Lucene 实现任意词搜索命中并返回位置信息
原文链接 http://codepub.cn/2017/10/13/lucene-implements-any-word-search-and-returns-location-information/
注:以下为加速网络访问所做的原文缓存,经过重新格式化,可能存在格式方面的问题,或偶有遗漏信息,请以原文为准。
背景
如果把这个标题拆分成两个来讲,那么每一个都很好解决,下文会进行详述,而如果把这两者看做是与条件并加上其它限制,则实现起来比较困难,本文就是要探讨在需求繁多的情况下,如何优雅地实现。比如需求如下
- 保留标点符号,否则去掉标点的话,在标点两边的词可能会匹配上,比如“你好,小甜甜”,去掉标点切分是『你|好|小|甜|甜』,那么『好小』有可能会命中,而如果切分成『你|好|,|小|甜|甜』,则『好小』无法命中
- 只要包含搜索词,要求对任意搜索词均可命中
- 比如“我爱你中国”,不同的分词工具会切分出不同的结果
- 『我|爱|你|中国』或者『我爱你|中国』或者『我|爱|你|中|国』等,那么要求搜索“我爱”或者“爱你”或者“你中”等都要命中
- 需要获取命中词的positions信息
- 还是以“我爱你中国”为例,如果搜“你中”,那么需要返回结果命中,并给出positions信息,例如start=2表示“你中”在原文本中是从第2个位置开始
- 可以设置slop
- 什么是slop,简单来说slop是指两个项的位置之间允许的最大间隔距离
- 为什么要设置slop呢?比如小黄文为了防止敏感词被屏蔽,会在敏感词中间加上干扰词,例如
性%=$虐待
,那么直接搜性虐待
无法命中 - 只要设置slop为3,即相当于搜
性[***]虐待
,这里面的[***]
就代表slop为3,可以匹配任意三个字
- 不要存储Field的值
- 如果可以存储的话,可以通过Document获取原文本,再用TokenStream分析该文本,使用QueryScore初始化TokenStream的分析结果,遍历每个token根据TokenScore的得分判断是否命中,若命中则输出位置信息或者起始偏移量即可
- 但是存储Field的值是需要占用硬盘空间的,当需要索引海量的文本的时候,会导致索引体积非常大,搜索性能变差
- 当然还可以通过将索引拆分成多份存储实现降低索引体积的目的,这也是一个方法,不过治标不治本
- 不要存储TermVectors
- 同样地如果可以存储的话,可以通过TermVectors获取Terms再遍历TermsEnum,获取PostingsEnum得到positions信息,缺点在于只能实现单个字(Term)的搜索匹配
- 但是存储TermVectors同样占用硬盘空间,为了缩小索引体积,不要存储
- 实现与逻辑,比如搜索“你中 & 我爱”表示两个词都要命中
- 实现或逻辑,比如搜索“你中 | 我爱”表示两个词至少要有一个命中
如果想要实现百分之百的任意词搜索命中,那么只能按字切分,因为没有任何分词工具能够保证切出来的词与搜索词是一致的。在上面说到为了达到匹配干扰词的目的,需要设置slop,但是会有一定的误判率,本来不该匹配的在设置slop之后也匹配上了。除了设置slop这种方式,还有一种方法,就是在索引阶段只保留汉字,其它的标点符号和干扰符号统统去掉,当然这也存在一定的误判率,而且获取的positions信息已经不是原文本中正确的positions信息了。两种方式的权衡与取舍可以根据业务需求而定,这两种方式都不会漏判,但是均会有误判。
简单需求的实现
如果并不要求满足上面所有的需求,而仅仅满足其中任何一个,那么实现起来都是非常简单的,以下代码均基于Lucene 5.5.0实现,示例如下。
仅要求任意搜索词命中
@org.junit.Test
public void testAnyMatch() throws IOException {
RAMDirectory ramDirectory = new RAMDirectory();
IndexWriter indexWriter = new IndexWriter(ramDirectory, new IndexWriterConfig(new StandardAnalyzer()));
Document document = new Document();
document.add(new TextField("content", "我爱你中国", Field.Store.NO));
indexWriter.addDocument(document);
indexWriter.commit();
IndexSearcher indexSearcher = new IndexSearcher(DirectoryReader.open(indexWriter));
PhraseQuery phraseQuery = new PhraseQuery.Builder().add(new Term("content", "你")).add(new Term("content", "中")).setSlop(0).build();
TopDocs search = indexSearcher.search(phraseQuery, Integer.MAX_VALUE);
System.out.println(search.totalHits);
//OR search like this
MultiPhraseQuery multiPhraseQuery = new MultiPhraseQuery();
Term first = new Term("content", "你");
Term second = new Term("content", "中");
multiPhraseQuery.add(new Term[]{first, second});
search = indexSearcher.search(multiPhraseQuery, Integer.MAX_VALUE);
System.out.println(search.totalHits);
}
可以存储Field的值
@org.junit.Test
public void testStoreFieldMatch() throws IOException, InvalidTokenOffsetsException {
RAMDirectory ramDirectory = new RAMDirectory();
IndexWriter indexWriter = new IndexWriter(ramDirectory, new IndexWriterConfig(new StandardAnalyzer()));
