如何在Elasticsearch中实现Term Centric+短语检索？

Elasticsearch的「multi-match + cross_fields」模式实现的是Term Centric检索，这个特性很实用，因为不需要per-field-IDF。但我想以Term Centric的方式执行短语检索，发现cross_fields和combined_fields都不支持短语检索，有没有其他实现方式？

我试过「multi-match + phrase」模式，但它采用的是per-field-IDF，不符合需求。

详细示例 #

1. 创建索引 #

PUT /testindex
{
    "settings": {
        "number_of_shards": 1,
        "number_of_replicas": 0
    },
    "mappings": {
      "properties": {
        "fn":    { "type": "text" },  
        "ln":  { "type": "text"  }
      }
    }
}

2. 添加10条相同文档 #

POST /testindex/_doc/{from 1 to 10}
{
    "fn": "hello world",
    "ln": "bye"
}

3. 添加另一条文档，将「hello world」放入ln字段 #

POST /testindex/_doc/100
{
    "fn": "bye",
    "ln": "hello world"
}

4. 执行短语检索 #

我希望对「hello world」执行短语检索，使用如下查询：

POST /testindex/_search
{
    "query": {
        "bool": {
            "must": {
                "multi_match": {
                    "query": "hello world",
                    "fields": ["fn", "ln"],
                    "type": "phrase"
                }
            }
        }
    }
}

但「multi-match+phrase」是字段中心式的，「hello world」在ln字段中的IDF值极高，导致ID为100的文档得分远高于其他文档，这不是我想要的结果：

{
    "took": 5,
    "timed_out": false,
    "_shards": {
        "total": 1,
        "successful": 1,
        "skipped": 0,
        "failed": 0
    },
    "hits": {
        "total": {
            "value": 11,
            "relation": "eq"
        },
        "max_score": 3.1015396,
        "hits": [
            {
                "_index": "testindex",
                "_type": "_doc",
                "_id": "100",
                "_score": 3.1015396,
                "_source": {
                    "fn": "bye",
                    "ln": "hello world"
                }
            },
            {
                "_index": "testindex",
                "_type": "_doc",
                "_id": "1",
                "_score": 0.26195967,
                "_source": {
                    "fn": "hello world",
                    "ln": "bye"
                }
            },
            ...
        ]
    }
}

内容的提问来源于stack exchange，提问作者Michael Liu