文本近似最近邻训练 (TextApproxNearestNeighborTrainBatchOp)

Java 类名：com.alibaba.alink.operator.batch.similarity.TextApproxNearestNeighborTrainBatchOp

Python 类名：TextApproxNearestNeighborTrainBatchOp

功能介绍

文本相似度是在字符串相似度的基础上，基于词，计算两两文章或者句子之间的相似度，文章或者句子需要以空格分割的文本，计算方式和字符串相似度类似：支持SimHashHamming，MinHash和Jaccard三种近似相似度计算方式，通过选择metric参数可计算不同的相似度。

该功能由训练和预测组成，支持计算1. 求最近邻topN 2. 求radius范围内的邻居。该功能由预测时候的topN和radius参数控制, 如果填写了topN，则输出最近邻，如果填写了radius，则输出radius范围内的邻居。

SimhashHamming（SimHash_Hamming_Distance)相似度=1-距离/64.0，应选择metric的参数为SIMHASH_HAMMING_SIM。

MinHash应选择metric的参数为MINHASH_SIM。

Jaccard应选择metric的参数为JACCARD_SIM。

参数说明

名称	中文名称	描述	类型	是否必须？	取值范围	默认值
idCol	id列名	id列名	String	✓
selectedCol	选中的列名	计算列对应的列名	String	✓
metric	距离类型	用于计算的距离类型	String		“SIMHASH_HAMMING_SIM”, “SIMHASH_HAMMING”, “MINHASH_JACCARD_SIM”, “JACCARD_SIM”	“SIMHASH_HAMMING_SIM”
numBucket	分桶个数	分桶个数	Integer			10
numHashTables	哈希表个数	哈希表的数目	Integer			10
seed	采样种子	采样种子	Long			0

代码示例

Python 代码

from pyalink.alink import *

import pandas as pd

useLocalEnv(1)

df = pd.DataFrame([
    [0, "a b c d e", "a a b c e"],
    [1, "a a c e d w", "a a b b e d"],
    [2, "c d e f a", "b b c e f a"],
    [3, "b d e f h", "d d e a c"],
    [4, "a c e d m", "a e e f b c"]
])

inOp = BatchOperator.fromDataframe(df, schemaStr='id long, text1 string, text2 string')

train = TextApproxNearestNeighborTrainBatchOp().setIdCol("id").setSelectedCol("text1").setMetric("SIMHASH_HAMMING_SIM").linkFrom(inOp)
predict = TextApproxNearestNeighborPredictBatchOp().setSelectedCol("text2").setTopN(3).linkFrom(train, inOp)
predict.print()

Java 代码

import org.apache.flink.types.Row;

import com.alibaba.alink.operator.batch.BatchOperator;
import com.alibaba.alink.operator.batch.similarity.TextApproxNearestNeighborPredictBatchOp;
import com.alibaba.alink.operator.batch.similarity.TextApproxNearestNeighborTrainBatchOp;
import com.alibaba.alink.operator.batch.source.MemSourceBatchOp;
import org.junit.Test;

import java.util.Arrays;
import java.util.List;

public class TextApproxNearestNeighborTrainBatchOpTest {
	@Test
	public void testTextApproxNearestNeighborTrainBatchOp() throws Exception {
		List <Row> df = Arrays.asList(
			Row.of(0, "a b c d e", "a a b c e"),
			Row.of(1, "a a c e d w", "a a b b e d"),
			Row.of(2, "c d e f a", "b b c e f a"),
			Row.of(3, "b d e f h", "d d e a c"),
			Row.of(4, "a c e d m", "a e e f b c")
		);
		BatchOperator <?> inOp = new MemSourceBatchOp(df, "id int, text1 string, text2 string");
		BatchOperator <?> train = new TextApproxNearestNeighborTrainBatchOp().setIdCol("id").setSelectedCol("text1")
			.setMetric("SIMHASH_HAMMING_SIM").linkFrom(inOp);
		BatchOperator <?> predict = new TextApproxNearestNeighborPredictBatchOp().setSelectedCol("text2").setTopN(3)
			.linkFrom(train, inOp);
		predict.print();
	}
}

运行结果

id	text1	text2
0	a b c d e	{“ID”:“[0,1,2]”,“METRIC”:“[0.953125,0.921875,0.90625]”}
1	a a c e d w	{“ID”:“[0,1,4]”,“METRIC”:“[0.9375,0.90625,0.859375]”}
2	c d e f a	{“ID”:“[0,1,4]”,“METRIC”:“[0.890625,0.859375,0.8125]”}
3	b d e f h	{“ID”:“[4,2,1]”,“METRIC”:“[0.9375,0.90625,0.890625]”}
4	a c e d m	{“ID”:“[1,0,4]”,“METRIC”:“[0.921875,0.921875,0.90625]”}