Conference,
Context-Sensitive Ranking Using Cross-Domain Knowledge for Chemical Digital Libraries
(09/2013 2013)
Abstract
Today, entity-centric searches are common tasks for information
gathering. But, due to the huge amount of available information the entity itself
is often not sufficient for finding suitable results. Users are usually searching
for entities in a specific search context which is important for their relevance
assessment. Therefore, for digital library providers it is inevitable to also consider
this search context to allow for high quality retrieval. In this paper we present
an approach enabling context searches for chemical entities. Chemical entities
play a major role in many specific domains, ranging from biomedical over
biology to material science. Since most of the domain specific documents lack
of suitable context annotations, we present a similarity measure using crossdomain
knowledge gathered from Wikipedia. We show that structure-based
similarity measures are not suitable for chemical context searches and introduce
a similarity measure combining entity- and context similarity. Our experiments
show that our measure outperforms structure-based similarity measures for
chemical entities. We compare against two baseline approaches: a Boolean retrieval
model and a model using statistical query expansion for the context term.
We compared the measures computing mean average precision (MAP) using a
set of queries and manual relevance assessments from domain experts. We were
able to get a total increase of the MAP of 30\% (from 31\% to 61\%). Furthermore,
we show a personalized retrieval system which leads to another increase
of around 10\%.
