Semantic Similarity in Heterogeneous Ontologies ∗

Elisa ChiabrandoDipartimento di Informatica,Università di Torino, Torino,

Italychiabrando@di.unito.it

Silvia LikavecDipartimento di Informatica,Università di Torino, Torino,

Italylikavec@di.unito.it

Ilaria LombardiDipartimento di Informatica,Università di Torino, Torino,

Italylombardi@di.unito.it

Claudia PicardiDipartimento di Informatica,Università di Torino, Torino,

Italypicardi@di.unito.it

Daniele Theseider DupréDipartimento di Informatica,

Università del PiemonteOrientale, Alessandria, Italy

dtd@mfn.unipmn.it

ABSTRACTRecent extensive usage of ontologies as knowledge bases thatenable rigorous representation and reasoning over heteroge-nous data poses certain challenges in their construction andmaintenance. Many of these ontologies are incomplete, con-taining many dense sub-ontologies. A need arises for a mea-sure that would help calculate the similarity between theconcepts in these kinds of ontologies. In this work, we in-troduce a new similarity measure for ontological conceptsthat takes these issues into account. It is based on con-ceptual specificity, which measures how much a certain con-cept is relevant in a given context, and on conceptual dis-tance, which introduces different edge lengths in the ontol-ogy graph. We also address the problem of computing simi-larity between concepts in the presence of implicit classesin ontologies. The evaluation of our approach shows animprovement over Leacock and Chodorow’s distance basedmeasure [6]. Finally, we provide two application domainswhich can benefit when the newly proposed similarity mea-sure is used.

1. INTRODUCTIONWith the fulfillment of Semantic Web 2.0 vision, which

advocates rigorous representation of data semantics that ismachine processable, more and more ontologies are beingconstructed and used as knowledge bases in various domains.Ontology utilization enables easy and precise data descrip-tion and integration and reasoning over data.

But the wide-spread use of ontologies does not come with-out a cost: very often ontologies are being developed by var-

∗This work has been supported by PIEMONTE Project - People In-

teraction with Enhanced Multimodal Objects for a New TerritoryExperience.

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$10.00.

ious people, thus making them heterogeneous and often in-complete. Moreover, very often they consist of various sub-ontologies that describe different sub-domains with varyinglevel of detail and different structure.

Several proposals to compute a measure of similarity be-tween concepts have been proposed in the literature, as dis-cussed in Sections 3 and 10. But there is a need for a measureof similarity between ontological concepts that would takeinto account the above mentioned issues (incompleteness,heterogeneity and different sub-ontologies). In this work wepropose such a measure. We first introduce the similaritymeasure for a simpler case of an ontology, where we only con-sider explicitly defined classes, followed by the more complexcase where implicit intersection classes are also considered.

The work presented in this paper is situated in the contextof the Social Semantic Web, and in particular it focuses onsemantically enhanced, thematic social networking services.

A thematic social networking service is a Web 2.0 applica-tion that allows users to get in touch and share diverse con-tent regarding a specific theme (i.e., the domain of the socialnetwork), such as for example books (aNobii, LibraryThing),music (Last.FM), pictures (Flickr), Web links (del.ici.ous),etc. A semantic enhancement consists of an ontological de-scription that describes and relates the items in the domain,as well as other relevant items.

In particular, this research is part of the PIEMONTEproject [9], which aims at endowing real-life objects in thedomain of gastronomy with social abilities, and developingand integrating a set of social networking and augmentedreality tools that allow people to interact with these socially-enhanced objects. The focus is on wine and food productsas significant elements of the cultural heritage of a territory.The supporting mechanism is provided with a knowledgebase made of a set of ontologies describing the different foodand wine domains as well as the places where products canbe tasted or bought, the geographical region where they be-long, and the actions that people can perform on them.

A specific framework in the project is devoted to allowingusers to tell facts about domain objects and their experiencewith them [7]. Such a framework should provide a notionof similarity of facts, so that when a user inserts facts tothe system, the system can provide her with similar factsinserted by other users (e.g. her friends). A proper notion

of similarity of concepts is useful both in such a context (forsimilarity of predicates in facts, and domain items mentionedin facts) and in applications that allow navigating in thenetwork of domain items.

The paper is organized as follows: in Section 2 we de-scribe the underlying ontology as a semantic description ofthe domain for our setting. Section 3 looks into details oftwo distance based similarity measures that can be used insuch a setting. In Section 4 we introduce conceptual speci-ficity as a measure of how much a certain concept is relevantin a given context. The main contribution is given in Sec-tion 5 where we introduce conceptual distance between theitems of the ontology, based on conceptual specificity. InSection 6 we employ this new distance measure to calcu-late the similarity between domain items. Evaluation of ourapproach is described in Section 7. Next, we introduce theproblem of implicit classes in Section 8, followed in Section 9by some possible application scenarios for the proposed simi-larity measure. Related work is given in Section 10, followedby conclusions in Section 11.