Document document = new Document();
document.add(new TextField("content", "我爱你中国", Field.Store.YES));
indexWriter.addDocument(document);
indexWriter.commit();
IndexSearcher indexSearcher = new IndexSearcher(DirectoryReader.open(indexWriter));
PhraseQuery phraseQuery = new PhraseQuery.Builder().add(new Term("content", "你")).add(new Term("content", "中")).setSlop(0).build();
TopDocs search = indexSearcher.search(phraseQuery, Integer.MAX_VALUE);
ScoreDoc[] scoreDocs = search.scoreDocs;
for (ScoreDoc scoreDoc : scoreDocs) {
String content = indexSearcher.doc(scoreDoc.doc).get("content");
TokenStream contentStream = new StandardAnalyzer().tokenStream("content", content);
CharTermAttribute charTermAttribute = contentStream.addAttribute(CharTermAttribute.class);
OffsetAttribute offsetAttribute = contentStream.addAttribute(OffsetAttribute.class);
QueryScorer queryScorer = new QueryScorer(phraseQuery);
queryScorer.setMaxDocCharsToAnalyze(Integer.MAX_VALUE);
TokenStream init = queryScorer.init(contentStream);
if (init != null) {
contentStream = init;
}
contentStream.reset();
queryScorer.startFragment(null);
int startOffset, endOffset;
for (boolean next = contentStream.incrementToken(); next && (offsetAttribute.startOffset() < Integer.MAX_VALUE); next = contentStream.incrementToken()) {
startOffset = offsetAttribute.startOffset();
endOffset = offsetAttribute.endOffset();
if (startOffset > content.length() || endOffset > content.length()) {
throw new InvalidTokenOffsetsException("Token " + charTermAttribute.toString() + " exceeds length of provided text sized " + content.length());
}
float res = queryScorer.getTokenScore();
if (res > Float.valueOf(0) && startOffset <= endOffset) {
System.out.println("hits: " + content.substring(startOffset, endOffset) + ", start: " + startOffset);
}
}
contentStream.close();
}
}
可以存储TermVectors的值
@org.junit.Test
public void testTermVectorsMatch() throws IOException, InvalidTokenOffsetsException {
RAMDirectory ramDirectory = new RAMDirectory();
IndexWriter indexWriter = new IndexWriter(ramDirectory, new IndexWriterConfig(new StandardAnalyzer()));
Document document = new Document();
FieldType fieldType = new FieldType();
fieldType.setIndexOptions(IndexOptions.DOCS_AND_FREQS_AND_POSITIONS_AND_OFFSETS);
fieldType.setStoreTermVectorPositions(true);
fieldType.setStoreTermVectors(true);
document.add(new Field("content", "我爱你中国", fieldType));
indexWriter.addDocument(document);
indexWriter.commit();
IndexSearcher indexSearcher = new IndexSearcher(DirectoryReader.open(indexWriter));
Term searchTerm = new Term("content", "中");
PhraseQuery phraseQuery = new PhraseQuery.Builder().add(searchTerm).setSlop(0).build();
TopDocs search = indexSearcher.search(phraseQuery, Integer.MAX_VALUE);
ScoreDoc[] scoreDocs = search.scoreDocs;
for (ScoreDoc scoreDoc : scoreDocs) {
Terms content = indexSearcher.getIndexReader().getTermVector(scoreDoc.doc, "content");
