Pig + Ansj 统计中文文本词频

---

最近特别喜欢用Pig，拥有能满足大部分需求的内置函数(built-in functions)，支持自定义函数(user defined functions, UDF)，能load 纯文本、avro等格式数据；illustrate看pig执行步骤的结果，describe看alias的schema；以轻量级脚本形式跑MapReduce任务，各种爽爆。

1. Word Count

较于中文，英文比较工整，可以根据空格、标点符号进行分词。

A = load '/user/.*/req-temp/text.txt' as (text:chararray);
B = foreach A generate flatten(TOKENIZE(text)) as word;
C = group B by word;
D = foreach C generate COUNT(B), group;

Pig的内置函数TOKENIZE用StringTokenizer来对英文文本进行分词(代码参看这里)，继承于抽象类EvalFunc<T>，返回DataBag词组。为了能统计单个词词频，需要用函数flatten对词组进行打散。抽象类EvalFunc<T>为用于pig语句foreach .. generate ..中的基类，以实现对数据字段的转换操作，其中exec()方法在pig运行期间被调用。

public class TOKENIZE extends EvalFunc<DataBag> {
    TupleFactory mTupleFactory = TupleFactory.getInstance();
    BagFactory mBagFactory = BagFactory.getInstance();

    @Override
    public DataBag exec(Tuple input) throws IOException {
        ...
        DataBag output = mBagFactory.newDefaultBag();
        ...
        String delim = " \",()*";
        ...
        StringTokenizer tok = new StringTokenizer((String)o, delim, false);
        while (tok.hasMoreTokens()) {
            output.add(mTupleFactory.newTuple(tok.nextToken()));
        }
        return output;
        ...
    }
}

2. Ansj中文分词

为了写Pig的UDF，需要添加maven依赖：

<dependency>
	<groupId>org.apache.hadoop</groupId>
	<artifactId>hadoop-common</artifactId>
	<version>${hadoop.version}</version>
	<scope>provided</scope>
</dependency>
    
<dependency>
	<groupId>org.apache.pig</groupId>
	<artifactId>pig</artifactId>
	<version>${pig.version}</version>
	<scope>provided</scope>
</dependency>
	
<dependency>
	<groupId>org.ansj</groupId>
	<artifactId>ansj_seg-all-in-one</artifactId>
	<version>3.0</version>
</dependency>

输入命令hadoop version得到hadoop的版本，输入pig -i得到pig的版本。务必要保证与集群部署的pig版本一致，要不然会报错：

ERROR org.apache.pig.tools.grunt.Grunt - ERROR 1066: Unable to open iterator for alias D

然后依葫芦画瓢，根据TOKENIZE.java修改，得到中文分词Segment.java：

package com.pig.udf;

public class Segment extends EvalFunc<DataBag> {

	TupleFactory mTupleFactory = TupleFactory.getInstance();
    BagFactory mBagFactory = BagFactory.getInstance();

    @Override
    public DataBag exec(Tuple input) throws IOException {
        try {
            if (input==null)
                return null;
            if (input.size()==0)
                return null;
            Object o = input.get(0);
            if (o==null)
                return null;
            DataBag output = mBagFactory.newDefaultBag();
            if (!(o instanceof String)) {
            	int errCode = 2114;
            	String msg = "Expected input to be chararray, but" +
                " got " + o.getClass().getName();
                throw new ExecException(msg, errCode, PigException.BUG);
            }
            
            // filter punctuation
            FilterModifWord.insertStopNatures("w");
            List<Term> words = ToAnalysis.parse((String) o);
            words = FilterModifWord.modifResult(words);
            
            for(Term word: words) {
            	output.add(mTupleFactory.newTuple(word.getName()));
            }
            return output;
        } catch (ExecException ee) {
            throw ee;
        }
    }

    @SuppressWarnings("deprecation")
    @Override
    public Schema outputSchema(Schema input) {
    ...
    }
    ...

ansj支持设置词性的停用词FilterModifWord.insertStopNatures("w");，如此可以去掉标点符号的词。将源代码打包后放在hdfs上，然后通过register jar包调用该UDF：

REGISTER ../piglib/udf-0.0.1-SNAPSHOT-jar-with-dependencies.jar
A = load '/user/.*/renmin.txt' as (text:chararray);
B = foreach A generate flatten(com.pig.udf.Segment(text)) as word;
C = group B by word;
D = foreach C generate COUNT(B), group;

截取人民日报社论的一段：

树好家风，严管才是厚爱。古人说：“居官所以不能清白者，率由家人喜奢好侈使然也。”要看到，好的家风，能系好人生的“第一粒扣子”。“修身、齐家”，才能“治国、平天下”，领导干部首先要“正好家风、管好家人、处好家事”，才能看好“后院”、堵住“后门”。“父母之爱子，则为之计深远”，与其冒着风险给子女留下大笔钱财，不如给子女留下好家风、好作风，那才是让子女受益无穷的东西，才是真正的“为之计深远”。

统计词频如下：

...
(3,能)
(2,要)
(2,计)
(1,与其)
(1,作风)
(1,使然)
(1,修身)
(1,厚爱)
(1,受益)
...

可见，ansj在不加载用户自定义词表的情况下，分词效果并不理想，不能对成语等词正确地分词。