Jianhan Zhu, Dawei Song, Marc Eisenstadt and Cristi Barladeanu
Knowledge Media Institute, The Open University, U.K.
Keywords: Meta search engines, collection fusion, metadata.
Abstract: We present a novel meta-search engine called DYNIQX for metadata based cross search in order to study
the effect of metadata in collection fusion. DYNIQX exploits the availability of metadata in academic
search services such as PubMed and Google Scholar etc for fusing search results from heterogeneous search
engines. Furthermore, metadata from these search engines are used for generating dynamic query controls
such as sliders and tick boxes etc for users to filter search results.
In the light of large scale powerful search engines
such as Google, which have achieved tremendous
success in recent years, thanks to their effective use
of the PageRank algorithm (Brin and Page 1998),
smart indexing, and efficiency in searching terabytes
of data (Ghemawat et al. 2003), how can traditional
professional, academic and library repositories
survive and keep their successes within their specific
domain? Even given the success of these big search
engines, in fact it is still very difficult for them to
work effectively with repositories that belong to
specific professional or proprietary domains. We
think there are two main reasons for this.
Firstly, due to legal or proprietary constraints,
sometimes search engines cannot get hold of full
content of information and may provide only the
link to the place where the information can
ultimately be found.
Secondly, big search engines, which tap into
heterogeneous resources, sometimes cannot perform
as well as some domain or context specific search
services (for example, in the context of arranging air
travel between London and New York, the British
Airways website will provide much better search
services than Google). We think that the key for
successful domain specific specialized search
services is to fully utilize the domain context and
metadata which describes the domain context.
A limitation of current niche search services is
the wide existence of information islands, which
results in a contextual “jump” for users when they
are searching different repositories (Awre et al.
2005). It is important to give users a unified search
interface, which has been successfully used by big
search engines.
We treat the problem of building a meta search
engine on top of a number of search engines as a
collection fusion problem as defined by Voorhees et
al. (1994; 1995). The research questions we would
like to answer are: How to generate a single ranked
search result list based on a number of ranked lists
from search engines? How to take into account
relevance of each result to the query and the original
rankings of the search results in the integrated
ranked list? How to integrate metadata in ranking?
After reviewing existing work, we found the
necessity for a meta-search system that can
seamlessly integrate multiple search engines of
different natures. To tackle this, we propose a novel
dynamic query meta-search system called DYNIQX
that integrates search results from multiple sources
by taking into account search results’ relevance to
the query, original rankings, and metadata, in
collection fusion, and provides a unified search
interface. DYNIQX also provides plug-in interfaces
for new search engines. DYNIQX can help facilitate
our investigation of current cross-search and
metadata-based search services, identification of
resources suitable for cross-search or metadata-
based search, and comparison of single source
search, cross-search, and metadata-based search.
Currently many niche search engines have adopted
what we call a linear/top-down/hierarchical
approach. For example, in the Intute search
Zhu J., Song D., Eisenstadt M. and Barladeanu C. (2008).
In Proceedings of the Third International Conference on Software and Data Technologies - ISDM/ABF, pages 312-315
DOI: 10.5220/0001896403120315
(http://www.intute.ac.uk), a popular search engine
among students for finding high quality educational
websites, a searcher may select from a list of subject
areas and/or resource types for his/her search, and
he/she is then taken to the result page. We think the
rigidity of this approach may limit the user to search
within the classification of the resources.
Additionally, there are many forms of metadata
which have not been fully exploited during the
search process.
To overcome the above limitations, we propose
to experiment with the dynamic query approach
based on Shneiderman’s philosophy (Shneiderman
1994) of letting users experiment in real time to tune
search results. Dynamic queries help users search
and explore large amounts of information by
presenting them with an overview of the datasets,
and then allow them quickly to filter out unwanted
information. “Users fly through information spaces
by incrementally adjusting a query (with sliders,
buttons, and other filters) while continuously
viewing the changing results.” A popular example of
this approach is that of Kayak.co.uk, a meta-search
engine which searches over 100 travel sites.
In DYNIQX, search results from a number of
search engines are fused into a single list by both the
relevance of each result to the search query based on
our indexing of top results returned from these
search engine, and the rankings of the result
provided by one or more search engines as below:
(|) (1 )(|) /(log( ()1)
fuse average
pqd pqd Rank d
∝− + +
where q is the query, p
(q|d) is the fused
conditional probability of document d used to rank it
in the final list, p(q|d) is the conditional probability
of d based on our index, λ is a parameter adjusting
the effect of the two components in the final
probability, and Rank
(d) is the average ranking
of document d given by search engines. In the
equation we take the log of the average ranking in
order to transform the linear distribution of the
rankings of d for integrating with the document
conditional probability.