2. SEMANTIC ORGANIZATION OFDOMAIN ITEMS

In this section we describe the ontology we use in our set-ting. The domain is modeled as an ontology O that containsa domain ontology and a predicate ontology. The domainontology defines and organizes the domain entities, whereasthe predicate ontology defines and organizes the predicates(Italian verbs) that can be used to describe the actions per-formed. Examples of predicates are: Drink, Walk, Listen,whereas examples of domain entities are: John Smith, Wine,Cabernet, March 23rd.

The domain and predicate ontologies are represented inthe Semantic Web recommendation OWL 2 as follows. Thedomain ontology includes:

• a product ontology, where some classes (e.g., Cheese,Wine) are defined in more detail, each with its own im-portant features (e.g., cow/goat/sheep milk for cheese);

• an actors ontology: persons (food producers, cooks,food sellers, restaurant owners) or virtual actors, i.e.other entities (restaurants, food shops) playing a rolein the gastronomic world; they are, of course, related tothe products by properties: producers produce prod-ucts, cooks and restaurants use them, sellers sell them;

• a geographical ontology, providing information aboutcommunities and larger areas that are places of originof products: the terroir (territory of origin), with itsclimate, type of ground and traditions, is in fact ofparamount importance for food quality (consider, e.g.,designated origin for a certain wine);

• a recipes ontology, with, at least, fundamental in-gredients for them; it is of course related to productsit uses, cooks and restaurants that serve the dish, theregion of origin in case of traditional recipes.

The predicate ontology contains two sorts of predicateclasses. Abstract predicate classes are non-lexicalizedconcepts, i.e. they are not associated with a specific verb(this is, actually, dependent on the specific natural lan-guage considered); they can be used to define role con-straints that are common to several predicates. However,

there can hardly be any interest in providing (direct) in-stances of abstract classes. On the other hand, concretepredicate classes are lexicalized and can be instantiated.Concrete predicate classes can have subclasses as well, sincethere may be verbs that further specify certain actions. Forexample, the predicate Devour is a subclass of the predicateEat.

3. DISTANCE BASED SIMILARITY MEA-SURES

In the literature there are three main approaches to mea-sure the similarity of two concepts in an ontology. The firstuses the definition of entropy, as in [12] (see Related workin Section 10 for more details). However, adopting this kindof approach requires a reasonably complete ontology, whichis something we do not want to assume. In the context ofSocial Web, the ontology of the domain is bound to grow de-pending on the social network usage, and it is fairly difficultto make any assumption about the completeness or even thehomogeneity of the semantic information.

Another approach is to use the ontology graph structure,more precisely, using directly the distance between nodes(usually, the number of edges or the number of nodes thatneed to be traversed in order to reach one node from theother), as introduced in [10]. Several measures of semanticsimilarity are based on such a distance; for a discussion andcomparison with entropy-based approaches see [2].

Finally, the third approach combines the information con-tent approach with edge counting based approach.

Due to the limitations of the ontologies we need to workwith, we focused on this second type of measure to define anew notion of semantic similarity. In our case it is difficultto obtain the frequencies of concepts in the taxonomy (andconsequently their probabilities) or have them provided apriori by domain experts. Hence, we take a closer look at twodistance based similarity measures that could be suitable tobe used in our setting.

• Leacock and Chodorow [6] use a similarity measurefor word sense disambiguation in a local context clas-sifier. The most similar nouns from the training setare substituted for the ambiguous ones in testing. Theauthors use the normalized path length in WordNet [4]between all the senses of the concepts being compared.The path length is measured in nodes rather than links,and the semantic similarity is computed as

simLC(a, b) = − log

(Np

2×MAX

)(1)

where Np is the number of nodes in path p from a tob and MAX is the maximum depth of the taxonomy.The length of the path between two same words (i.e.between members of the same synset) is 1.

Modifying slightly this measure we obtain:

simLCd(a, b) = − log

(dist(a, b)

2×MAX

)(2)

where dist(a, b) is the distance from a to b.

The disadvantage of this similarity measure is thatmany pairs of non-similar words are estimated as sim-ilar, due to the equal edge lengths in their hierarchy.

• Wu and Palmer’s [17] similarity measure accountsfor the depths of the given words in the taxonomy andof their common subsumer, which characterizes theircommonalities. Their measure is based on number ofnodes on the paths between the compared nodes. Theconceptual similarity between two nodes a and b, withthe first subsuming node c, is computed as:

simWP(a, b) =2Nc

N(a) + N(b)

(3)

where Nn is the number of nodes on the path from theroot to the node n.

This measure can be slightly modified, in order to ac-count for edge distances:

simWPd(a, b) =2dist(c)

dist(a) + dist(b)(4)

where dist(n) is the depth of n, i.e., its distance fromthe root.

