an ontology, in our solution, enables similarity checks when there are no direct overlaps between user profiles and, therefore, result in more accurate similarity measurements. blem of automated routing of conference papers to their rev somewhat related problem to that of expert finding. Most of the current approaches to that problem use a group of papers authored by reviewers to determine their user profile and perform routine content matching(similar to personalization)to determine whether a paper is fit to be reviewed by that user [12]. The expert finding task introduced by TREC 2005 [13] requires one to provide a ranked list of the candidate experts based on the web data provided. Our attem handle the problem of choosing the best expert description of a hypothetical expert(set of topics with weights) profiles of candidate experts Use of ontologies to derive new concepts that are likely to be of interest to the user through semantic spreading activation networks has been studied as well [14-17, 5]. Pre- ious studies have shown that the spreading process improves accuracy and overco the challenges caused by inherent relationships and Polysemy in word sense disam- biguation process [15, 16] and ontology mapping [17]. We use this spreading process to facilitate the semantic similarity computation. We build on the spreading process used in [5] to learn user preferences in order to drive a personalized multimedia search. The learning process utilizes ontologies as a means to comprehend user interests (in BOw format)and establishes the need to consider related concepts to improve search quality While the results in [5] suggest that personalized search is of better quality in comparison to normal search, they do not show whether the consideration of related terms contributes to these improvements. On the other hand, we show that our spreading process indeed improves the accuracy of our new similarity measures and in the particular context of user profile matching A number of approaches have already been proposed to determine the similarity between two ontology concepts(or words). These determine similarity by: measurin the path distance between them [18], evaluating shared information between them [19] recursively matching sub-graphs[20), combining information from various sources [21]. are only able to determine closeness between two concepts(or words), we compute similarity between two weighted sets of concepts(or words). One of our algorithms us the simple path measure described in [18]over a bipartite graph to determine such a set intersection Ve now compare with other works that use San based IR techniques. One of our imilarity measures is similar to the one discussed in[25] but differs in the treatment of the results of the activation process. While the previous work utilizes the results of the activation to rank documents with respect to a query, our work maps an aggregate of the activation results to a similarity value. Knowledge from an ontology is used to extend the bow with terms that share important relationships with original terms to improve document retrieval is presented in [4]. Our work on set spreading is somewhat similar to this but we further explore the notion of computing similarity by optimal concept matching in bipartite graphs and using SAN 3 Background In this section, we formally define and explain some terms used in the rest of the documentan ontology,in our solution, enables similarity checks when there are no direct overlaps between user profiles and, therefore, result in more accurate similarity measurements. The problem of automated routing of conference papers to their reviewers is a somewhat related problem to that of expert finding. Most of the current approaches to that problem use a group of papers authored by reviewers to determine their user profile and perform routine content matching (similar to personalization) to determine whether a paper is fit to be reviewed by that user [12]. The expert finding task introduced by TREC 2005 [13] requires one to provide a ranked list of the candidate experts based on the web data provided. Our attempt is to handle the problem of choosing the best expert given a description of a hypothetical expert (set of topics with weights) and a set of user profiles of candidate experts. Use of ontologies to derive new concepts that are likely to be of interest to the user through semantic spreading activation networks has been studied as well [14–17, 5]. Pre￾vious studies have shown that the spreading process improves accuracy and overcomes the challenges caused by inherent relationships and Polysemy in word sense disam￾biguation process [15, 16] and ontology mapping [17]. We use this spreading process to facilitate the semantic similarity computation. We build on the spreading process used in [5] to learn user preferences in order to drive a personalized multimedia search. The learning process utilizes ontologies as a means to comprehend user interests (in BOW format) and establishes the need to consider related concepts to improve search quality. While the results in [5] suggest that personalized search is of better quality in comparison to normal search, they do not show whether the consideration of related terms contributes to these improvements. On the other hand, we show that our spreading process indeed improves the accuracy of our new similarity measures and in the particular context of user profile matching. A number of approaches have already been proposed to determine the similarity between two ontology concepts (or words). These determine similarity by: measuring the path distance between them [18], evaluating shared information between them [19], recursively matching sub-graphs [20], combining information from various sources [21], analysing structure of the ontology [22], and combining content analysis and web search [23]. A few other measures are evaluated in [24]. While all these approaches are only able to determine closeness between two concepts (or words), we compute similarity between two weighted sets of concepts (or words). One of our algorithms use the simple path measure described in [18] over a bipartite graph to determine such a set intersection. We now compare with other works that use SAN based IR techniques. One of our similarity measures is similar to the one discussed in [25] but differs in the treatment of the results of the activation process. While the previous work utilizes the results of the activation to rank documents with respect to a query, our work maps an aggregate of the activation results to a similarity value. Knowledge from an ontology is used to extend the BOW with terms that share important relationships with original terms to improve document retrieval is presented in [4]. Our work on set spreading is somewhat similar to this but we further explore the notion of computing similarity by optimal concept matching in bipartite graphs and using SAN. 3 Background In this section, we formally define and explain some terms used in the rest of the document