c# - 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

示例

1. query1 = "红狐狸"
2. query2 = "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

解释query1

{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)
}

解释query2

{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

这是由于查询规范。评分算法的这一特性试图使一个查询与下一个查询的分数大致可比。

计算如下:

queryNorm = 1 / sumOfSquaredWeights^½

地点:

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，提升项为分数贡献了更大的相对权重，这是所有提升的真正目的。