in the ontology. For example, Wordnet based spreading can be tuned to employ this termination condition when path from individual terms to the root suffices to terminate the spreading In less rigorous ontologies such as the Wikipedia category graph may not be able to support this condition as there may not be a single root. In such the spreading process is terminated if there exists at least one path from every node that the complete details of the two spreading algorithms in our technical report (4?cribe belongs to the smallest of the two profiles to the nodes in the other profile. We describe 5 Similarity Computation In this section, we describe the complete details of our variant metrics to compute semantic similarity using ontologies 5.1 Set-based measure Set spreading process enriches the profiles by appending the related terms in to capture all the relationships between the terms. For set spreading, the same similarity technique defined in Equation 1 is applicable to compute similarity b the extended BOws or BOCs Set spreading-based similarity computation begins by measuring similarity of the original profiles, and proceeds by incrementally extending he profiles until termination while computing the similarity between profiles at every Iteration 5.2 SAN-based measure This similarity computation metric is inspired by the abundant work that exists in the area of semantic search especially by techniques that process a San (e. g, [25, 15). We focus on similarity computation techniques that use a SAn resulting from graph spreading process(see figure 2a for an overview of SAN structure). Following the construction of the semantic network the similarity values are computed either by reducing the graph to a bipartite graph or by activating the graph with an activation strategy. We have implemented both these techniques for evaluation. A brief introduction to the activation process is presented below. For a more detailed discussion the reader is pointed to [15] (a)SAN Building Process (b) Bipartite Graph Matching(Hungarian Algorithm) Fig 2: SAN-based Similarity Computationsin the ontology. For example, Wordnet based spreading can be tuned to employ this termination condition when path from individual terms to the root suffices to terminate the spreading. In less rigorous ontologies such as the Wikipedia category graph may not be able to support this condition as there may not be a single root. In such a case, the spreading process is terminated if there exists at least one path from every node that belongs to the smallest of the two profiles to the nodes in the other profile. We describe the complete details of the two spreading algorithms in our technical report [26]. 5 Similarity Computation In this section, we describe the complete details of our variant metrics to compute semantic similarity using ontologies. 5.1 Set-based Measure Set spreading process enriches the profiles by appending the related terms in order to capture all the relationships between the terms. For set spreading, the same cosine similarity technique defined in Equation 1 is applicable to compute similarity between the extended BOWs or BOCs. Set spreading-based similarity computation begins by measuring similarity of the original profiles, and proceeds by incrementally extending the profiles until termination while computing the similarity between profiles at every iteration. 5.2 SAN-based measure This similarity computation metric is inspired by the abundant work that exists in the area of semantic search especially by techniques that process a SAN (e.g., [25, 15]). We focus on similarity computation techniques that use a SAN resulting from graph spreading process (see figure 2a for an overview of SAN structure). Following the construction of the semantic network the similarity values are computed either by reducing the graph to a bipartite graph or by activating the graph with an activation strategy. We have implemented both these techniques for evaluation. A brief introduction to the activation process is presented below. For a more detailed discussion the reader is pointed to [15]. First Profile Second Profile . . . . . . . . . -- Profile Terms -- Terms related by r1 Relation (r1) -- Terms related by r2 Relation (r2) (a) SAN Building Process Optimal Bipartite Matching 1 2 3 4 5 6 7 1 4 5 7 E = {1,2,3,4,5,6,7} M = {1,4,5,7} (b) Bipartite Graph Matching (Hungarian Algorithm) Fig. 2: SAN-based Similarity Computations