4. CONCEPTUAL SPECIFICITYIn the following discussion, we consider an ontology as a

rooted, directed acyclic graph of inverse IS-A (subclass-of orinstance-of) relations, i.e., the arc is directed from class tosubclass or instance.

In general, the specificity of a concept is associated withthe depth of the corresponding node in the ontology, whilethe distance between two concepts is associated with thelength of the shortest path between the two nodes. How-ever, due to the lack of completeness and homogeneity ofthe ontology we use, there are two issues that should betaken into account:

• The domain ontology is obtained by putting togetherseveral sub-ontologies detailing different aspects of thesocial network domain. In the gastronomic domainfor example, we put together an ontology of wines, anontology of places where wines are sold or tasted, andan ontology of producers. This is generally done byadding some fairly abstract concepts to the hierarchythat have no practical significance with respect to thedomain itself - the best example is the Thing node thatis usually the ontology root.

• Due to the lack of homogeneity among the differentsub-ontologies, it can easily happen that concepts thathave the same depth in the graph are perceived byusers as having different conceptual specificity. More-over, the perceived specificity may vary depending onthe user context. For example, if the context is a winefair, the Wine concept is not specific at all, since ev-erything at the fair has probably something to do withwine. On the other hand, if the context is a Farmers’Market, then Wine becomes more specific, being oneamong many other things that can be found on themarket stalls.

Given the above considerations, we propose to distinguishtwo types of nodes in the ontology:

• Ground nodes: representing the notions perceived aspractically relevant in the domain. These nodes shouldhave a finite specificity value, possibly depending onthe context. The nodes with the same specificity value

should be perceived by users as similarly relevant froman ontological point of view in the considered context.1

Moreover, specificity should be monotonic with respectto depth.

• Sky nodes: representing abstract notions that are notconsidered relevant enough. These nodes should havea specificity value equal to −∞.

Figure 1 depicts an example of this partition.

Sky

Ground

Figure 1: An example of ontology partition intoGround and Sky concepts.

In order to partition the ontology into the Sky and Groundsets of nodes, we require that a domain expert pre-selectsa set S of surface nodes (highlighted in Figure 1) which isthe basis for defining Ground nodes and their depth. Thenodes in S represent the first domain-relevant concepts oneencounters while traversing the ontology.

Definition 1 Given the ontology O as a rooted, directedacyclic graph 〈N,E〉, and given a set S of surface nodes,the set of Ground nodes in O is the set

Ground = reach(S)

of nodes that are reachable from nodes in S.

A finite specificity value is assigned directly to surfacenodes, with 0 as a default value. As explained above, suchan assignment may be context dependent: in the example,Wine could have specificity value 0 in general, but a negativespecificity value in the context of a wine fair.

In defining the specificity for a ground node n, we con-sider the paths from surface nodes to n. We consider thateach step in such a path adds some specificity to a conceptcorresponding to node n with respect to its ancestors. Sincewe do not want to burden the modeler with measuring suchadditional specificity for each step, nor we expect to have,e.g., statistical information on the frequency of occurrence ofthe concepts, we consider each step as a unit of “additionalspecificity”.

For example, suppose that there are two paths of differentlength from surface nodes of specificity 0 to node n: theshorter one, covers, of course, the same specificity distance

1Ontologically relevant means that it offers a useful distinction in

relation to a theme. For example, even if a person does not considergenre an important factor in judging a book, he or she cannot dismissgenre as an irrelevant concept in the litterature domain, if only to beable to say that “genre does not matter”.

in fewer steps. However, in our interpretation, this meansthat some of its paths are longer than the minimal unit, andthen we use the length of the longer path. If a surface nodes has a non-zero specificity, we have to add it to the lengthof the path from s to n. We summarize the definition ofspecificity as follows:

Definition 2 Given the ontology O as a rooted, directedacyclic graph 〈N,E〉, a set of surface nodes S, and a functionspecS : S → N providing the specificity of surface nodes, theconceptual specificity value of a node n ∈ N is defined asfollows:

• if n ∈ S, then spec(n) = specS(n).

• if n ∈ Sky, then spec(n) = −∞ .

• if n ∈ Ground \ S, then

spec(n) = max{spec(s) + num(p) | s p−→ n, s ∈ S}

where sp−→ n means that p is a path from s to n and num(p)

is the number of edges traversed on the path p from s to n(including the node n).

5. CONCEPTUAL DISTANCEThe semantic distance between two ontological concepts

is often based on their shortest connecting path (in termsof number of edges) in the ontology graph. However, in ourcase, such a distance also depends on the specificity of thetraversed nodes, since we want to account for the fact thattwo specific nodes further down in the ontology graph aremore similar than two more general nodes higher up in thehierarchy.

