DeepLearning4j是一个java的神经网络框架,便于java程序员使用神经网络来完成一些机器学习工程。

不管什么机器学习框架,NLP是一个不能不谈的领域,DL4J也提供了nlp的相关实现。其中入门的例子就是从一大堆文字中找到最相关的词。

我们先来看看官方的demo,然后再模仿一个类似的程序,只不过是阅读中文的小说。

官方的demo叫Word2VecRawTextExample,我们直接新建一个java的maven项目,pom.xml如下:

<?xml version="1.0" encoding="UTF-8"?>
<project xmlns="http://maven.apache.org/POM/4.0.0"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
<modelVersion>4.0.0</modelVersion>

<groupId>com.tianyalei</groupId>
<artifactId>wolf_ml_mnist</artifactId>
<version>1.0-SNAPSHOT</version>

<properties>
<project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
<nd4j.version>1.0.0-beta</nd4j.version>
<dl4j.version>1.0.0-beta</dl4j.version>
<datavec.version>1.0.0-beta</datavec.version>
<logback.version>1.1.7</logback.version>
<scala.binary.version>2.10</scala.binary.version>
</properties>

<dependencies>
<!--神经网络的实现方法-->
<dependency>
<groupId>org.deeplearning4j</groupId>
<artifactId>deeplearning4j-core</artifactId>
<version>${dl4j.version}</version>
</dependency>
<!--ND4J库的CPU版本,驱动DL4J-->
<!--nd4j-cuda-9.1-platform,写成这个就是GPU-->
<dependency>
<groupId>org.nd4j</groupId>
<artifactId>nd4j-native-platform</artifactId>
<!--<artifactId>nd4j-cuda-9.1-platform</artifactId>-->
<version>${nd4j.version}</version>
</dependency>
<dependency>
<groupId>org.deeplearning4j</groupId>
<artifactId>deeplearning4j-nlp</artifactId>
<version>${dl4j.version}</version>
</dependency>
<dependency>
<groupId>org.deeplearning4j</groupId>
<artifactId>deeplearning4j-ui_2.11</artifactId>
<version>${dl4j.version}</version>
</dependency>
<dependency>
<groupId>ch.qos.logback</groupId>
<artifactId>logback-classic</artifactId>
<version>${logback.version}</version>
</dependency>
<!--中文分词start-->
<!--<dependency>
<groupId>org.fnlp</groupId>
<artifactId>fnlp-core</artifactId>
<version>2.1-SNAPSHOT</version>
</dependency>-->
<dependency>
<groupId>net.sf.trove4j</groupId>
<artifactId>trove4j</artifactId>
<version>3.0.3</version>
</dependency>
<dependency>
<groupId>commons-cli</groupId>
<artifactId>commons-cli</artifactId>
<version>1.2</version>
</dependency>
<!--中文分词end-->
</dependencies>
</project>
package com.tianyalei.nlp;

import org.deeplearning4j.models.word2vec.Word2Vec;
import org.deeplearning4j.text.sentenceiterator.BasicLineIterator;
import org.deeplearning4j.text.sentenceiterator.SentenceIterator;
import org.deeplearning4j.text.tokenization.tokenizer.preprocessor.CommonPreprocessor;
import org.deeplearning4j.text.tokenization.tokenizerfactory.DefaultTokenizerFactory;
import org.deeplearning4j.text.tokenization.tokenizerfactory.TokenizerFactory;
import org.nd4j.linalg.io.ClassPathResource;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

import java.util.Collection;

/**
* Created by agibsonccc on 10/9/14.
*
* Neural net that processes text into wordvectors. See below url for an in-depth explanation.
* https://deeplearning4j.org/word2vec.html
*/
public class Word2VecRawTextExample {

private static Logger log = LoggerFactory.getLogger(Word2VecRawTextExample.class);

public static void main(String[] args) throws Exception {

// Gets Path to Text file
String filePath = new ClassPathResource("raw_sentences.txt").getFile().getAbsolutePath();

log.info("Load & Vectorize Sentences....");
// Strip white space before and after for each line
SentenceIterator iter = new BasicLineIterator(filePath);
// Split on white spaces in the line to get words
TokenizerFactory t = new DefaultTokenizerFactory();

