Semantic Web ontologies are fast-growing knowledge sources on the Web. Searching relevant concepts from this large repository is a challenging problem. The current Semantic Web search engines provide either (1) coarse-grained search over ontologies or (2) very ne-grained search over individuals. We believe searching and ranking concepts across ontologies provides an ideal granularity for certain tasks such as ontology population and web page annotation. Towards this objective, we propose a novel approach of indexing concepts using ontology axioms in an inverted le structure and ranking them using a dynamic ranking algorithm. Our proposed method is generic and domain-independent. A preliminary evaluation indicates that our proposed method is e ective, outperforming the search function of BioPortal, a large and widely-used ontology repository.
The current breed of semantic web search engines can be broadly grouped into 2
categories: (1) those that search over ontologies, and (2) those that search over
individual resources. The former may be too coarse-grained as a large ontology
may contain hundreds of thousands or even millions of concepts. On the other
hand, the latter approach may be too ne-grained { many resources may be
relevant and returning them individually may not be the best approach. We describe
an approach of retrieving relevant concepts from semantic web ontologies. We
propose a novel technique of indexing concepts using axioms in ontologies. Our
system supports semi-structured queries where names of concepts and relevant
properties can be speci ed.
Semantic search engines such as Sindice [
Our interface provides keyword query input as well as allows to select contextual information around concepts in an ontology corpus. Given a concept in an ontology, all its contextual features are indexed using an inverted le structure. Such features include the concept's label, ID, URI, synonyms, data and object properties used in axioms about the concept, sub classes, super classes, equivalent classes. This approach enriches concept search by disambiguating a concept from those with similar names. For example, if heart concept is searched in context of diseases using our approach, results related to diseases of heart will be ranked higher, while results in other contexts such as functionality will be ranked lower. We now explain the ranking algorithm based on contextual features.
Let , , represent weights of concept label, data properties (i = 1 to m) and object properties (j = 1 to n) of the concept respectively. Let represent weights of context features (k = 1 to t) like synonyms, provenance of the concept. The weights , , and are currently are assigned values based on heuristics. In future we plan to learn these weights using machine learning algorithms. The weight of concept c in the ontology corpus, denoted Wc, is calculated as follows: (1) (2) m n t Wc = :[ + : X i + : X j + : X k] i=1 j=1
k=1 if exact match similarity(x; y)
where x and y represent 2 strings =
(1 4
For evaluation purposes we compare our system1 with the search function on BioPortal,2 a large and widely-used biomedical ontology repository. In our experiment a large portion of ontologies, 252 out of 348 in total, were downloaded from BioPortal and indexed. Together these ontologies contain more than 660,000 classes. 1 Available at http://qassist.cse.iitb.ac.in/LOD/ 2 http://bioportal.bioontology.org/ Algorithm 1: Ranking Algorithm
Data: Query Tokens Q = Qc; Qd1 ; ::Qdm ; Qo1 ; :::Qon ; Qf1 ; ::Qft , Concepts C = C1; C2:::Cn 21 fRoerseua0lct,h: Weleemige0hn,tt Cofi C20o,Cncedpot 0W, cWc 0; 3456 if si mfo(rQecaf;cohlr+abdi=eaitlf1a( Cstpoirimom)p)e(rdQ>toyd0iof;thCdped(nCoi)) > 0 then 7 +
Two metrics widely-used in information retrieval, normalized discounted
cumulative gain (NDCG) and mean average precision (MAP) [
(a) without property infomation (b) without property infomation (c) with property infomation Semantic Web search is primarily divided into two types - one which allows keyword query capability and other which needs SPARQL query input. The latter gives exact results due to precise input queries. This requires user to have technical knowledge about writing a SPARQL query. We present an approach of searching for concepts using semistructured keyword queries that incorporates contextual features to improve precision. A preliminary evaluation and a comparison with BioPortal's search function shows the e ectiveness of our system. In future we will investigate the incorporation of ontology reasoning to include implicit contextual features. Currently the ranking algorithm derives feature weights heuristically. Going ahead we will learn the weights using machine learning methods. In addition to enriched concept search, our further work will also include property search across ontologies.