TermsEnum iterator = content.iterator();
BytesRef bytesRef;
while ((bytesRef = iterator.next()) != null) {
PostingsEnum postings = iterator.postings(null, PostingsEnum.ALL);
if (postings.nextDoc() != Spans.NO_MORE_DOCS) {
for (int i = 0; i < postings.freq(); i++) {
if (searchTerm.text().equals(bytesRef.utf8ToString())) {
System.out.println("hits: " + bytesRef.utf8ToString() + ", start: " + postings.nextPosition());
}
}
}
}
}
}
复杂需求的实现
实现某一个简单的需求就不再举例了,下面要讲解如何实现复杂的需求,也就是说,要同时满足上面的需求列表,而不仅仅是只满足其中的某一条。首先需要解决的就是分词之后保留标点符号的问题,在Lucene中,我并没有找到原生的支持保留标点符号的Analyzer,于是只能自己造轮子了。
保留标点符号的分词器
import lombok.extern.log4j.Log4j2;
import org.apache.lucene.analysis.Analyzer;
import org.apache.lucene.analysis.TokenStream;
import org.apache.lucene.analysis.Tokenizer;
import org.apache.lucene.analysis.core.LowerCaseFilter;
import org.apache.lucene.analysis.pattern.PatternTokenizer;
import java.util.regex.Pattern;
@Log4j2
public class ReservePunctuationAnalyzer extends Analyzer {
public ReservePunctuationAnalyzer() {
}
@Override
protected TokenStreamComponents createComponents(String fieldName) {
final Tokenizer source;
source = new PatternTokenizer(Pattern.compile(""), -1);
TokenStream result = new LowerCaseFilter(source);
return new TokenStreamComponents(source, result);
}
}
分词测试如下
@Test
public void test() throws IOException {
String input = "你好,小甜甜。";
TokenStream test = new ReservePunctuationAnalyzer().tokenStream("test", input);
CharTermAttribute charTermAttribute = test.addAttribute(CharTermAttribute.class);
OffsetAttribute offsetAttribute = test.addAttribute(OffsetAttribute.class);
test.reset();
while (test.incrementToken()) {
System.out.println("token:[" + charTermAttribute + "], offset:[" + offsetAttribute.startOffset() + "]");
}
test.close();
}
分词结果输出如下
token:[你], offset:[0] token:[好], offset:[1] token:[,], offset:[2] token:[小], offset:[3] token:[甜], offset:[4] token:[甜], offset:[5] token:[。], offset:[6]
任意词搜索命中并返回positions信息
下面再来解决在不存储Field、不存储TermVectors的情况下,如何实现任意词搜索命中并返回positions信息,同时还可以设置slop的值。要实现这些功能就需要用到SpanQuery及其一系列的子类,先来看一张继承关系图,这些都是即将要使用到的类。
SpanQuery的doc注释很简单,就一句话“Base class for span-based queries”,基于跨度查询的基类。而真正具有实际作用的是其各个子类。
import org.apache.lucene.document.Document;
import org.apache.lucene.document.Field;
import org.apache.lucene.document.FieldType;
import org.apache.lucene.document.LongField;
import org.apache.lucene.index.*;
import org.apache.lucene.search.IndexSearcher;
import org.apache.lucene.search.spans.SpanNearQuery;
import org.apache.lucene.search.spans.SpanTermQuery;
import org.apache.lucene.search.spans.SpanWeight;
import org.apache.lucene.search.spans.Spans;
import org.apache.lucene.store.RAMDirectory;
import java.io.IOException;
import java.util.List;
import static org.apache.lucene.search.spans.SpanNearQuery.newOrderedNearQuery;
/**
* <p>
* Created by wangxu on 2017/10/13 14:29.