/*
CommonPreprocessor will apply the following regex to each token: [\d\.:,"'\(\)\[\]|/?!;]+
So, effectively all numbers, punctuation symbols and some special symbols are stripped off.
Additionally it forces lower case for all tokens.
*/
t.setTokenPreProcessor(new CommonPreprocessor());

log.info("Building model....");
Word2Vec vec = new Word2Vec.Builder()
//是一个词在语料中必须出现的最少次数。本例中出现不到五次的词都不予学习。
.minWordFrequency(5)
//是网络在处理一批数据时允许更新系数的次数。迭代次数太少,网络可能来不及学习所有能学到的信息;迭代次数太多则会导致网络定型时间变长。
.iterations(1)
//指定词向量中的特征数量,与特征空间的维度数量相等。以500个特征值表示的词会成为一个500维空间中的点。
.layerSize(100)
.seed(42)
.windowSize(5)
//告知网络当前定型的是哪一批数据集
.iterate(iter)
//将当前一批的词输入网络
.tokenizerFactory(t)
.build();

log.info("Fitting Word2Vec model....");
vec.fit();

log.info("Writing word vectors to text file....");

// Prints out the closest 10 words to "day". An example on what to do with these Word Vectors.
log.info("Closest Words:");
Collection<String> lst = vec.wordsNearestSum("day", 10);
//Collection<String> lst = vec.wordsNearest(Arrays.asList("king", "woman"), Arrays.asList("queen"), 10);
log.info("10 Words closest to 'day': {}", lst);
}
}

这就是NLP的helloworld级的入门项目,目标是从给定的raw_sentences.txt中找到与day最相近的词,将资源放到resource中,运行该程序即可。

运行结果:

让DeepLearning4j阅读小说并给出关联度最高的词_DeepLearning4j

可以看到,day的最相近的词有week、night、year等,还算非常靠谱。至于原理呢,大家可以在文档里去搜索day这个词,看看它的附近的词和用法,然后再去搜索week、night等词的旁边的词和用法,就大概知道怎么回事了。

该文用的相关资源去我项目里找https://github.com/tianyaleixiaowu/wolf_ml_mnist

从代码的注释中可以看看基本的概念,下面我们来让它学习一下中文的小说,并给出最接近的词。

和英文自带空格分词不同,中文是额外需要一个中文分词器的,不然中文全是一句一句的,根本分不开。所以我们在让机器学习读中文前,需要先将中文句子分成一个个的词语。

让DeepLearning4j阅读小说并给出关联度最高的词_nlp_02

中文分词器很多,论使用的简易程度和效果,还是复旦的NLP比较靠谱,https://github.com/FudanNLP/fnlp。

GitHub上面有文档,讲怎么使用的,这里我直接简单说一下,下载models里的三个.m文件,和libs里的fnlp-code.jar,将jar添加为工程的依赖lib即可。至于复旦nlp额外需要的两个jar,我已经放在pom.xml里了。

让DeepLearning4j阅读小说并给出关联度最高的词_DeepLearning4j_03

然后就可以使用fnlp来对文档进行分词了。我们选择的文档是天龙八部tlbb.txt,这是没分词时的样子。

让DeepLearning4j阅读小说并给出关联度最高的词_maven_04

分词的代码

package com.tianyalei.nlp.tlbb;

import java.io.*;

/**
* 运行后将得到一个分词后的文档
* @author wuweifeng wrote on 2018/6/29.
*/
public class FenCi {
private FudanTokenizer tokenizer = new FudanTokenizer();

public void processFile() throws Exception {
String filePath = this.getClass().getClassLoader().getResource("text/tlbb.txt").getPath();
BufferedReader in = new BufferedReader(new FileReader(filePath));

File outfile = new File("/Users/wuwf/project/tlbb_t.txt");
if (outfile.exists()) {
outfile.delete();
}
FileOutputStream fop = new FileOutputStream(outfile);

// 构建FileOutputStream对象,文件不存在会自动新建
String line = in.readLine();
OutputStreamWriter writer = new OutputStreamWriter(fop, "UTF-8");
while (line != null) {
line = tokenizer.processSentence(line);
writer.append(line);
line = in.readLine();
}
in.close();
writer.close(); // 关闭写入流,同时会把缓冲区内容写入文件
fop.close(); // 关闭输出流,释放系统资源
}

public static void main(String[] args) throws Exception {
new FenCi().processFile();
}
}
package com.tianyalei.nlp.tlbb;

import org.fnlp.ml.types.Dictionary;
import org.fnlp.nlp.cn.tag.CWSTagger;
import org.fnlp.nlp.corpus.StopWords;
import org.fnlp.util.exception.LoadModelException;

import java.io.IOException;
import java.util.List;

