ElasticStack系列之十三 & 联想补全策略
Posted 星火燎原智勇
tags:
篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了ElasticStack系列之十三 & 联想补全策略相关的知识,希望对你有一定的参考价值。
业务需求
1. 实现搜索引擎前缀搜索功能(中文,拼音前缀查询及简拼前缀查询功能)
2. 实现摘要全文检索功能,及标题加权处理功能(按照标题权值高内容权值相对低的权值分配规则,按照索引的相关性进行排序,列出前20条相关性最高的文章)
前缀搜索
中文搜索:
1. 搜索“刘”,匹配到“刘德华”、“刘斌”、“刘德志”
2. 搜索“刘德”,匹配到“刘德华”、“刘德志”
小结:搜索的文字需要匹配到集合中所有名字的子集。
全拼搜索:
1. 搜索“li”,匹配到“刘德华”、“刘斌”、“刘德志”
2. 搜索“liud”,匹配到“刘德华”、“刘德”
3. 搜索“liudeh”,匹配到“刘德华”
小结:搜索的文字转换成拼音后,需要匹配到集合中所有名字转成拼音后的子集
简拼搜索:
1. 搜索“w”,匹配到“我是中国人”,“我爱我的祖国”
2. 搜索“wszg”,匹配到“我是中国人”
小结:搜索的文字取拼音首字母进行组合,需要匹配到组合字符串中前缀匹配的子集
解决方案
方案一:
将 “like” 搜索的字段的 中、英简拼、英全拼 分别用索引的三个字段来进行存储并且 不进行分词,最简单直接,检索索引数据的时候进行通配符查询(like查询),从这三个字段中分别进行搜索,查询匹配的记录然后返回。
优势:存储格式简单,倒排索引存储的数据量最少。
缺点:like 索引数据的时候开销比较大 prefix 查询比 term 查询开销大得多
方案二:
将 中、中简拼、中全拼 用一个字段衍生出三个字段(multi-field)来存储三种数据,并且分词器filter 采用 edge_ngram 类型对分词的数据进行分词处理存储到倒排索引中,当检索索引数据时,检索所有字段的数据。
优势:格式紧凑,检索索引数据的时候采用 term 全匹配规则,也无需对入参进行分词,查询效率高。
缺点:采用以空间换时间的策略,但是对索引来说可以接受。采用衍生字段来存储,增加了存储及检索的复杂度,对于三个字段搜索会将相关度相加,容易混淆查询相关度结果
方案三:
将索引数据存储在一个不需分词的字段中(keyword), 生成倒排索引时进行三种类型倒排索引的生成,倒排索引生成的时候采用 edge_ngram 对倒排进一步拆分,以满足业务场景需求,检索时不对入参进行分词。
优势:索引数据存储简单,检索索引数据的时只需对一个字段采用 term 全匹配查询规则,查询效率极高。
缺点:采用以空间换时间的策略——比方案二要少,对索引数据来说可以接受。
ES 针对这一业务场景解决方案还有很多种,先列出比较典型的这三种方案,选择方案三来进行处理。
准备工作
- pinyin分词插件安装及参数解读
- ElasticSearch edge_ngram 使用
- ElasticSearch multi-field 使用
- ElasticSearch 多种查询特性熟悉
代码
myself_settings.json:
{
"refresh_interval":"2s",
"number_of_replicas":1,
"number_of_shards":2,
"analysis":{
"filter":{
"autocomplete_filter":{
"type":"edge_ngram",
"min_gram":1,
"max_gram":15
},
"pinyin_first_letter_and_full_pinyin_filter" : {
"type" : "pinyin",
"keep_first_letter" : true,
"keep_full_pinyin" : false,
"keep_joined_full_pinyin": true,
"keep_none_chinese" : false,
"keep_original" : false,
"limit_first_letter_length" : 16,
"lowercase" : true,
"trim_whitespace" : true,
"keep_none_chinese_in_first_letter" : true
},
"full_pinyin_filter" : {
"type" : "pinyin",
"keep_first_letter" : true,
"keep_full_pinyin" : false,
"keep_joined_full_pinyin": true,
"keep_none_chinese" : false,
"keep_original" : true,
"limit_first_letter_length" : 16,
"lowercase" : true,
"trim_whitespace" : true,
"keep_none_chinese_in_first_letter" : true
}
},
"analyzer":{
"full_prefix_analyzer":{
"type":"custom",
"char_filter": [
"html_strip"
],
"tokenizer":"keyword",
"filter":[
"lowercase",
"full_pinyin_filter",
"autocomplete_filter"
]
},
"chinese_analyzer":{
"type":"custom",
"char_filter": [
"html_strip"
],
"tokenizer":"keyword",
"filter":[
"lowercase",
"autocomplete_filter"
]
},
"pinyin_analyzer":{
"type":"custom",
"char_filter": [
"html_strip"
],
"tokenizer":"keyword",
"filter":[
"pinyin_first_letter_and_full_pinyin_filter",
"autocomplete_filter"
]
}
}
}
}
myself_mapping.json
{
"test_type": {
"properties": {
"full_name": {
"type": "text",
"analyzer": "full_prefix_analyzer"
},
"age": {
"type": "integer"
}
}
}
}
工程目录:
测试项目代码:
