elasticsearch同义词和搜索相关度优化

虽然docker方便.但是由于163镜像版本太低.所以在本地搭了一个es.版本6.3.0;

首先说下这次研究的方向: 搜索框搜西红柿,那么番茄是出不来的. 搜索米的时候.虾米条比大米的相关度还高
暂时考虑的方案如下:
1.加入同义词
2.相关度优化.将字段排优先级,名字的相关度提升
3.未雨绸缪,增加类似某宝某东一样的,销量高的排名也提升一点,具体如下

同义词插件

插件1 :同义词文件配置方式
插件2 :数据库配置方式
原本想使用插件2. 两种方式都是可以动态加载同义词词库.所以配置好之后不需要修改词库就重启es,
但是插件2的大佬一直没有更新,所以我在6.3.0中加入插件报错我没有处理好暂时先使用插件1

插件2
项目支持6.x版本.但是tag上没有,因此clone到本地.修改版本.然后maven打包.得到插件放到es的plugins下重启es
github上有使用的实例.我的测试用例,这个成功之后,暂时先放一边.先搞下相关度优化

DELETE prod
PUT /prod
{
    "index" : {
        "analysis" : {
            "analyzer" : {
                "synonym" : {
                    "tokenizer" : "ik_max_word",
                    "filter" : ["remote_synonym"]
               }
            },
            "filter" : {
                "remote_synonym" : {
                    "type" : "dynamic_synonym",
                    "synonyms_path" : "synonym.txt",    //文件没有就创建一个
                    "interval": 30          //由于本地跑.所以30s重新查一次.线上肯定不可以30s
                },
                "local_synonym" : {
                    "type" : "dynamic_synonym",
                    "synonyms_path" : "synonym.txt"
                }
            }
        }
    }
}

将项目中需要优化的几个字段抽取出来自己定义了一个精简版的demo

PUT /prod/demo/_mapping
{
 
      "properties":{
        "name":{
          "type":"text",
           "analyzer": "synonym"
        },
        "description":{
           "type":"text",
           "analyzer": "ik_max_word"
        },
        "brandName":{
           "type":"text",
           "analyzer": "ik_max_word"
        },
        "labelName": {
           "type": "text",
           "analyzer": "ik_max_word"
        },
         "menuCategoryNamePath": {
           "type": "text",
           "analyzer": "ik_max_word"
        },
        "num":{
          "type": "integer"
        }
  }
}

原先的查询代码:

QueryBuilders.multiMatchQuery(((SearchProductReq) req).getSearchContent(),
                    PRODUCT_NAME,
                    BRAND_NAME,
                    DESCRIPTION,
                    LABEL_NAME,
                    MENU_CATEGORY_NAME_PATH);

等同于:

GET prod/demo/_search
{
  "query":{
    "multi_match": {
      "query": "米",
      "fields": ["name","description","brandName","labelName","menuCategoryNamePath"]
    }
  }
}

修改后,排分还待修正.暂时按这样的分数:

   QueryBuilder  queryBuilder = QueryBuilders.boolQuery()
                    .should(QueryBuilders.matchQuery("name", "米").boost(0.8f))
                    .should(QueryBuilders.matchQuery("brandName", "米").boost(0.6f))
                    .should(QueryBuilders.matchQuery("labelName", "米").boost(0.6f))
                    .should(QueryBuilders.matchQuery("menuCategoryNamePath", "米").boost(0.2f))
                    .should(QueryBuilders.matchQuery("description", "米").boost(0.4f));


        FieldValueFactorFunctionBuilder factorFunctionBuilder = new FieldValueFactorFunctionBuilder("num");
        factorFunctionBuilder.factor(0.1f);
        factorFunctionBuilder.modifier(FieldValueFactorFunction.Modifier.LOG1P);
        FunctionScoreQueryBuilder boostMode = QueryBuilders
                .functionScoreQuery(queryBuilder, factorFunctionBuilder)
                .boostMode(CombineFunction.SUM);
        SearchRequestBuilder requestBuilder = client.prepareSearch(ESConstant.PRODUCT_INDEX)
                .setTypes(ESConstant.PRODUCT_TYPE);
        requestBuilder.setQuery(boostMode);

等同于:

GET prod/demo/_search
{
  "query":{
    "function_score": {
      "query": {
          "bool": {
      "should": [
        {
          "match": {
            "name": {"query": "大米","boost":0.8}
          }
        },
        {
          "match": {
            "brandName": {"query": "大米","boost":0.6}
          }
        },
        {
          "match": {
            "labelName": {"query": "大米","boost":0.6}
          }
        },
        {
          "match": {
            "description": {"query": "大米","boost":0.5}
          },
        {
          "match": {
            "menuCategoryNamePath": {"query": "大米","boost":0.2}
          }
        }
      ]
    }
      },
      "field_value_factor": {
        "field": "num",
        "modifier": "log1p",
        "factor": 0.1
      },
      "boost_mode": "sum"
    }
  }
}

提升相关度使用function_score参考官方文档

加入同义词,这个需要修改mapping.

PUT /prod/demo/_mapping
{
 
      "properties":{
        "name":{
          "type":"text",
           "analyzer": "synonym"
        },
        "description":{
           "type":"text",
           "analyzer": "synonym"
        },
        "brandName":{
           "type":"text",
           "analyzer": "synonym"
        },
        "labelName": {
           "type": "text",
           "analyzer": "synonym"
        },
         "menuCategoryNamePath": {
           "type": "text",
           "analyzer": "synonym"
        },
        "num":{
          "type": "integer"
        }
  }
}

这时候在config的synonym.txt文件中增加同义词.比如比如插入数据:


POST /prod/demo/2
{
   "name":"大米",
    "description":"稻香大米",
    "brandName":"COCO",
    "labelName":"大米",
   "menuCategoryNamePath":"食品|饮料",
    "num":3
}
POST /prod/demo/3
{
  "name":"虾米条",
   "description":"虾米条",
    "brandName":"COCO",
    "labelName":"虾米条",
    "menuCategoryNamePath":"食品|零食",
     "num":1
}
POST /prod/demo/4
{
   "name":"惠宜 珍珠米 10KG",
   "description":"惠宜 珍珠米 10KG",
    "brandName":"COCO",
    "labelName":"惠宜",
    "menuCategoryNamePath":"食品|饮料",
     "num":5
}
POST /prod/demo/5
{
   "name":"口口牌 泰国进口 泰国茉莉香米",
   "description":"口口牌 泰国进口 泰国茉莉香米",
    "brandName":"COCO",
    "labelName":"口口牌",
    "menuCategoryNamePath":"食品|饮料",
     "num":3
}

ik分词器不会拆大米.所以大米和米是两个条件. 如果把米/大米设置成近义词.name这时候.搜米和大米都能收到上述商品.并且相关度也有有一定变化.

首先没有淘宝京东那么智能,但是由于我刚接触es.暂时先这样优化下,
测试数据.很假,主要测试下效果.