Semantic relevance among information resources can play an effective role in information retrieval, and there are several different approaches to measure semantic similarities. However, they are usually limited measuring similarities among structured concepts, and some also considers the similarities between individual resources but they are limited to to textual documents. This paper aims to extend the similarity to relevance so that even the relations between completely different resources can also be considered, and explore the possibility of measuring such semantic relevance not limited to a specific structure. The method presented in this paper is a novel measure for semantic relevance among resources that are organised within an ontological structure.
There have long been various researches to enhance information retrieval and knowledge sharing, and ontology-based semantic technology opened a new possibility of improvement. Upon retrieving information, usually a list of results is provided based on the given query. However, the search result is generally concentrating on the relevance between the query and each resource, and to find other resources that are similar or relevant to a certain resource, users need to go back to the previous results list or initiate another query. Since each resource representing different knowledge is semantically connected to each other, measuring and utilising the semantic relevance between those resources can provide more flexible range of results which are not strictly limited to the given query. Such measure can also be used to provide a guideline or recommendation on what other resources are related and needed to be considered when the user is examining a certain resource.
In this paper, we propose a methodology to measure semantic relevance between resources, based on ontological representation. First, generic aspects of semantic similarity measure are briefly discussed with some existing methodologies. Then, an ontology-based methodology for measuring semantic relevance between resources is proposed and described in detail, and it is concluded with a discussion.
Semantic similarity is the likeness of the semantic content among documents or
the meaning among a set of structured terms, and there are several methods
to discover the semantic similarities among concepts or resources. One of the
oldest but still effective approaches is text-based technique, including
vectorbased model such as latent semantic analysis[1]. Also, there are several researches
using direct relations between objects, based on either tree structure[3] or more
generic and complex graph structure[2]. Their approach is typically (
The method for semantic relevance measure proposed in this paper aims to calculate the relevance between resources as numeric values based on the semantic relations between concepts. 3.1
To measure semantic relevance between resources, all resources should be semantically structured and represented. An ontology would be the ideal representation for this purpose. Such ontology can be constructed in two different ways, depending on the purpose of the system and the type of resources being managed. Firstly, the ontology can be designed such that some classes contain resources as their instances whereas other classes represent other semantic concepts. For example, an ontology can have a class called Documents containing all the resources as documents form, and another class called Person containing personal profiles which are not regarded as resources hence not the target objects for searching but used as additional semantic information to produce, for instance, personalised search results. Secondly, all instances in all classes can represent resources. This is appropriate when everything described within the ontology is regarded as a resource. In this case, for above example, personal profiles are also regarded as resources hence they are also the target objects for searching.
Once the fundamental structure of the ontology - the concepts, attributes and properties - is defined, a set of rules should be defined to specify additional semantic information. The rules can be defined in the usual way, but for our methodology, some additional rules may be required to concentrate on representing the meaningful semantic relations between resources. As a simple example, suppose we have an ontology with a class Resource containing all the resources as its instances and another class Topic containing the topic hierarchy of topics as its instances(Fig. 1). Here, one may define a rule between a resource R1 which is about Topic T1 and another resource R2 which is about Topic T2, a subtopic of T1(Fig. 2). This example rule can be defined in F-Logic[7] as following: R2[isSubT opicResourceOf
R1] ← R1 : Resource ∧ R2 : Resource ∧ T1 : T opic ∧ T2 : T opic ∧ R1[isAbout
T2]
The fundamental classes(concepts), properties(attributes and relations) and rules should be defined by domain experts, and the resources(instances) can be created by either a limited number of experts or the users themselves depending on the purpose and functionalities of the system. 3.2
Once the ontology is defined, the next step is assigning a numeric value for each relation or rule, representing the semantic distance between concepts(classes). The semantic distance is the opposite term of semantic relevance - in other words, the higher the value is, the less relevant the given two concepts are. Depending on the structure of the ontology, the resources can be stored in a single class, multiple classes or all classes may contain resources. In any case, the ultimate concern at this stage is defining the semantic distance among classes containing resources. Distances between other class can also be defined, but it would only be an intermediate step to assign the right value to rules.
Deciding the right value for each relation or rule can be tricky. The values can instantiated manually by domain experts, and some historical evaluation methods can be applied to update those values. Also, there are several researches on detecting the semantic similarity between concepts in both tree-based and graph-based structure, some of which are briefly mentioned in Section 2. However, the detailed methods for defining semantic distance between concepts are not the scope of this paper hence not listed here. 3.3
Semantic Relevance Measure between Resources To infer the semantic relevance between resources(instances), we propose a graphbased relevance measure to exploit various relations and rules between resources. The final semantic relevance value can be determined via the following five steps: Graph Creation The ontology is interpreted into a directed graph G = (V, E) where: – V is a set of nodes, representing all resources; – E is a set of edges, representing relations and rules between resources. Since additional rules as well as relations are interpreted as edges, it is possible that multiple edges e1(x, y), e2(x, y), . . . , en(x, y) are created between two nodes x ∈ V and y ∈ V . If there is an edge whose final node is its initial node, such edge is ignored hence not included in the graph.
∀x ∈ V :
e(x, x) ∈/ E
For each edge ei(x, y) ∈ E where x, y ∈ V , a numeric value dei(x,y) is assigned as its label to represent the distance from node x to node y. This value is derived from the distance between concepts(Section 3.2). The relevance value between two nodes is the inverse of the distance value.
relevance =
1
distance
(
n 1 Rxy = X
i=1 dei(x,y) The distance value de0(x,y) of the merged edge e0(x, y) is: de0(x,y) =
1
Rxy
By applying this edge merging algorithm above, a new graph with no multiple
edges with same direction between two adjacent nodes can be obtained. Fig. 3
shows an example illustration of merging multiple edges.
(
Single Path Distance In a weighted graph, the weight of a path is the sum of the weights of each visited edges, and this can be applied to calculate the distance of a single path in our graph. Therefore, for a path P (a1an) which visits n nodes a1, a2, . . . , an ∈ V , the distance of the path DP (a1an) can be defined as follows: n−1 DP (a1an) = X de0(akak+1)
k=1
However, it is found that it can produce undesirable results, and the desired semantic distance value of a path should be greater than the simple sum of the edge distances. More precisely, the value added should become greater every time it passes another node. Therefore, the above equation is replaced by the following one with a value k.
n−1 DP (a1an) = X(k × de0(akak+1))
k=1 Merging Multiple Paths The last step is to merge multiple paths between two nodes so that the final relation value can be obtained. This method is fundamentally the same as merging multiple edges.
For n paths P1, P2, . . . , Pn between two non-adjacent nodes x, y ∈ V , the
semantic relation value Rxy is as follows:
(