为什么 Lucene 提升查询的分数低于相同的普通查询？

Question

我在 lucene 中测试 boost 运算符并发现了奇怪的行为

例子

查询 1 = "red fox"
查询2 = "red^1.2 fox"

当我针对文本测试查询时：

"wonderful red fox"

查询 2 的分数低于查询 1。但我希望 query2 应该获胜。

以下查询说明

查询1的解释

{0,4339554 = (MATCH) sum of:
  0,2169777 = (MATCH) weight(content:fox in 0), product of:
    0,7071068 = queryWeight(content:fox), product of:
      0,3068528 = idf(docFreq=1, maxDocs=1)
      2,304384 = queryNorm
    0,3068528 = (MATCH) fieldWeight(content:fox in 0), product of:
      1 = tf(termFreq(content:fox)=1)
      0,3068528 = idf(docFreq=1, maxDocs=1)
      1 = fieldNorm(field=content, doc=0)
  0,2169777 = (MATCH) weight(content:red in 0), product of:
    0,7071068 = queryWeight(content:red), product of:
      0,3068528 = idf(docFreq=1, maxDocs=1)
      2,304384 = queryNorm
    0,3068528 = (MATCH) fieldWeight(content:red in 0), product of:
      1 = tf(termFreq(content:red)=1)
      0,3068528 = idf(docFreq=1, maxDocs=1)
      1 = fieldNorm(field=content, doc=0)
}

查询2的解释

{0,4313012 = (MATCH) sum of:
  0,2396118 = (MATCH) weight(content:fox^1.25 in 0), product of:
    0,7808688 = queryWeight(content:fox^1.25), product of:
      1,25 = boost
      0,3068528 = idf(docFreq=1, maxDocs=1)
      2,035813 = queryNorm
    0,3068528 = (MATCH) fieldWeight(content:fox in 0), product of:
      1 = tf(termFreq(content:fox)=1)
      0,3068528 = idf(docFreq=1, maxDocs=1)
      1 = fieldNorm(field=content, doc=0)
  0,1916894 = (MATCH) weight(content:red in 0), product of:
    0,6246951 = queryWeight(content:red), product of:
      0,3068528 = idf(docFreq=1, maxDocs=1)
      2,035813 = queryNorm
    0,3068528 = (MATCH) fieldWeight(content:red in 0), product of:
      1 = tf(termFreq(content:red)=1)
      0,3068528 = idf(docFreq=1, maxDocs=1)
      1 = fieldNorm(field=content, doc=0)
}

我想知道为什么提升查询的分数低于普通查询？

Answer 1

这是查询规范的问题。评分算法的这一功能试图使一个查询与下一个查询的分数大致具有可比性。

计算如下：

queryNorm = 1 / sumOfSquaredWeights^½

其中：

sumOfSquaredWeights = query boost² · ∑ ( idf · term boost )²

如果您从解释中删除该因素，只需将最终分数除以查询范数，您就会发现第二个查询确实获得了更高的分数：

query1 --> .4339554 / 2.304384 = 0.1883
query2 --> .4313012 / 2.035813 = 0.2119

但更重要的一点是：您不应该过多地阅读比较一个查询与下一个查询的分数。分数仅与生成它们的查询真正相关。您可以在解释中看到提升项对分数贡献了更大的相对权重，这是所有提升的真正目的。

为什么 Lucene 提升查询的分数低于相同的普通查询？

Why lucene boosted query has lower score than same normal query?

c#

lucene

lucene.net