为概率分布

Parameters:

freq -

unLog

public void unLog()

toLog

protected void toLog()

train

public void train(Collection<int[][]> samples)

训练

Parameters:

samples - 数据集 int[i][j] i=0为观测，i=1为状态，j为时序轴

estimateEmissionProbability

protected void estimateEmissionProbability(Collection<int[][]> samples,
                                           int max_state,
                                           int max_obser)

估计状态发射概率

Parameters:

samples - 训练样本集

max_state - 状态的最大下标

max_obser - 观测的最大下标

estimateTransitionProbability

protected void estimateTransitionProbability(Collection<int[][]> samples,
                                             int max_state)

利用极大似然估计转移概率

Parameters:

samples - 训练样本集

max_state - 状态的最大下标，等于N-1

estimateStartProbability

protected void estimateStartProbability(Collection<int[][]> samples,
                                        int max_state)

估计初始状态概率向量

Parameters:

samples - 训练样本集

max_state - 状态的最大下标

generate

public abstract int[][] generate(int length)

生成样本序列

Parameters:

length - 序列长度

Returns:

序列

generate

public List<int[][]> generate(int minLength,
                              int maxLength,
                              int size)

生成样本序列

Parameters:

minLength - 序列最低长度

maxLength - 序列最高长度

size - 需要生成多少个

Returns:

样本序列集合

predict

public abstract float predict(int[] o,
                              int[] s)

预测（维特比算法）

Parameters:

o - 观测序列

s - 预测状态序列（需预先分配内存）

Returns:

概率的对数，可利用 (float) Math.exp(maxScore) 还原

predict

public float predict(int[] o,
                     Integer[] s)

预测（维特比算法）

Parameters:

o - 观测序列

s - 预测状态序列（需预先分配内存）

Returns:

概率的对数，可利用 (float) Math.exp(maxScore) 还原

similar

public boolean similar(HiddenMarkovModel model)

similar

protected static boolean similar(float[] A,
                                 float[] B)

deepCopy

protected static Object deepCopy(Object object)