A very simple example of this is the following: if weconsider the concept Drink and its two child nodes Wine

and Milk, we have that the shortest path between Wine

and Milk has the length 2. If we consider the Cabernet

concept (descendant of Wine) and its child nodes Caber-

net Wonder 2004 and Cabernet Merveille 2007, we havethat again the shortest path between the two children hasthe length 2. However, since Cabernet Wonder 2004 andCabernet Merveille 2007 are children of a more specificnode, their conceptual distance is perceived as smaller thanthe distance between Wine and Milk.

This problem is due to the fact that in the standard edgecounting approaches to similarity, all the edges are usuallyconsidered of the same length, thus representing uniformdistances between the nodes. In the taxonomies with certainvery dense sub-taxonomies, like ours, this is a problem, sinceit does not reflect the fact that the descendants lower downin the hierarchy are considered conceptually closer than theones higher up,

For this reason we define an edge length function that isparameterized with respect to the specificity of its sourcenode (i.e., its depth in the concept hierarchy), ensuring thatedge lengths decrease exponentially when going deeper un-derground (while being infinite for sky nodes means that weare not even interested in such distances). In particular:

Definition 3 Given an edge e : s → t, the edge length isgiven by

len(e) = k−spec(s).

where k ∈ N is a constant (k ≥ 2).

We then compute the conceptual distance between anytwo nodes as follows:

Definition 4 Given two nodes n1, n2 ∈ O, their conceptualdistance is the length of the shortest path2 connecting themvia an ancestor node:

dist(n1, n2) = min{len(p1) + len(p2) |∃g such that g

p1−→ n1, gp2−→ n2}.

Notice that whenever a path from n1 to n2 crosses a Skynode, then dist(n1, n2) = ∞. This corresponds to the casewhen n1 and n2 belong to different sub-ontologies, and aretherefore concepts of a different sort (e.g. a Wine and aRestaurant). We are not claiming that there is no relation-ship between these two concepts, but that they are ontolog-ically distant.

6. SEMANTIC SIMILARITY REVISITEDIn this section, we propose a new measure of similarity be-

tween ontological concepts, based on the conceptual speci-ficity (introduced in Section 4) and the conceptual distance(defined in Section 5). The new similarity measure is ob-tained by adapting Leacock and Chodorow’s definition ofsimilarity [6], given in Section 3, with the modified notionof distance in Section 5.

Definition 5

Given two domain entities n1 and n2 (nodes in the domainontology), we define their similarity as:

simLCd(n1, n2) = − log

(dist(n1, n2)

2×MAX

)(5)

where MAX is the maximum length of a path from a surfacenode to a terminal node in the ontology. When n1 = n2 thendist(n1, n2) = 0 and the nodes n1 and n2 become infinitelyclose.

As we would see in Section 7, this new measure of simi-larity brings some improvements over the original measureof Leacock and Chodorow [6]. It resolves the problem ofthe same distance between the nodes further down and thenodes further up in the hierarchy: the more specific con-cepts become less distant, and consequently more similar,than the more general concepts.

7. EVALUATION

7.1 Goals of the experimentThis section describes a simple experiment we conducted

in order to evaluate our proposal of measuring the seman-tic similarity between ontological concepts. In particular,our main goal was to compare human judgements with theresults computed by the system. Moreover, we wanted to an-alyze the performance of Leacock and Chodorow’s similaritymeasure [6] and Wu and Palmer’s similarity measure [17] us-ing a standard distance between the nodes in the ontologyand using our modified conceptual distance.

2considering the graph undirected.

7.2 Description of the experimentA total of twenty persons were chosen among the contacts

and colleagues of the authors, according to an availabilitysampling strategy.3 Half of the subjects were used as a ref-erence group, the other half was used as a control group tomeasure the correlation of judgements of different humansubjects, as in [12]. All subjects were native Italian speak-ers. They were asked to rate the similarity of 28 pairs ofItalian verbs from the system ontology, assigning the valueson the 5-point scale from 0 to 4 (0 meaning not similar atall, 4 meaning very similar) to the verb-pairs.

The pairs range from the ones expected to be classifiedas highly similar by human subjects (and measured as rel-atively highly similar), such as (Italian for) Talk-Speak, tothe other extreme of pairs classified as almost not similar,e.g. Die-Chat.

The ordering of pairs was random for each subject.

7.3 Results and discussionTable 1 reports the correlation between the reference group

human judgements and the control group human judgements,as well as the correlation between the reference group humanjudgements with the following similarity measures:

• simWPd: Wu and Palmer’s measure [17] with the stan-dard (uniform) edge distance;

• simWPd: Wu and Palmer’s measure with our concep-tual (exponentially decreasing) edge distance;

• simLCd: Leacock and Chodorow’s measure [6] with stan-dard (uniform) edge distance;

• simLCd: Leacock and Chodorow’s measure with ourconceptual (exponentially decreasing) edge distance.