* </p>
* <p>
* Description: TODO
* </p>
*
* @author Wang Xu
* @version V1.0.0
* @since V1.0.0 <br/>
* WebSite: http://codepub.cn <br>
* Licence: Apache v2 License
*/
public class SpanNearQueryDemo {
@org.junit.Test
public void test() throws IOException {
String input = "现有的中文分词算法可分为三大类:基于字符串匹配的类基分词方法、基于理解的分词方法和基于统计的分词方法。";
RAMDirectory ramDirectory = new RAMDirectory();
IndexWriterConfig indexWriterConfig = new IndexWriterConfig(new ReservePunctuationAnalyzer());
try (IndexWriter indexWriter = new IndexWriter(ramDirectory, indexWriterConfig)) {
Document document = new Document();
FieldType fieldType = new FieldType();
fieldType.setIndexOptions(IndexOptions.DOCS_AND_FREQS_AND_POSITIONS);
Field field = new Field("title", input, fieldType);
LongField IDX = new LongField("IDX", 1, Field.Store.YES);
document.add(field);
document.add(IDX);
indexWriter.addDocument(document);
input = "计算机算法是很难很复杂滴";
document = new Document();
field = new Field("title", input, fieldType);
IDX = new LongField("IDX", 2, Field.Store.YES);
document.add(field);
document.add(IDX);
indexWriter.addDocument(document);
input = "计算机算法可以大幅度提升程序性能";
document = new Document();
field = new Field("title", input, fieldType);
IDX = new LongField("IDX", 3, Field.Store.YES);
document.add(field);
document.add(IDX);
indexWriter.addDocument(document);
indexWriter.commit();
IndexSearcher indexSearcher = new IndexSearcher(DirectoryReader.open(ramDirectory));
SpanTermQuery first = new SpanTermQuery(new Term("title", "类"));
SpanTermQuery second = new SpanTermQuery(new Term("title", "基"));
SpanNearQuery spanNearQuery = newOrderedNearQuery("title").addClause(first).addClause(second).build();
SpanWeight weight = spanNearQuery.createWeight(indexSearcher, true);
List<LeafReaderContext> leaves = indexSearcher.getIndexReader().getContext().leaves();
for (LeafReaderContext leaf : leaves) {
Spans spans = weight.getSpans(leaf, SpanWeight.Postings.POSITIONS);
while (spans.nextDoc() != Spans.NO_MORE_DOCS) {
Document doc = leaf.reader().document(spans.docID());
while (spans.nextStartPosition() != Spans.NO_MORE_POSITIONS) {
System.out.println("doc id = " + spans.docID() + ", doc IDX= " + doc.get("IDX") + ", start position = " + spans.startPosition() + ", end " +
"position = " + spans.endPosition());
}
}
}
//================================================================
// 输出结果是
// doc id = 0, doc IDX= 1, start position = 24, end position = 26
//================================================================
System.out.println();
//修改slop,设置1,默认是0
spanNearQuery = newOrderedNearQuery("title").addClause(first).addClause(second).setSlop(1).build();
weight = spanNearQuery.createWeight(indexSearcher, true);
leaves = indexSearcher.getIndexReader().getContext().leaves();
for (LeafReaderContext leaf : leaves) {
Spans spans = weight.getSpans(leaf, SpanWeight.Postings.POSITIONS);
while (spans.nextDoc() != spans.NO_MORE_DOCS) {
Document doc = leaf.reader().document(spans.docID());
while (spans.nextStartPosition() != spans.NO_MORE_POSITIONS) {
System.out.println("doc id = " + spans.docID() + ", doc IDX= " + doc.get("IDX") + ", start position = " + spans.startPosition() + ", end " +
