以前对于文本类型的数据,都是通过tf-idf进行处理的,这个可以参见以前写的博客,这里就不在详细介绍了。最近项目组老大跟我说了word2vec这种文本型特征提取的方式。特地给我讲解了一下俩者之间的区别:
一个词经过tf-idf处理之后,是一个数字,如果是相近的词语,它是无法区分的。Word2Vec就不一样了,比如研究和科研这俩个词,经过Word2Vec处理之后,是向量的形式。科研:[1,0,0,1,0],研究[1,0,0,0.8,0]。是可以判断是否相近的。
对于概念,我在spark官网翻译了一段话:计算一系列词的分布式向量。分布式的主要优点是相近的词在向量空间中是相近的,使泛化的新模式更容易和模型的评估更强大。分布式向量显示在许多自然语言处理应用中是有用的。命名实体类别,消歧。解析,标注和机器翻译。
代码如图:
package com.iflytek.features
import org.apache.spark.ml.feature.Word2Vec
import org.apache.spark.ml.linalg.Vector
import org.apache.spark.sql.Row
import org.apache.spark.sql.SparkSession
object wordtovec {
def main(args: Array[String]): Unit = {
val spark=SparkSession.builder().appName("pca").master("local").getOrCreate()
// Input data: Each row is a bag of words from a sentence or document.
val documentDF = spark.createDataFrame(Seq(
"Hi I heard about Spark".split(" "),
"I wish Java could use case classes".split(" "),
"Logistic regression models are neat".split(" ")
).map(Tuple1.apply)).toDF("text")
import org.apache.spark.ml.feature.Word2Vec
import org.apache.spark.ml.linalg.Vector
import org.apache.spark.sql.Row
import org.apache.spark.sql.SparkSession
object wordtovec {
def main(args: Array[String]): Unit = {
val spark=SparkSession.builder().appName("pca").master("local").getOrCreate()
// Input data: Each row is a bag of words from a sentence or document.
val documentDF = spark.createDataFrame(Seq(
"Hi I heard about Spark".split(" "),
"I wish Java could use case classes".split(" "),
"Logistic regression models are neat".split(" ")
).map(Tuple1.apply)).toDF("text")
// Learn a mapping from words to Vectors.
val word2Vec = new Word2Vec()
.setInputCol("text")
.setOutputCol("result")
.setVectorSize(3)
.setMinCount(0)
val model = word2Vec.fit(documentDF)
val result = model.transform(documentDF)
result.select("result").take(3).foreach(println)
val vecs=model.getVectors
vecs.foreach { x => println(x.apply(0)+":"+x.apply(1))}
val synonyms =model.findSynonyms("are", 3)
synonyms.select("word", "similarity").foreach { x => println(x.apply(0)+":"+x.apply(1)) }
}
}
result.select("result").take(3).foreach(println)
val vecs=model.getVectors
vecs.foreach { x => println(x.apply(0)+":"+x.apply(1))}
val synonyms =model.findSynonyms("are", 3)
synonyms.select("word", "similarity").foreach { x => println(x.apply(0)+":"+x.apply(1)) }
}
}
result的输出结果:
[[-0.028139343485236168,0.04554025698453188,-0.013317196490243079]]
[[0.06872416580361979,-0.02604914902310286,0.02165239889706884]]
[[0.023467857390642166,0.027799883112311366,0.0331136979162693]]
[[0.06872416580361979,-0.02604914902310286,0.02165239889706884]]
[[0.023467857390642166,0.027799883112311366,0.0331136979162693]]
vecs的输出结果:
heard:[-0.053989291191101074,0.14687322080135345,-0.0022512583527714014]
are:[-0.16293057799339294,-0.14514029026031494,0.1139335036277771]
neat:[-0.0406828410923481,0.028049567714333534,-0.16289857029914856]
classes:[-0.1490514725446701,-0.04974571615457535,0.03320947289466858]
I:[-0.019095497205853462,-0.131216898560524,0.14303986728191376]
regression:[0.16541987657546997,0.06469681113958359,0.09233078360557556]
Logistic:[0.036407098174095154,0.05800342187285423,-0.021965932101011276]
Spark:[-0.1267719864845276,0.09859133511781693,-0.10378564894199371]
could:[0.15352481603622437,0.06008218228816986,0.07726015895605087]
use:[0.08318991959095001,0.002120430115610361,-0.07926633954048157]
Hi:[-0.05663909390568733,0.009638422168791294,-0.033786069601774216]
models:[0.11912573128938675,0.1333899050951004,0.1441687047481537]
case:[0.14080166816711426,0.08094961196184158,0.1596144139766693]
about:[0.11579915136098862,0.10381520539522171,-0.06980287283658981]
Java:[0.12235434353351593,-0.03189820423722267,-0.1423865109682083]
wish:[0.14934538304805756,-0.11263544857501984,-0.03990427032113075]
are:[-0.16293057799339294,-0.14514029026031494,0.1139335036277771]
neat:[-0.0406828410923481,0.028049567714333534,-0.16289857029914856]
classes:[-0.1490514725446701,-0.04974571615457535,0.03320947289466858]
I:[-0.019095497205853462,-0.131216898560524,0.14303986728191376]
regression:[0.16541987657546997,0.06469681113958359,0.09233078360557556]
Logistic:[0.036407098174095154,0.05800342187285423,-0.021965932101011276]
Spark:[-0.1267719864845276,0.09859133511781693,-0.10378564894199371]
could:[0.15352481603622437,0.06008218228816986,0.07726015895605087]
use:[0.08318991959095001,0.002120430115610361,-0.07926633954048157]
Hi:[-0.05663909390568733,0.009638422168791294,-0.033786069601774216]
models:[0.11912573128938675,0.1333899050951004,0.1441687047481537]
case:[0.14080166816711426,0.08094961196184158,0.1596144139766693]
about:[0.11579915136098862,0.10381520539522171,-0.06980287283658981]
Java:[0.12235434353351593,-0.03189820423722267,-0.1423865109682083]
wish:[0.14934538304805756,-0.11263544857501984,-0.03990427032113075]
synonyms的输出:
classes:0.8926231541787831
I:0.8009102388269147
Hi:0.24258211195587995
I:0.8009102388269147
Hi:0.24258211195587995
getVectors:得到语料中所有词及其词向量
transform:将训练语料中,一行,也就是一个句子,表示成一个向量。它的处理方式是,对句子中所有的词向量取平均作为句子的向量表示,最native的表示方法。
findSynonyms("are",3):得到的是与词are相近的3个词。
经过我的实际测试,发现使用Word2Vector可以提高各项评价指标,大家也不妨试试啊。
spark2.0的分类、回归、聚类算法我都测试了一遍,只有分类的朴素贝叶斯是行不通的(特征值不接受负的),其他的都是行的通的。