Information Technology Reference
In-Depth Information
Chapter 6
Relational Ranking
Abstract
In this chapter, we introduce a novel task for learning to rank, which
does not only consider the properties of each individual document in the ranking
process, but also considers the inter-relationship between documents. According to
different relationships (e.g., similarity, preference, and dissimilarity), the task may
correspond to different real applications (e.g., pseudo relevance feedback, topic dis-
tillation, and search result diversification). Several approaches to solve this new task
are reviewed in this chapter, and future research directions along this line are dis-
cussed.
As shown in previous chapters, in most cases it is assumed that the ranked list is
generated by sorting the documents according to their scores outputted by a scoring
function. That is, the hypothesis
h
can be represented as sort
◦
f(
x
)
, and function
f
works on a single document independent of other documents. However, in some
practical cases, the relationship between documents should be considered, and only
defining the scoring function
f
on single documents is not appropriate. For example,
in the task of topic distillation [
3
], if a page and its parent page (in the sitemap) are
similarly relevant to the query, it is desirable to rank the parent page above the child
page. In the scenario of pseudo relevance feedback, it is assumed that documents
that are similar in their content should be ranked close to each other even if their
relevance features are different. In the scenario of search result diversification, it
is not a good idea to rank documents that are topically very similar all on the top
positions.
In the literature, there are several attempts on introducing inter-relationship be-
tween documents to the ranking process. For ease of reference, we call such kinds of
ranking “relational ranking”. In this chapter, we will first introduce a unified frame-
work for relational ranking [
6
,
7
], and then introduce several pieces of work that
targets specifically at search result diversification [
1
,
4
,
8
,
9
].
