使用马尔可夫逻辑网(MLNs)进行分类实验

3.实践

该例子为一个3类文本分类问题，使用MLN目前唯一的Alchemy作为实验平台：

依照Alchemy用户手册给出的语法和规则，定义各个文件如下：

1) three.mln

//predicate declarations

HasToken(doc, token)

Class0(doc)

Class1(doc)

Class2(doc)

//formulas

HasToken(d, +t) => Class0(d)

HasToken(d, +t) => Class1(d)

HasToken(d, +t) => Class2(d)

2) three-train.db

Class0(Doc37261)

Class0(Doc37913)

...

HasToken(Doc37261, ASheep)

HasToken(Doc37261, AVisualization)

HasToken(Doc37261, AXref)

...

3) three-test.db

HasToken(Doc39063, ASystems)

HasToken(Doc39063, AWet)

HasToken(Doc39063, ATIF)

...

权重学习命令： ../../bin/learnwts -d -i three.mln -o three-out.mln -t three-train.db -ne HasToken,Class0,Class1,Class2

学习到的规则及其权重如下：

//predicate declarations

HasToken(doc,token)

Class0(doc)

Class1(doc)

Class2(doc)

// 1.80043  HasToken(d,ASheep) => Class0(d)

1.80043  !HasToken(a1,ASheep) v Class0(a1)

// 2.96212  HasToken(d,AVisualization) => Class0(d)

2.96212  !HasToken(a1,AVisualization) v Class0(a1)

...

推理事实命令： ../../bin/infer -i three-out.mln -e three-test.db -r three.results -q Class0,Class1,Class2

推理结果为：

注：该次试验是失败的，效果也N不好，因为没有对特征进行任何处理，连词干化都没做

...

Class0(Doc61551) 4.9995e-05

Class0(Doc61552) 4.9995e-05

...

最后文件组织如下图：

分别给定训练集路径和测试集路劲，转换语料库为Alchemy格式的数据，数据的组织如下图：

转换格式的python脚本如下：

#encoding:utf-8

import os

import re

tp = re.compile(r'\W')

mlnfile = "three.mln"

traindb = "three-train.db"

testdb  = "three-test.db"

validb  = "three-validate.db"

#对指定文件路径生成token列表

def genetoken(filepath):

    lines = open(filepath).readlines()

    tokens = list()

    line_tokens = [token for line in lines for token in tp.split(line) if len(token.strip())>0]

    map(tokens.append, line_tokens)

    tokens = map(upperToken, tokens)   

    return ["A"+token for token in list(set(tokens))]

#把token首字符变大写

def upperToken(token):

    token = token[0].upper()+token[1:] if len(token) > 1 else token.upper()

    return token

def getTitle(title):

    return "Doc"+title

if __name__ == '__main__':

    train_path = 'train'#train set path

    test_path  = 'test' #test set path

    classIndex = dict()#类别索引

    for index, subfile in enumerate(os.listdir(train_path)):

        className = "Class" + str(index)

        classIndex[subfile] = "Class" + str(index)

    #生成训练集

    for subdir in os.listdir(train_path):

        className = classIndex[subdir]#类别名

        subdir = os.path.join(train_path, subdir)#子目录，即类别目录

        titledb = [className + "(" + getTitle(title) + ")" for title in os.listdir(subdir)]

        #如Class1(Doc1111), Class2(Doc2222)

        open(traindb,'a').write('\n'.join(titledb) + '\n\n')

    for subdir in os.listdir(train_path):

        className = classIndex[subdir]#类别名

        subdir = os.path.join(train_path, subdir)#子目录，即类别目录

        for docfile in os.listdir(subdir):

            titleName = getTitle(docfile)

            tokens = genetoken(os.path.join(subdir, docfile))

            tokendb = ["HasToken(" + titleName + ", " + token + ")" for token in tokens]

            open(traindb,'a').write('\n'.join(tokendb)+ '\n\n')   

    #生成验证集

    for subdir in os.listdir(test_path):

        className = classIndex[subdir]#类别名

        subdir = os.path.join(test_path, subdir)

        titledb = [className + "(" + getTitle(title) + ")" for title in os.listdir(subdir)]#如Class1(Doc1111), Class2(Doc2222)

        open(validb,'a').write('\n'.join(titledb) + '\n\n')

        for docfile in os.listdir(subdir):

            titleName = getTitle(docfile)

            tokens = genetoken(os.path.join(subdir, docfile))

            tokendb = ["HasToken(" + titleName + ", " + token + ")" for token in tokens]

            open(testdb,'a').write('\n'.join(tokendb) + '\n\n')