Similarity method CorrelationControl Group 0.9060

simWPd 0.7942simWPd 0.6371simLCd 0.7637simLCd 0.8558

Table 1: Correlation results for semantic similarity

The correlation for the replication experiment with thehuman subjects (i.e. the comparison of the two groups ofhuman subjects) is similar to the one reported in [12]. Thebest results with respect to similarity measures is obtainedusing our modification of Leacock and Chodorow’s originalmeasure, with exponentially decreasing edge distances.

Even though in our experiment the original Wu and Palmer’smeasure provides better results than the original Leacockand Chodorow’s measure, our modified notion of distancedoes not fit well with Wu and Palmer’s measure (as long aswe are interested in a linear correlation with human judge-ments). In fact, in simWPd(a, b), where the denominatoris equivalent to dist(a, c) + dist(b, c) + 2dist(c), replacingdist with dist assigns a higher weight to the upper edges,and then the term 2dist(c) dominates the term dist(a, c) +dist(b, c). As a result, while simWPd(a, b) tends to use most

3Even though non-random samples are not statistically representa-

tive, they are often used in psychology research and usability testing,during early evaluation phases.

of the range [0, 1], when using the modified distance, thesimilarity values obtained with the measure simWPd(a, b) aresqueezed towards 1.

8. IMPLICIT RELEVANT CLASSESThe definitions for specificity and distance given in the

previous sections are based on the structure of the ontologyas a graph. The underlying idea is that the ontology explic-itly includes all and only the concepts that are considered“important”, i.e. relevant for such structural measures. How-ever, it may be the case that several orthogonal structuresare important, e.g., in the case of cheese:

• based on the type of milk: cow’s milk cheese, goatcheese, sheep cheese;

• based on freshness, from fresh cheese to aged cheese;

• based on texture, from soft cheese to hard cheese.

At the same time, some other features (e.g. shape) will bepresent in the ontology, as a part of knowledge about typesof cheese, but would not be considered relevant for classifi-cation of types of cheese (cylindric cheeses etc).

In this case, we can assume that most intersections of rel-evant classes are relevant, without expecting the modeler toexplicitly introduce all such intersections (fresh goat cheese,semi-hard cow’s milk cheese, etc)4.

Suppose that C1 and C2 are two types of a Fresh Goat

Cheese, as in Figure 2. Both the specificities of C1 and C2

and their conceptual distance from one another, as definedpreviously, are affected by the presence or absence of theclass Fresh Goat Cheese. The presence of the class Fresh

Figure 2: An example of intersection

Goat Cheese would add specificity to C1 and C2 and reducetheir distance because the path through Fresh Goat Cheese

would be shorter (significantly shorter, given the exponentialdecrease of distance) than the paths through the less specificFresh Cheese and Goat Cheese.

Let us discuss in detail how specificity could be redefinedfor ground nodes, if all such intersections were present. Con-sider a ground node n and let P (n) = {p1, . . . , pm} be theset of all its parents having no descendants in P (n) (seeFigure 3).

Without the intersection classes, the length of the longestpath from surface nodes to n would be obtained by adding

4Some of these intersections, e.g. hard fresh cheese, may be useless

in the sense that they fail to have subclasses and instances in thedomain; however, this does not affect the modified definitions below,since we only consider the intersections that are superclasses of someexisting class.

Figure 3: Modified specificity

1 to the specificity of the parent with the highest specificity.The presence of all the intersections would increase, how-ever, from 1 to m the number of edges from any pi to n.Therefore, Definition 2 of conceptual specificity is modifiedby defining the specificity for ground nodes inductively usingthe specificity of surface nodes as follows:

Definition 6 For a non-surface ground node n, letP (n) = {p1, . . . , pm} ={p | p is a parent node of n, p has no subclass in P(n)}.Then

spec(n) = max{spec(pi) | i = 1, . . . ,m}+ m.

In order to modify the definition of conceptual distancebetween two nodes n1 and n2, while simulating the presenceof all intersection classes, we have to consider the minimalupper bounds {g1, . . . , gm} of the nodes n1 and n2 (FreshCheese and Goat Cheese in in the above example). In gen-eral, such upper bounds might have different depths and thepaths to n1 and n2 might involve several edges. With theaddition of all intersection classes, the shortest path (in thesense of Definition 4) would go through the single minimalupper bound

g = g1⋂

. . .⋂

gm.

The length of the paths gp1−→ n1, g

p2−→ n2, whose sum

len(p1) + len(p2) = dist(n1, n2)

can be computed from the specificity of g and the numberof edges in the paths.

The specificity of g would be the one of the most specificgi, increased by m− 1 (similarly to Definition 6).

