5. QUERY REWRITING
作用:
- query rewriting is the task of altering a given query so that it will get better results and, more importantly, to help solve the recall problem.
- can treat it as a machine translation problem: language of user queries(S) <=> language of web documents(T)
5.1 Methodology
两个阶段:
- learning phase: learns phrase-level translations from queries to documents;
- decoding phase: generates candidates for a given query;
Learning Phase =>
此阶段存在的困难:获取大量query - 可以提高相关度的rewritten query训练数据;
困难原因:1)好的翻译模型需要超大量的双语文本;2)编辑不能很好的选择什么样的query可以提高相关性;
解决方案:
- 使用click graphs(加权无向图:queries和doc是nodes,edges代表queries和document的点击,权重是点击数)
- 使用文章title作为对应的rewritten query(因为相对于文章body,文章title与query更加相似)
- 根据得到的query-title配对,we follow the common steps for a typical phrase-based matching translation framework to learn phrase-level translations;
Decoding Phase =>
作用:
每个query(q)都有很多分词的方法得到多个phrase,而且每个phrase都有很多translation,这导致将出现成百上千的候选rewritten_query;
=》decoding phase将在这些候选词中挑出最可靠的rewritten_query(qw);
公式:(待添加)
hi(qc,q)代表第i个feature function;λi指定该function的权重,λi可以被人工指定或者通过loss function学习得到;
特征函数:
对于每对(qc,q),本论文包含3种类型的feature function:Query feature functions, Rewrite query feature functions, Pair feature functions;
(Query feature functions)
h1 - number of words in q;h2 - number of stop words in q;h3 - language model score of the query q;h4 - query frequency of q;h5 - average length of words in q;
(Rewrite query feature functions)
h6 - number of words in qc;h7 - number of stop words in qc;h8 - language model score of the query qc;h9 - query frequency of qc;h10 - average length of words in qc;
(Pair feature functions)
h11 - Jaccard similarity of URLs shared by q and qc in the query-URL graph;
h12 - difference between the frequencies of q and qc;
h13 - word-level cosine“余弦” similarity between q and qc;
h14 - difference between the number of words between q and qc;
h15 - number of common words in q and qc;
h16 - difference of language model scores between q and qc;
h17 - difference of the number of stop words between q and qc;
h18 - difference of the average length of words between q and qc;
=》经实验,发现h11, h12, h13是最重要的三个feature functions;
5.2 Ranking Strategy
根据original query和rewritten query,有两种排序策略:
Replace the original query with the rewritten query (未采用)=>
评估:直接采用replace的方式很冒险,一些低质量的rewrites会对相关度造成负面影响;
Blending mode(采用) =>
方法:
1)分别使用original query(q)和rewritten query(qc)从搜索引擎中获取top-N个文档,并记录下两次获得的文档的序列和分值(O, R);
2)从O和R中取交集:若文档D同时出现在O和R中,D的最终分数未max(O, R);
3)在此基础上根据各文档的分值进行排序,选择top-N作为original query搜索的最终结果;
两种排序策略的评估:
两种方法都能对tail query的搜索相关度进行显著的提高;
但是由于rewritten query可能改变original query的目的,Replace策略的结果不如Blending Mode的好;