/**
* @author wuweifeng wrote on 2018/6/29.
*/
public class FudanTokenizer {
private CWSTagger tag;

private StopWords stopWords;

public FudanTokenizer() {
String path = this.getClass().getClassLoader().getResource("").getPath();
System.out.println(path);
try {
tag = new CWSTagger(path + "models/seg.m");
} catch (LoadModelException e) {
e.printStackTrace();
}

}

public String processSentence(String context) {
return tag.tag(context);
}

public String processSentence(String sentence, boolean english) {
if (english) {
tag.setEnFilter(true);
}
return tag.tag(sentence);
}

public String processFile(String filename) {
return tag.tagFile(filename);
}

/**
* 设置分词词典
*/
public boolean setDictionary() {
String dictPath = this.getClass().getClassLoader().getResource("models/dict.txt").getPath();

Dictionary dict;
try {
dict = new Dictionary(dictPath);
} catch (IOException e) {
return false;
}
tag.setDictionary(dict);
return true;
}

/**
* 去除停用词
*/
public List<String> flitStopWords(String[] words) {
try {
return stopWords.phraseDel(words);
} catch (Exception e) {
e.printStackTrace();
return null;
}
}

}

然后运行一下,过一会就得到了分词后的文档tlbb_t.txt,将分词后的拷贝到resource下,将来机器学的就是分词后的文档。

让DeepLearning4j阅读小说并给出关联度最高的词_xml_05

package com.tianyalei.nlp.tlbb;

import org.deeplearning4j.models.embeddings.loader.WordVectorSerializer;
import org.deeplearning4j.models.word2vec.Word2Vec;
import org.deeplearning4j.text.sentenceiterator.BasicLineIterator;
import org.deeplearning4j.text.sentenceiterator.SentenceIterator;
import org.deeplearning4j.text.tokenization.tokenizer.preprocessor.CommonPreprocessor;
import org.deeplearning4j.text.tokenization.tokenizerfactory.DefaultTokenizerFactory;
import org.deeplearning4j.text.tokenization.tokenizerfactory.TokenizerFactory;
import org.nd4j.linalg.io.ClassPathResource;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

import java.io.File;
import java.io.IOException;
import java.util.Collection;

/**
* @author wuweifeng wrote on 2018/6/29.
*/
public class Tlbb {
private static Logger log = LoggerFactory.getLogger(Tlbb.class);

public static void main(String[] args) throws IOException {
String filePath = new ClassPathResource("text/tlbb_t.txt").getFile().getAbsolutePath();
log.info("Load & Vectorize Sentences....");

SentenceIterator iter = new BasicLineIterator(new File(filePath));

TokenizerFactory t = new DefaultTokenizerFactory();
t.setTokenPreProcessor(new CommonPreprocessor());

log.info("Building model....");
Word2Vec vec = new Word2Vec.Builder().
minWordFrequency(5)
.iterations(1)
.layerSize(100)
.seed(42)
.windowSize(5)
.iterate(iter)
.tokenizerFactory(t)
.build();
log.info("Fitting Word2Vec model....");
vec.fit();
log.info("Writing word vectors to text file....");

// Write word vectors to file
log.info("Writing word vectors to text file....");

WordVectorSerializer.writeWordVectors(vec, "tlbb_vectors.txt");
WordVectorSerializer.writeFullModel(vec, "tlbb_model.txt");
String[] names = {"萧峰", "乔峰", "段誉", "虚竹", "王语嫣", "阿紫", "阿朱", "木婉清"};
log.info("Closest Words:");

for (String name : names) {
System.out.println(name + ">>>>>>");
Collection<String> lst = vec.wordsNearest(name, 10);
System.out.println(lst);
}


}

}

代码和之前的demo区别不大,运行后,就能看到这几个人的关联度最高的词了。

让DeepLearning4j阅读小说并给出关联度最高的词_DeepLearning4j_06