Xiulei Liu , Xia Hou Junyang Yu, Ying Gao Yue Zhang Yingying Zhang

1 Beijing Key Laboratory of Internet Culture and Digital Dissemination Research, Beijing Information Science and Technology University,100101, China

2 State key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, 100876, China

3 Software School, Henan Univesity, 475001, China

* The corresponding author is Junyang Yu, email: jyyu@henu.edu.cn

I. INTRODUCTION

Many ontologies represent semantic sensors data[1][2]. However, the heterogeneity between the different semantic sensor ontologies makes service operators difficult, such as semantic inferring, non-linear inverted index establishing, service composing via Internet of Things(IoT).

In general, two main issues need to be addressed in order to share, reuse, and interoperate the knowledge represented through different sensor ontologies for providing better services via the IoT. The first one is how to align different sensor ontologies; the second one is how to access the alignments via the IoT.

To address the first issue, many solutions for aligning sensor ontologies are proposed as shown in [3], [4], [5], [6], [12], [13], [14].The existing technologies mostly use element information (i.e. element names text), structure information (i.e. concept structure information) and instances information to perform sensor ontology alignment by calculating coefficients in the [0, 1] range[15],[20]. In this article, we put forward the method based on lexical analysis by processing the senses of words in annotations through WordNet and semantic analysis by extending structural subsumption reasoning algorithms.

To address the second issue, most solutions express alignments by using an expressive alignment language. However, many IoT service require common standards for expressing and accessing alignments to enable the easy and seamless integration of data via the IoT.The Simple Knowledge Organization System is a general data model and widely accepted standard for sharing and connecting knowledge organization systems. It can be utilized to exchange knowledge between IoT services[26][29]. We adopt the SKOS model to define the integration sensor ontology to express association links between entities from different sensor ontologies. Our solution (called ALOWS)combines the description logics (DL)-based ontology alignment solution with the SKOS model to produce alignments and make them to be utilized via the IoT.

The work of this paper contains the following four points: Firstly, we propose a comprehensive framework. At lexical analysis stage,it analyzes annotations in n sensor ontology to define the representation of an entity image by composing the appropriate sense and its extension of each word and then identifies and filters candidates. At semantic analysis stage,it parses various constructors and axioms in a sensor ontology to make the implied semantic and lexical information to be read-off easily by rephrasing entities into normal forms, and then compares the syntactical structure of normal forms, by adjusting the allowable degree of discrepancy between entities through extending structural subsumption reasoning algorithms, to infer alignments. Secondly, ALOWS converts the alignments obtained from the previous stage into property assertions in the SKOS model to construct the integration sensor ontology which can be accessed via the IoT. Thirdly, we check the consistency of the integration sensor ontology to ensure that it satisfies the integrity conditions in the SKOS model. Eventually, we propose an evaluation of ALOWS to compare with 13 methods and make an example of the integration sensor ontology.

The rest of the paper is as below.

Section 1 discusses background and related work. Section 2 presents a sensor ontology alignment framework. Section 3 explains the definition of the integration sensor ontology in the SKOS model. Section 4 demonstrates the results of the experiments. Section 5 provides work summary and some prospects of future work.

This paper presents that utilizing the mapping rules and thresholds determines the SKOS mapping properties between entities.

II. BACKGROUND AND RELATED WORK

2.1 Background

We only focus on sensor ontologies in the OWL-DL form. An OWL-DL ontology includes multiformconstructors,axiomsand facts, that are corresponded to DL syntaxes,meanwhile,annotations(an entity label and comment) are the source of non-semantic information. Horrockset al. in[19]have shown that the OWL-DL is corresponded toSHOIN(D) which isSHOINwith support of data values, data types and datatype properties.

This paper has assumptions shown below.The subClassOf (C1?C2) and subPropertyOf(R1?R2) are calledinclusion axioms, and we assume that C1and R1are named concepts or properties. The equivalentClass (C1≡ C2) and equivalentProperty (R1≡R2) are calledequality axioms, and we assume that one of C1and C2must be named, and also one of R1and R2must be named. The disjointWith (C1⊥C2) is calleddisjoint axioms, and we assume that one of C1and C2must be named. We also have supposed that no cycles for entities (cyclic concepts and properties) exist in a sensor ontology.