The number of edges in p1 (and similarly for p2) is com-puted as follows. We consider the set I = {i1, . . . , ih} of theintermediate nodes on the paths from any gj (j = 1, . . . ,m)to n1, excluding those that are defined as intersections oftwo ancestor nodes in the paths (to avoid counting themtwice, since we are simulating the presence of intersections).Nodes that lie between g and n1 are those that can be ob-tained by intersecting g in all possible ways with the nodesfrom I.

In particular, node n1 can be reached from g traversing h+1 edges, at any step intersecting with one of the h membersof I, e.g. through the nodes g∩i1, g∩i1∩i2, . . ., g∩i1 . . .∩ih.

In this case we calculate len(p1) as:

len(p1) =

h∑j=0

k−spec(g)+j

9. APPLICATION DOMAINSIn this section we present two applications which would

benefit from using a measure of similarity of two domainentities.

The main application developed as a part of the PIEMON-TE project is called WantEat Mobile [9]. It is a mobileapplication developed for Apple iPhone whose main goal isto allow communication between people and “enhanced ob-jects”. Objects are enhanced in that (i) they do not merelycollect user-generated content, but they rather synthesizeand merge it in a sort of “personal profile”, and (ii) theycan establish social relations with other objects (or people).Social relations between objects can develop and grow de-pending on user-generated content (e.g. tags, comments,etc.) but also depending on the semantic description of theapplication domain.

(a) (b)

Figure 4: Two screenshots of the Wheel interface

WantEat Mobile’s main feature is the Wheel [1, 3], anintelligent browsing system that allows users to navigate thesocial network and discover some territory through the linksbetween its products, its places, and the people that workor live in it. Once the user gets in touch with an object, letus say a type of wine, this object is placed at the center ofthe wheel interface, surrounded by the members of its socialnetwork (or some of them, according to the user model andpreferences, and a set of filters the user may set), dividedin four sectors according to their type. The user can dragan element from a wheel sector to the center, making it thefocus of attention and navigating its social network. Figure 4shows (a) the wheel interface, and (b) the act of moving anobject to the wheel center.

In order for objects to be able to synthesize their ownpersonal profile and present it to users, a collective storybuilding framework has been studied. This frameworkallows users to tell about their experience with the domainobjects, at the same time offering a suitable representationto perform automated reasoning on the information acquired[7]. In this framework users can compose objects, people andother items (such as time-stamps or customized labels) intofacts by joining them to a leading predicate and giving themdifferent roles (e.g., subject, object, etc.) with respect to it.

The fact database can be used by the system, togetherwith the ontology knowledge base, to (i) aggregate infor-mation concerning the given object; (ii) derive relations be-tween objects or objects and people, and (iii) search for thefacts similar to the facts inserted by the user that may be of

a certain interest to her. To this aim we introduced the mea-sure of pertinence, that given two facts f1 and f2, providesan evaluation of “how much they have in common”. In par-ticular, our definition of pertinence is based on: conceptualsimilarity between domain entities and between predicates,social relatedness of people participating in the facts, pro-vided by the user model and co-location which provides anapproximation of the things happening at the same time andplace.

In both cases, the same ontology is being used (describedin Section 2). This ontology is not (and does not aim atbeing) complete, in the sense that some sections of a givendomain could be left out as no participants in the social net-work are interested in them. In addition, we cannot gener-ally assume that for a given class (e.g. Bettelmatt Cheese)all the instances (e.g. Bettelmatt Cheeses from differentproducers) are present, as this depends on the producersthat join the network. Finally, our ontology contains a num-ber of very dense sub-ontologies, developed with a particularlevel of detail, as opposed to the other more general ones.

Both WantEat Mobile and the storytelling framework relystrongly on their semantic knowledge base and aim at inte-grating this “expert” information with user-provided con-tent. Hence, the notion of conceptual similarity, measuringhow much two “notions” described in the ontology are simi-lar to each other, is quite central to their proper functioning.This is relevant both to the Wheel feature, as similar objectscan easier become friends, and to the storytelling framework,as in order to find similar stories told by different users wefirst need to be able to asses how similar the things they talkabout are.

10. RELATED WORKIn his seminal paper, Resnik [12] introduces the notion of

semantic similarity based on information content in an IS-Ataxonomy. The information content of a class in a taxon-omy is given by the negative logarithm of the probability ofoccurrence of the class in a text corpus. This means thatthe more abstract classes provide less information content,as opposed to more concrete and detailed classes lower downin the hierarchy. Concept probabilities are computed as rel-ative frequencies (each noun in the text corpus is countedas an occurrence of each class that contains it.). The closestclass that subsumes both compared concepts, called a mostinformative subsumer, provides the shared information forboth, and gives the measure of their similarity:

sim(a, b) = maxc∈S(a,b)[− log p(C)] (6)

where p(c) is the probability of encountering an instance ofconcept c, and S(a, b) is the set of all concepts that subsumea and b. Resnik argues that his approach shows better per-formance results than edge-counting approach, using humansimilarity judgements as a benchmark.