"position = " + spans.endPosition());
}
}
}
//================================================================
// 输出结果是
// doc id = 0, doc IDX= 1, start position = 14, end position = 17
// doc id = 0, doc IDX = 1, start position = 24, end position = 26
//================================================================
}
}
}
实现逻辑与查询
通过上面的图,想必你也知道,Lucene官方并不对SpanAndQuery提供支持,在Lucene的官方讨论组中,有人发起过支持SpanAndQuery的issue,但是一直没有获得官方的回应。不过已经有商业公司实现了这种搜索技术,公司名称是SearchTechnologies,API参见SpanAndQuery,但是并不开源(So Sad),我没有找到其实现的具体源码,如果你知道的话,烦请告知我一下。
既然官方不予支持,那就只能自己造轮子了,逻辑上来讲,也不复杂,有两种方式可以实现。
第一种方式,从词的角度,例如『爱你』和『你中』两个搜索词实现逻辑与,那么只要分别地把每一个搜索词都单独搜一下,最后在命中结果中取交集,就可以实现逻辑与的功能,代码写起来也很简单,在此不予示例。
第二种方式,从Term的角度,例如『爱你』如果按字切分,那么能够切成两个Term,分别是『爱』和『你』,这时候使用SpanNearQuery构造查询语句,加上一个很大的slop,但是不管slop多大,它总是有上限的,万一两个Term之间的距离超过slop,同样无法命中,所以说这种实现方式是存在漏洞的,除非你确定你的两个Term之间的距离不会超过某个具体的slop值,那么可以使用之。
注意在使用SpanNearQuery获取positions信息的时候,你不能够同时保证按字切分,又可以在两个搜索词之间设置slop值,这是因为如果用BooleanQuery去包装两个SpanNearQuery,那么将丢失positions信息。如果不按字切分,那么切出来的某个词就是一个Term,即将『爱你』和『你中』看成是两个Term,这时候是可以设置slop值的。如果按字切分,那么切成『爱|你』和『你|中』,实现逻辑与,如果设置slop值,相当于是『爱|slop值|你|slop值|你|slop值|中』,已经与逻辑与不匹配了,逻辑与的本意是『爱|slop值为0|你|slop为任意值|你|slop值为0|中』,请细细体会。
实现逻辑或查询
官方已经对或逻辑提供了支持,就是SpanOrQuery,直接操练起来即可。
import org.apache.lucene.document.Document;
import org.apache.lucene.document.Field;
import org.apache.lucene.document.FieldType;
import org.apache.lucene.document.LongField;
import org.apache.lucene.index.*;
import org.apache.lucene.search.IndexSearcher;
import org.apache.lucene.search.spans.*;
import org.apache.lucene.store.RAMDirectory;
import java.io.IOException;
import java.util.List;
import static org.apache.lucene.search.spans.SpanNearQuery.newOrderedNearQuery;
/**
* <p>
* Created by wangxu on 2017/10/13 14:29.
* </p>
* <p>
* Description: TODO
* </p>
*
* @author Wang Xu
* @version V1.0.0
* @since V1.0.0 <br/>
* WebSite: http://codepub.cn <br>
* Licence: Apache v2 License
*/
public class SpanOrQueryDemo {
@org.junit.Test
public void test() throws IOException {
String input = "现有的中文分词算法可分为三大类:基于字符串匹配的类基分词方法、基于理解的分词方法和基于统计的分词方法。";
RAMDirectory ramDirectory = new RAMDirectory();
IndexWriterConfig indexWriterConfig = new IndexWriterConfig(new ReservePunctuationAnalyzer());
try (IndexWriter indexWriter = new IndexWriter(ramDirectory, indexWriterConfig)) {
Document document = new Document();
FieldType fieldType = new FieldType();
fieldType.setIndexOptions(IndexOptions.DOCS_AND_FREQS_AND_POSITIONS);
Field field = new Field("title", input, fieldType);
LongField IDX = new LongField("IDX", 1, Field.Store.YES);
document.add(field);
document.add(IDX);
indexWriter.addDocument(document);
input = "计算机算法是很难很复杂滴";
document = new Document();
field = new Field("title", input, fieldType);
IDX = new LongField("IDX", 2, Field.Store.YES);
document.add(field);
document.add(IDX);
indexWriter.addDocument(document);
input = "计算机算法可以大幅度提升程序性能";
document = new Document();
field = new Field("title", input, fieldType);
IDX = new LongField("IDX", 3, Field.Store.YES);
document.add(field);
document.add(IDX);
indexWriter.addDocument(document);
indexWriter.commit();
IndexSearcher indexSearcher = new IndexSearcher(DirectoryReader.open(ramDirectory));