public class PrefixTest { @Test public void testCreateIndex() throws Exception{ TransportClient client = ESConnect.getInstance().getTransportClient(); //定义索引 BaseIndex.createWithSetting(client,"baidu_index","esjson/baidu_settings.json"); //定义类型及字段详细设计 BaseIndex.createMapping(client,"baidu_index","baidu_type","esjson/baidu_mapping.json"); } @Test public void testBulkInsert() throws Exception{ TransportClient client = ESConnect.getInstance().getTransportClient(); List<Object> list = new ArrayList<>(); list.add(new BulkInsert(12l,"我们都有一个家名字叫中国",12)); list.add(new BulkInsert(13l,"兄弟姐妹都很多景色也不错 ",13)); list.add(new BulkInsert(14l,"家里盘着两条龙是长江与黄河",14)); list.add(new BulkInsert(15l,"还有珠穆朗玛峰儿是最高山坡",15)); list.add(new BulkInsert(16l,"我们都有一个家名字叫中国",16)); list.add(new BulkInsert(17l,"兄弟姐妹都很多景色也不错",17)); list.add(new BulkInsert(18l,"看那一条长城万里在云中穿梭",18)); boolean flag = BulkOperation.batchInsert(client,"baidu_index","baidu_type",list); System.out.println(flag); } }
接下来查看下定义的分词器效果:
http://192.168.20.114:9200/baidu_index/_analyze?text=刘德华AT2016&analyzer=full_prefix_analyzer
得到的结果内容为:
{
"tokens": [
{
"token": "刘",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "刘德",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "刘德华",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "刘德华a",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "刘德华at",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "刘德华at2",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "刘德华at20",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "刘德华at201",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "刘德华at2016",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "l",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "li",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "liu",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "liud",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "liude",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "liudeh",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "liudehu",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "liudehua",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "l",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "ld",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "ldh",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "ldha",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "ldhat",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "ldhat2",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "ldhat20",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "ldhat201",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
},
{
"token": "ldhat2016",
"start_offset": 0,
"end_offset": 9,
"type": "word",
"position": 0
}
]
}
看到以上结果,则表明大功告成了!
以上是关于ElasticStack系列之十三 & 联想补全策略的主要内容,如果未能解决你的问题,请参考以下文章
ElasticStack系列之十六 & ElasticSearch5.x index/create 和 update 源码分析
ElasticStack系列之十七 & 大文本搜索性能提升方案
ElasticStack系列之十五 & query cache 引起性能问题思考
ElasticStack系列之十一 & 同步 mysql 数据的实践与思考
ElasticStack系列之十九 & bulk时 index 和 create 的区别
ElasticStack系列之十四 & ElasticSearch5.x bulk update 中重复 id 性能骤降