DYNIQX provides a novel way of meta-
searching a number of search engines in terms that
high quality search results from a number of search
engines are integrated, metadata from heterogeneous
sources are unified for filtering and searching these
high quality search results, high quality results based
on a number of queries covering a topic are all
integrated in DYNIQX, and features such as
metadata-driven controls and term clouds are used
for facilitating search.
The architecture of our DYNIQX system is
shown in Figure 1. In Figure 1, first, a user sends a
query to the DYNIQX system. The query is
processed and translated into the appropriate form
for each search service, e.g., PubMed. For each
query, each search engine, e.g., Intute, PubMed, or
Google Scholar, returns a ranked list of search
results. Results from all these ranked lists are
indexed and searched by Lucene (Hatcher and
Gospodnetic 2004).
Unlike typical search engines where the user can
only specify one query at a time, in DYNIQX, the
user can specify a number of queries describing
different aspects of a search topic, e.g., “bird flu”,
“avian influenza”, and “H5N1” etc for finding
documents on “bird flu”. Each query is translated
into the appropriate form for each search service,
e.g., PubMed. For each query, each search engine
returns a ranked list of results. Results are ranked by
their overall relevance scores to the topic in a single
ranked list. For each result, its overall relevance
score integrates the relevance of the result to the
queries based on the Lucene index, rankings and
relevance scores of the result by each search engine,
and metadata associated with the result. Metadata
from these heterogeneous sources are unified for
filtering results. This is illustrated in the DYNIQX
search interface shown in Figure 2.
In Figure 2, in Section A, a user adds a number
of search queries shown in Section B. Statistics of
search results from different search engines are
shown in a table in Section B. The user can
select/deselect search engines in Section E for meta-
search. Once search results are retrieved from search
engines, the user can view a single ranked list in
Section G. When more results arrive, the user clicks
a refresh button in Section A to refresh the single
ranked list. Based on the significance of terms
measured by their document frequencies, a term
cloud is displayed in Section F for filtering result. In
Section D, the user can exclude/include queries in
the meta-search. Metadata associated with search
results are used for re-ranking search results in
Section C.
In this paper, we propose a novel metadata based
search engine called DYNIQX which fuses
information from data collections of heterogeneous
nature. Metadata from multiple sources are
integrated for generating dynamic controls in the
forms of sliders and tick boxes etc for the users to
further filter and rank search results. Since the effect
of metadata in IR has not been sufficiently studied
Figure 1: Architecture of DYNIQX.
Figure 2: DYNIQX dynamic query interface.
ICSOFT 2008 - International Conference on Software and Data Technologies
previously, our work provides insights into how to
integrate metadata with mostly content based
information retrieval systems. Our preliminary user
evaluation reported in (Zhu et al. 2008) shows that
DYNIQX can help users to complete real world
information search tasks more effectively and
efficiently than individual search engines. In the
future, we will carry out more formal user evaluation
of DYNIQX and study the effect of different ranking
algorithms in collection fusion in DYNIQX.
The work reported in this paper is funded in part by
the JISC (Joint Information Systems Committee)
funded DYNIQX (Metadata-based DYNamIc Query
Interface for Cross(X)-searching content resources)
Awre, C. et al. (2005) The CREE Project: investigating
user requirements for searching within institutional
environments. D-Lib Magazine, October 2005, 11(10).
Brin, S., and Page, L. (1998) The Anatomy of a Large-
Scale Hypertextual Web Search Engine. Computer
Networks 30(1-7): 107-117
Ghemawat, S. et al. (2003) The Google File System. In
ACM Symp. on Operating Systems Principles.
Hatcher, E., and Gospodnetic, O. (2004) Lucene in Action.
Manning Publications Co, ISBN: 1932394281.
Shneiderman, B. (1994) Dynamic Queries for Visual
Information Seeking. IEEE Software 11(6): 70-77.
Voorhees, E.M. et al. (1994) The Collection Fusion
Problem. In Text REtrieval Conference (TREC).
Voorhees, E.M. et al. (1995) Learning Collection Fusion
Strategies. In SIGIR: 172-179.
Zhu, J., Song, D., Eisenstadt, M., Barladeanu, C., and
Rüger, S. (2008) DYNIQX: A novel meta-search
engine for metadata based cross search. In First IEEE
International Conference on the Applications of
Digital Information and Web Technologies (ICADIWT
2008), VSB- Technical University of Ostrava, Czech