SpanTermQuery first = new SpanTermQuery(new Term("title", "类"));
SpanTermQuery second = new SpanTermQuery(new Term("title", "基"));
SpanNearQuery spanNearQueryFirst = newOrderedNearQuery("title").addClause(first).addClause(second).build();
first = new SpanTermQuery(new Term("title", "算"));
second = new SpanTermQuery(new Term("title", "法"));
SpanNearQuery spanNearQuerySecond = newOrderedNearQuery("title").addClause(first).addClause(second).build();
SpanOrQuery spanOrQuery = new SpanOrQuery(spanNearQueryFirst, spanNearQuerySecond);
SpanWeight weight = spanOrQuery.createWeight(indexSearcher, true);
List<LeafReaderContext> leaves = indexSearcher.getIndexReader().getContext().leaves();
for (LeafReaderContext leaf : leaves) {
Spans spans = weight.getSpans(leaf, SpanWeight.Postings.POSITIONS);
while (spans.nextDoc() != Spans.NO_MORE_DOCS) {
Document doc = leaf.reader().document(spans.docID());
while (spans.nextStartPosition() != Spans.NO_MORE_POSITIONS) {
System.out.println("doc id = " + spans.docID() + ", doc IDX= " + doc.get("IDX") + ", start position = " + spans.startPosition() + ", end " +
"position = " + spans.endPosition());
}
}
}
//================================================================
// 输出结果是
// doc id = 0, doc IDX= 1, start position = 7, end position = 9
// doc id = 0, doc IDX= 1, start position = 24, end position = 26
// doc id = 1, doc IDX= 2, start position = 3, end position = 5
// doc id = 2, doc IDX= 3, start position = 3, end position = 5
//================================================================
}
}
}
实现SpanAllNearQuery
这是一个附加功能,因为目前还没有碰到这样的需求,但是这种查询实现起来非常有意思,所以在此简单讲解一下。这个问题的来源是有人提了个issue,请求官方支持SpanAllNearQuery,但是同样地,官方不理不睬。果然公益的就是拽啊,完全不倾听用户的需求,不像商业公司,为了赚用户的钱,只要用户有需求,就尽力实现。
那么这个需求是什么样的呢?简单表示如下a WITHIN 5 WORDS OF (b AND c)
,还可以把它换一种方式理解(a WITHIN 5 WORDS OF b) AND (a WITHIN 5 WORDS OF c)
,就是说我要查询,在a的前面5个或者后面5个token中出现b和c的所有结果集。要实现这个功能,需要借助于SpanNotQuery和SpanOrQuery,SpanOrQuery在实现或逻辑中已经介绍过了,那么SpanNotQuery又是什么意思呢?举例如下
SpanNotQuery(a, b, 5, 5)表示在a的前5个或者后5个token中不能出现b SpanNotQuery(a, c, 5, 5)表示在a的前5个或者后5个token中不能出现c
下面先从逻辑上先实现这个需求,要获得在a的前面或后面5个token中出现b和c,需要将其反转理解,先查询在a的前面5个或者后面5个token中不能出现b『SpanNotQuery(a, b, 5, 5)』或者在a的前面5个或者后面5个token中不能出现c的结果『SpanNotQuery(a, c, 5, 5)』,再用SpanOrQuery来组合『SpanNotQuery(a, b, 5, 5)』和『SpanNotQuery(a, c, 5, 5)』实现或逻辑,最后用SpanNotQuery排除掉SpanOrQuery的结果集,那么剩下的就是在a的前面5个或者后面5个能出现b也能出现c的结果。
import org.apache.lucene.analysis.core.WhitespaceAnalyzer;
import org.apache.lucene.document.Document;
import org.apache.lucene.document.Field;
import org.apache.lucene.document.IntField;
import org.apache.lucene.document.TextField;
import org.apache.lucene.index.DirectoryReader;
import org.apache.lucene.index.IndexWriter;
import org.apache.lucene.index.IndexWriterConfig;
import org.apache.lucene.index.Term;
import org.apache.lucene.search.IndexSearcher;
import org.apache.lucene.search.ScoreDoc;
import org.apache.lucene.search.TopDocs;
import org.apache.lucene.search.spans.SpanNotQuery;
import org.apache.lucene.search.spans.SpanOrQuery;
import org.apache.lucene.search.spans.SpanTermQuery;
import org.apache.lucene.store.RAMDirectory;
import java.io.IOException;
/**
* <p>
* Created by wangxu on 2017/06/16 16:02.