A very appealing approach to measuring semantic similar-ity is given by Jiang and Conrath [5] where edge countingapproach is improved with information content one. Thestrength of a link connecting a child to its parent is the dif-ference between information content values of the parent andthe child. The weight of a link, in addition to link strength,takes into account local and average densities, depth of theparent node in the hierarchy and link type. Then the dis-tance between two nodes is calculated as the shortest path

linking the two nodes, using weighted links to traverse thepaths. Experimental results show that this combined ap-proach performs better than information content approach.

Lin [8] proposes an information-theoretic definition of sim-ilarity derived from a set of assumptions about similarity. Itis measured as the ratio between the amount of informationthat two concepts have in common and the amount of infor-mation needed to fully describe them, thus taking into ac-count similarities, as well as differences, between comparedterms. His similarity measure is not tied to a particularknowledge representation and is applicable to any applica-tion with a probabilistic model. This allows using his mea-sure in the applications in which similarity measure couldnot be introduced before.

For an evalution of these similarity measures, see [2].Smeaton and Quigley [13] describe an application that re-

trieves the image captions based on the user query. Startingfrom a corpus of image captions and a collection of queries,they index the queries and the captions by the words oc-curring in them and then use semantic similarity betweenindex terms to calculate the query-caption similarity. Theword-word similarity is determined using a set of hierarchicalconcept graphs derived from WordNet [4] and the measureof semantic similarity is based on the work of Resnik [11].Among the experimental runs of different set similarity al-gorithms, the runs most interesting for our work are: (i) therun introducing a word-word similarity threshold to elimi-nate the words with low similarity and (ii) the run in whichthe most similar caption term for each query term and themost similar query term for each caption term are calculatedand used in overall sum of similarity values.

In the application described by Tudhope and Taylor [16],the similarity measure is used to improve automatic genera-tion of links in hypermedia navigation, based on three mea-sures of similarity: subject, temporal and spatial. In caseof subject and spatial dimensions, they calculate the short-est paths connecting two terms. Each traversal between twodirectly connected terms has a corresponding cost factor as-sociated to it and in the subject dimension the cost factordepends on the type of relationship and the depth in thehierarchy. This ensures that the siblings deeper in the hi-erarchy are semantically closer than the siblings higher up.In the spatial similarity measure a variation of a branch andbound search algorithm produces multiple paths to a solu-tion where semantic similarity is below zero. The temporalsimilarity measure calculates similarity between two timeperiods or points in time. As in our case, they employ amaximal set similarity algorithm [15] which sums the maxi-mum similarities for each term with respect to the membersof the other set and normalizes them.

Sussna [14] uses the notion of semantic distance betweensemantic network nodes in order to disambiguate word sensesin document indexing, using WordNet [4] to provide the se-mantics. In his edge weighting algorithm, he considers eachedge as consisting of two inverse relations. He gives a weightrange, depending on the number of descendants of the sametype, to each of the nine possible relations that connectnouns in WordNet. For each edge the two inverse weightsare averaged and the result is divided by the depth of thedeeper node (depth-relative scaling). Semantic distance orrelatedness between two nodes is calculated as the weight ofthe shortest path connecting the two nodes. Then, multiplesense words appearing next to each other in the document

are disambiguated by choosing the senses with minimal pair-wise distance between senses (calculating each combinationof possible senses). This approach, called mutual constraintamong terms, is very similar to our definition of similarity.

11. CONCLUSIONS AND FUTURE WORKFinding an adequate measure of similarity to be used in

heterogenous (and very often incomplete) ontologies is achallenging task. In this work we addressed this problem,proposing a new approach to measuring the similarity be-tween the concept of the ontology, based on the conceptualdistance between elements of the domain, and in particular,on their conceptual specificity.

Conceptual specificity is a measure of relevance of the do-main concepts in a certain context. It allows for the partitionof the ontology into relevant and non-relevant parts whichbrings the focus only to the important concepts. The con-cept of conceptual specificity can be introduced into existingsimilarity measures in order to use conceptual (weighted)distance.

We distinguish two different kinds of ontologies: ontolo-gies with underlying graph structure which includes all therelevant classes and ontologies where implicit relevant classeshave to be taken into account.

We tested two newly obtained measures of similarity forontologies which do not take into account implicit classes,more precisely, we tested the ontology of italian verbs. Wealso compared our similarity measures to the original mea-sures of Leacock and Chodorow [6] and Wu and Palmer [17],using both, the standard (uniform) edge distance and ourconceptual (exponentially decreasing) edge distance. Ourexperiment showed that the best results are obtained us-ing our conceptual (exponentially decreasing) edge distancewith Leacock and Chodorow’s measure.