Semantics in an OWL-DL sensor ontology can be seen as the knowledge-base (KB) (see[21], [22] for more details). When aligning sensor ontologies, we usually use implicit informations, so we need to analyze the KB including: TBox, PBox and ABox. Because this paper does not focus on individual information, we only analysis TBox and PBox in the KB. So entities only refer to concepts and properties in this paper.

It is necessary to align a datatype property to an object property respectively from different sensor ontologies in the OWL-DL form.This paper does not differ a datatype property and an objec propertyt, and abstracts them asproperties(roles). In this case, the OWL-DL syntaxes are the same to theSHOINsyntaxes.Based on the considerations above, datatype properties are converted into object properties as roles by steps below: first convert the range(datatype) of a datatype property as a concept;then asserte the value in this datatype as an individual of the corresponded concept; finally convert this datatype property into an object property. The converted sensor ontology is denoted asOcwhich only contains axioms and constructors from theSHOINsyntaxes. In the rest of this paper, only refer to the sensor ontologies in this form.

The SKOS is one of application of the Resource Description Framework (RDF) which permits concepts to be composed, published,linked with data via IoT and integrated into other concept schemes. In this paper, we only focus on some vocabularies in the SKOS model for our purpose, includingskos:Concept,skos:ConceptScheme,skos:inScheme, andSKOS mapping properties. The skos:Concept and skos:ConceptScheme are each instances of the owl:Class. The skos:inScheme and SKOS mapping properties are each instances of the owl:ObjectProperty. In our paper, we use the concept assertions of the skos:ConceptScheme to represent the aligned sensor ontologies, use the concept assertions of the skos:Concept to represent entities (SKOS conceptfor short)in the aligned sensor ontologies, and use the skos:inScheme to represent that which scheme a SKOS concept is taken from. The SKOS mapping properties describe relations between the SKOS concepts from the aligned sensor ontologies.

2.2 Related work

More and more attention has been paid to the study of ontology alignment. [3], [4], [5], [12],[13], [14], [52], [53][60]shown several current methods of ontology alignment which can be devided into different classifications by different views. In general , the basic information used by most solutions are as follows:

? Lexical information, such as terminology acquired from entity labels or comments.

? Structural information, which is part of ontology represented language, such as Sub-Class, is-a relation and SupProperty.

? Semantic information, such as model-theoretic semantics and RDF compatible semantics, obtained from syntaxes. This information is implied in constructors and axioms. Usually, formal semantics has many uses. These formal semantic can be used by tableau algorithms do resoning, and be used by structural subsumption algorithms to solve non-standard inference problems.

? External information, which can be obtained from other resources, (e.g., WordNet and Linked-Open Data1http://linkeddata.org/).

? Individual information which usually implied in concrete instances.

Most solutions focuse on lexical information, structural information and external resources. Some inference technologies will use the semantic information. Adoptting what kind of technologies depends on what kind of solution by using informations above. Some of the technologies contain similarity flooding[23],[28],[49], coefficient computation[25],[43], graph matching[16],[31],[32],[34], formal concept analysis[17],[33], machine learning[18],[19],[36], Bayesian decision theory[44], vector space models[27][61], hybrid solutions[30],[35],[45],[56], Dempster Shafter theory of evidence[9],[57], association rule paradigms[8], Markov networks[42],[58],reasoning mechanisms technologies[15],[37],[38],[43], [46], [59].

Analyzing data obtained from external resources by means of using machine learning techniques, lexical information, structural information, and instances expressed in a sensor ontology, GLUE[36]offers an automatic ontology alignment method. The automated semantic aligning of ontologies (ASMOV)[43]gets calculate similarity by lexical and structural expressions of two ontologies. SOBOM[23]generates candidates by using the four kinds of information as showed above. Then by similarity values which generated by the similarity flooding algorithms to compute similarity values of candidates, then we can get the difference between concepts. ALOWS makes use of structural information indirectly,namely, using semantic information including the structural information like sup- and subrelations.

DSSim[9]combines lexical matchers with the graph to represent the structure information of ontologies. It uses the Dempster Shafer theory of evidence to increase the correctness of alignments, it also uses synonyms from WordNet. AROMA[8]relies on implication intensity measures to select rules between entities by two criteria permitting to respectively assess the implication quality and the generativity of the rules. Lily[34]extracts semantic sub-graphs by using a graph model by analyzing the semantic presentation of elements in an ontology graph. The purpose of Semantic description documents which are constituted by the semantic sub-graphs is to gets the literal similarity between ontologies. In order to address the aligning problem without rich literal information, Lily employs similarity propagation algorithms with propagation conditions.