* </p>
* <p>
* Description: TODO
* </p>
*
* @author Wang Xu
* @version V1.0.0
* @since V1.0.0 <br/>
* WebSite: http://codepub.cn <br>
* Licence: Apache v2 License
*/public class SpanAllNearQueryDemo {
@org.junit.Test
public void test() throws IOException {
RAMDirectory ramDirectory = new RAMDirectory();
IndexWriter indexWriter = new IndexWriter(ramDirectory, new IndexWriterConfig(new WhitespaceAnalyzer()));
Document document = new Document();
document.add(new TextField("key", "X b X X X X a X X X X c X", Field.Store.YES));//命中
document.add(new IntField("IDX", 1, Field.Store.YES));
indexWriter.addDocument(document);
document = new Document();
document.add(new TextField("key", "X X X X X b a c X X X X X", Field.Store.YES));//命中
document.add(new IntField("IDX", 2, Field.Store.YES));
indexWriter.addDocument(document);
document = new Document();
document.add(new TextField("key", "X b X X X X a a X X X X c", Field.Store.YES));//不命中,不能同时以两个a为中心,两个a必选其一
document.add(new IntField("IDX", 3, Field.Store.YES));
indexWriter.addDocument(document);
document = new Document();
document.add(new TextField("key", "X b X X X X X a X X X X c", Field.Store.YES));//不命中
document.add(new IntField("IDX", 4, Field.Store.YES));
indexWriter.addDocument(document);
document = new Document();
document.add(new TextField("key", "X b X X X X a X X X X X c", Field.Store.YES));//不命中
document.add(new IntField("IDX", 5, Field.Store.YES));
indexWriter.addDocument(document);
document = new Document();
document.add(new TextField("key", "b X X X X X a X X X X X c", Field.Store.YES));//不命中
document.add(new IntField("IDX", 6, Field.Store.YES));
indexWriter.addDocument(document);
document = new Document();
document.add(new TextField("key", "b X X X X X a X X X X X X", Field.Store.YES));//不命中
document.add(new IntField("IDX", 7, Field.Store.YES));
indexWriter.addDocument(document);
document = new Document();
document.add(new TextField("key", "X X X X X X a X X X X X X", Field.Store.YES));//不命中
document.add(new IntField("IDX", 8, Field.Store.YES));
indexWriter.addDocument(document);
indexWriter.commit();
IndexSearcher indexSearcher = new IndexSearcher(DirectoryReader.open(indexWriter));
SpanTermQuery a = new SpanTermQuery(new Term("key", "a"));
SpanTermQuery b = new SpanTermQuery(new Term("key", "b"));
SpanTermQuery c = new SpanTermQuery(new Term("key", "c"));
SpanOrQuery exclude = new SpanOrQuery(new SpanNotQuery(a, b, 5, 5), new SpanNotQuery(a, c, 5, 5));
//排除在a的前5个或者后5个不能出现b也不能出现c的document,那么剩下的就是在a的前5个token或者后5个token能够出现b和c的document
SpanNotQuery spanNotQuery = new SpanNotQuery(a, exclude);
TopDocs search = indexSearcher.search(spanNotQuery, Integer.MAX_VALUE);
ScoreDoc[] scoreDocs = search.scoreDocs;
for (ScoreDoc scoreDoc : scoreDocs) {
System.out.println("hist IDX: " + indexSearcher.doc(scoreDoc.doc).get("IDX"));
}
indexSearcher.getIndexReader().close();
indexWriter.close();
}
}
SpanNearQuery实现通配符查询
import com.yuewen.nrzx.character.analyzer.ReservePunctuationAnalyzer;
import org.apache.lucene.document.Document;
import org.apache.lucene.document.Field;
import org.apache.lucene.document.FieldType;
import org.apache.lucene.document.LongField;
import org.apache.lucene.index.*;
import org.apache.lucene.search.IndexSearcher;
import org.apache.lucene.search.TopDocs;
import org.apache.lucene.search.WildcardQuery;
import org.apache.lucene.search.spans.SpanMultiTermQueryWrapper;
import org.apache.lucene.search.spans.SpanNearQuery;