The most imminent next step to take is to test our simi-larity measure for the ontologies with implicit classes. Thisis not a trivial task, since we would like to test it on a veryspecific ontology (for example, the Cheese ontology), and forthat we need to include the experts in the field.

Further next step is to employ the given similarity mea-sure in the WantEat Mobile application [9, 3] and see if theproposed items satisfy the user expectations. Another do-main where we want to test our similarity measure is in thedigital storytelling framework [7], where the similarity mea-sure would be used to calculate the similarity of the conceptsused to build facts, and consequently to calculate the per-tinence of facts regarding a certain theme in a domain of aSocial Networking Service.

In our work we only consider IS-A links in the ontology.Considering also PART-OF/MEMBER-OF links would def-initely contribute to the accuracy of measured similarity be-tween domain items.

12. REFERENCES[1] G. Biamino, P. Grillo, I. Lombardi, A. Marcengo,

A. Rapp, R. Simeoni, and F. Vernero. “The Wheel”:an innovative visual model for interacting with asocial web of things. In Visual Interfaces to the Socialand Semantic Web, VISSW ’11, Palo Alto, CA, USA,2011. Accepted for publication.

[2] A. Budanitsky and G. Hirst. EvaluatingWordNet-based Measures of Lexical Semantic

Relatedness. Computational Linguistics, 32(1):13–47,March 2006.

[3] E. Chiabrando, R. Furnari, P. Grillo, S. Likavec, andI. Lombardi. Dynamic interface reconfiguration basedon different ontological relations. In 14th InternationalConference on Human-Computer Interaction, HCII’11, LNCS. Springer-Verlag, 2011. Accepted forpublication.

[4] C. Fellbaum, editor. WordNet: An Electronic LexicalDatabase. The MIT Press, Cambridge, MA; London,1998.

[5] J. Jiang and D. Conrath. Semantic similarity based oncorpus statistics and lexical taxonomy. In Proc.International Conference on Research inComputational Linguistics, pages 19–33, 1997.

[6] C. Leacock and M. Chodorow. Combining local contextand WordNet similarity for word sense identification,pages 305–332. In C. Fellbaum (Ed.), MIT Press, 1998.

[7] S. Likavec, I. Lombardi, A. Nantiat, C. Picardi, andD. Theseider Dupre. Threading facts into a collectivenarrative world. In R. Aylett, M. Y. Lim, S. Louchart,P. Petta, and M. Riedl, editors, 3rd JointInternational Conference on Interactive DigitalStorytelling, ICIDS ’10, volume 6432 of LNCS, pages86–97, Edinburgh, UK, 2010. Springer-Verlag.

[8] D. Lin. An information-theoretic definition ofsimilarity. In Proc. 15th International Conference onMachine Learning ICML ’98, pages 296–304, SanFrancisco, CA, USA, 1998. Morgan KaufmannPublishers Inc.

[9] PIEMONTE Team. WantEat: interacting with socialnetworks of intelligent things and people in the worldof enogastronomy. In Interacting with Smart ObjectsWorkshop 2011, Palo Alto, CA, USA, 2011. Acceptedfor publication.

[10] R. Rada, H. Mili, E. Bicknell, and M. Blettner.Development and application of a metric on semanticnets. IEEE Transactions on Systems Management andCybernetics, 19(1):17–30, 1989.

[11] P. Resnik. Semantic classes and syntactic ambiguity.In Proc. Workshop on Human Language TechnologyHLT ’93, pages 278–283. Association forComputational Linguistics, 1993.

[12] P. Resnik. Semantic similarity in a taxonomy: Aninformation-based measure and its application toproblems of ambiguity in natural language. Journal ofArtificial Intelligence Research, 11:95–130, 1998.

[13] A. F. Smeaton and I. Quigley. Experiments on usingsemantic distances between words in image captionretrieval. In Proc. 19th Annual International ACMSIGIR Conference on Research and Development inInformation Retrieval SIGIR ’96, pages 174–180, NewYork, NY, USA, 1996. ACM.

[14] M. Sussna. Word sense disambiguation for free-textindexing using a massive semantic network. In Proc.2nd International Conference on Information andKnowledge Management CIKM ’93, pages 67–74, NewYork, NY, USA, 1993. ACM.

[15] C. Taylor. Semantic modeling and navigation in amuseum hypermedia system, 1995. PhD Thesis,University of Glamorgan.

[16] D. Tudhope and C. Taylor. Navigation via similarity:

Automatic linking based on semantic closeness.Information Processing and Management,33(2):233–242, 1997.

[17] Z. Wu and M. Palmer. Verbs semantics and lexicalselection. In Proc. 32nd annual meeting onAssociation for Computational Linguistics, pages133–138, Morristown, NJ, USA, 1994. Association forComputational Linguistics.