According to deductions, most of the semantic solutions in ontology alignment can be divided into two main categories: DL reasoning and Propositional Satisfiability problem(SAT).

The input of current SAT deciders is Conjunctive Normal Form (CNF) formulas[15].Semantics syntaxes are not used well by CNF in DL. This leads to that ontology alignment usually not use the result of the methods based on CNF (such as S-Match[13]) in the OWLDL form. S-Match provides the distinction between two notions. The image of a label is context-insensitive concerning only the senses of words in this label by using Word-Net. That focuses on lexical information in the label. The image of an entity in schemes is context-sensitive. The logic expression of the image is calculated as the intersection of all labels of the images from the root entity to this entity itself. S-Match inputs the images of the two entities from different schemes into SAT deciders to infer if there is any semantic relation between them. But it cannot exploit the semantics implied in an OWL-DL ontology. ALOWS, S-Match and DLOM[37]are similar in utilizing WordNet, the difference is ALOWS and DLOM obtain the suitable sense of each word, they use the extended ones as well. ALOWS and DLOM not only focus on comment and an entity label, but also use the logics and concepts in the range and domain of properties which are used to increase scopes and accuracy when processing properties.DLOM utilizes semantic information with the help of reasoners, but ALOWS exploits semantic information by extending structural subsumption algorithms (see [6], [11], [21]for details about structural subsumption algorithms).

In general, the reasoners based on DL reasoning adopt tableau algorithms which use negation to decrease the subsumption to (un)satisfiability of concept descriptions. Such methods as ILIADS (Integrated Learning In Alignment of Data and Schema) [46], Ctx-Match [47], Kosimap (Knowledge Organization System Implicit Mapping)[48]and DLOM,used this technique to process all sorts of syntaxes in an OWL-DL ontology and use semantics to some extent.

ILIADS first defines similarity measures only based on the usage count of senses in identifies candidates and WordNet. But,ALOWS, S-Match and DLOM make the best of relations in WordNet, which acts as a significant role in aligning ontologies. Because of executing a constant numberNof inference rules, ILIADS cannot check whether all restrictions are consistent or not, it is not comprehensive. DLOM validates this Through the use of reasoners directly. ALOWS ensures this by extending structural subsumption reasoning algorithms.

CtxMatch submits expressions into a reasoner. It directly obtains relations between entities from the output of this reasoner. By using reasoners, Kosimap uses logical consequences embedded in ontologies to calculate three different types of similarity measures for each pair of entities.

ILIADS, Kosimap and DLOM utilize the consequences from reasoners as one of the factors for calculating similarity values. S-Match and CtxMatch input the expressions derived from lexical analysis stage into reasoners to infer association links. They reflect semantics in an ontology to some extent. These have the advantages over the methods which only take structural information as similarity measures.However, they don’t exploit the implied semantics enough expressed by constructors and axioms in an ontology (such as ?, ?, ? and?). For example, two groups of axioms are given:

?Book?Reference, Reference?Publication;

?Book?Publication, Publication?Reference;

they can produce the same consequence from reasoners, namelyBook?Publication; however the two groups reflect different semantics. So, this approach of using reasoners does not always reflect all semantics in ontologies as shown in the example above. ALOWS makes the semantics implied in an ontology to be read-off easily by rephrasing entities into normal forms and then compares the syntactic structure of normal forms to infer alignments.

III. ALOWS FRAMEWORK

The main work flow of ALOWS, as shown in Fig. 1, is divided into two major stages: alignment stage and construction stage. The input of the system is different sensor ontologies,and outputs the integration sensor ontology in the SKOS model, which reflects the automatically defined association links between entities from the inputted ontologies. In this paper, the example ontologies, namelySensor Ontology leftandSensor Ontology right, are separately adopted benchmark dataset from Folders 101 and 302 in the Ontology Alignment Evaluation Initiative (OAEI). So we uses＜left (or right):an entity label＞to cite entities inSensor Ontology leftorSensor Ontology right.

The alignment stage has two sub-stages:lexical analysis sub-stage and semantic analysis sub-stage are as shown in Section 3.1 and Section 3.2 respectively. The alignment stage follows a novel method based on the key idea that knowledge engineers express the meaning of an entity by words in entity labels and comments and then interpret them by the semantics of an entity which derives from restrictions such as ?, ?, ≥ and ?. The alignment stage generatesa set of association links(calledan alignment), which are a four-tuple as defined below:

whereCandDare from different ontologies and the same entity type (attributes or concepts);Relationincludes ? (beIncluded), ≡(equivalent), and ? (include); Similarity Scoreis calculated as shown Section 4.1.