import org.apache.lucene.search.spans.SpanQuery;
import org.apache.lucene.search.spans.SpanTermQuery;
import org.apache.lucene.store.RAMDirectory;
import java.io.IOException;
public class SpanNearQueryAndWildcardQueryDemo {
@org.junit.Test
public void test() throws IOException {
String input = "现有的中文分词算法可分为三大类:基于字符串匹配的类基分词方法、基于理解的分词方法和基于统计的分词方法。";
RAMDirectory ramDirectory = new RAMDirectory();
IndexWriterConfig indexWriterConfig = new IndexWriterConfig(new ReservePunctuationAnalyzer());
try (IndexWriter indexWriter = new IndexWriter(ramDirectory, indexWriterConfig)) {
Document document = new Document();
FieldType fieldType = new FieldType();
fieldType.setIndexOptions(IndexOptions.DOCS_AND_FREQS_AND_POSITIONS);
Field field = new Field("title", input, fieldType);
LongField IDX = new LongField("IDX", 1, Field.Store.YES);
document.add(field);
document.add(IDX);
indexWriter.addDocument(document);
input = "计算机算法是很难很复杂滴";
document = new Document();
field = new Field("title", input, fieldType);
IDX = new LongField("IDX", 2, Field.Store.YES);
document.add(field);
document.add(IDX);
indexWriter.addDocument(document);
input = "计算机算法可以大幅度提升程序性能";
document = new Document();
field = new Field("title", input, fieldType);
IDX = new LongField("IDX", 3, Field.Store.YES);
document.add(field);
document.add(IDX);
indexWriter.addDocument(document);
indexWriter.commit();
IndexSearcher indexSearcher = new IndexSearcher(DirectoryReader.open(ramDirectory));
// 用?和*均可以实现SpanNearQuery的通配符查询,但是注意*在通配符查询中表示可以匹配0个或多个字符
// 但是在SpanQuery中只能匹配相当于slop=1的情形,不能匹配slop大于1的情形
SpanTermQuery first = new SpanTermQuery(new Term("title", "复"));
SpanQuery wildcard = new SpanMultiTermQueryWrapper<>(new WildcardQuery(new Term("title", "?")));
SpanTermQuery last = new SpanTermQuery(new Term("title", "滴"));
SpanNearQuery spanNearQuery = new SpanNearQuery.Builder("title", true).addClause(first).addClause(wildcard).addClause(last).build();
TopDocs search = indexSearcher.search(spanNearQuery, Integer.MAX_VALUE);
System.out.println("IDX: " + indexSearcher.doc(search.scoreDocs[0].doc).get("IDX"));
wildcard = new SpanMultiTermQueryWrapper<>(new WildcardQuery(new Term("title", "*")));
spanNearQuery = new SpanNearQuery.Builder("title", true).addClause(first).addClause(wildcard).addClause(last).build();
search = indexSearcher.search(spanNearQuery, Integer.MAX_VALUE);
System.out.println("IDX: " + indexSearcher.doc(search.scoreDocs[0].doc).get("IDX"));
}
}
}
实验
服务器 CPU 以及内存信息
$ cat /proc/cpuinfo | grep name | cut -f2 -d: | uniq -c
24 Intel(R) Xeon(R) CPU E5-2420 v2 @ 2.20GHz
$ free -g
||total|used|free|shared|buffers|cached| |--|--|--|--|--|--| |Mem|62|56|6|0|0|1| |-/+ buffers/cache|54|8||||| |Swap|1|0|1|||||
在公司内部,仅仅索引了十分之一的文档(Document数量:20023911),鉴于没有存储Field,也没有存储TermVectors,索引不算太大,简单测试了下,如果存储TermVectors的话,索引会从112GB增长到162GB,如果再存储Field的话,那么索引要超过200GB。此处的实验只是简单的单次搜索,没有测试与逻辑和或逻辑情况下的搜索情况。搜索阶段实验的结果如下所示
线程数目 | 搜索总次数 | 命中次数 | 搜索总耗时 | 平均单次耗时 | 搜索加构建Query耗时 | 平均单次耗时 | 索引大小 |
---|---|---|---|---|---|---|---|
1 | 18820 | 18139 | 36001,698ms | 1912.95ms | 36011,989ms | 1913.50ms | 112GB |
5 | 18820 | 18139 | 30775,283ms | 1635.24ms | 30785,538ms | 1635.79ms | 112GB |
10 | 18820 | 18139 | 21637,515ms | 1149.71ms | 21647,953ms | 1150.26ms | 112GB |
50 | 18820 | 18139 | 21572,506ms | 1146.25ms | 21583,101ms | 1146.82ms | 112GB |
索引阶段并没有做详细完备的实验,只是简单拉取了一点数据,记录如下,仅供参考。索引2080550个Document,统计从数据库拉取数据加更新数据到索引耗时2133s,平均每次耗时0.001s。只计算更新数据到索引中,不计算从数据库拉取数据耗时464s,平均每次耗时0.0002s,可见索引速度也是很快的,这全部得益于Lucene的优良设计。