So the lexical analysis sub-stage analyzes annotations in a sensor ontology to produce candidates which reflect lexical relations between entities from different ontologies.The semantic analysis sub-stage parses constructors and axioms to exploit semantics in an onology by extending structural subsumption algorithms (usually be used to solve non-standard inference problems). So, to infer alignments between entities from different ontologies, ALOWS focuses on both lexical information in annotations and semantics in axioms and constructors.

The construction stage as shown in Section 4 defines the integration sensor ontology,which satisfies the consistency and integrity conditions of the SKOS model, when deriving alignments by the previous stage. The integration sensor ontology can be accessed via the IoT and provides an interconnection between two ontologies using a set of mapping properties in the SKOS model.

3.1 Lexical analysis

One of the main difficulties in aligning entities from different ontologies is that the design of ontologies has certain background knowledge and usually associated with specific context.The information does not usually be a part of a sensor ontology specification. The lack of background knowledge makes it difficulty when aligning ontologies because it generates too many ambiguities in interpreting and analysing the senses of words [15]. ALOWS discusses the extension of vocabularies using WordNet for tacking this problem. Groups nouns, adjectives, verbs, and adverbs into sets of cognitive synonyms in WordNet can be denoted as ?synsets?. Each synset expresses a different concept and gives the general and short definitions of this concept.

The Tokenizer module in Fig. 1 parses all entity labels and comments into words, e.g., if an entity label is “softCopyFormat”, it turns into＜soft, copy, format＞.

It also processes the plural, past participle and present participle of a word (see [41] for more details).

By using a statistical approach with the help of WordNet, the Synset Finder module discovers thesuitablesynset(sense) of each word represented in a natural language form extrated from a sensor ontology (see [37] for more details). The Synset Extension module extends the suitable sense of each word into a set of theextendedsynsets for this word, each of which lives a unique relationship with the suitable synset of this word in WordNet (such asrelated, derived_from_adj, pertainymandderivationally). It is contribute to include the possible synsets of a word in a existing context rather than just finding the suitable synset.Despite this function can not guarantee to find the best senses of words, it expands the range of the synset representation for each word.For instance, there are two statements:＜Book,publishedBy, IEEE＞;＜Book, hasPublisher,IEEE＞,here, ?publish? and ?publisher? have similar role or effect. So the extension of synonyms ?publish? will contain the suitable synset of ?publisher?. We denote the suitablesynsetplus a word’s extension as a group ofusefulsynsets of this word.

When determining the suitable synsets and their extension of all words, what we need is to find any two synsets between where they have relation from each sensor ontology, and define the sematic relations, then construct the Synset Relation Ontology based on these relations. When inferring if two entities can be a candidate, we have to get relations between synsets, which compose an entity image, by reasoning over the Synset Relation Ontology.

The Representation of an Entity Image module computes the image of an entity,namely what the knowledge engineer expresses in this case we ignore the effect of DL,expressed asIM(E) whereEis an entity. The formal definition ofIM(E) is described as below:

In Eq. (1),La (E)is a group of words in an entity label, andCo (E)is a set of words in an entity comment. We take the conceptleft: Bookas an example. They should beLa (left: Book) ={Book},Co (left: Book)={collection, book,written, monograph,or,text}represents a set, each element of which is consisted of theielements chosen from thenelements ofA. We can convert this into combination questions. For example,{monograph}, {written} }.

In Eq. (2) ,emeans an element ofW(E)in Eq. (1);Wimeans an element ofe;Wi.USmeans a suit of the helpful synsets ofWi;simeans an element in this set;means the intersection of several synsets.

When defining an entity image, we should identify candidates, which express sense logical relations between entities, by using reasoners over the Synset Relation Ontology(as shown in the dotted line in Fig. 1) in the Candidate Generator module. GivenIM(E1)andIM(E2), we reason the relation between them to form an original candidate if there is any. This step produces a group of original candidates.

However, not all candidates are accurate,maybe some of them are redundant. We filter out such candidates and keep more relevant ones in the Candidate Filter module. The filtered candidate set is denoted asMCF.

3.2 Semantic analysis

Axioms and constructors are used to restrict an entity further when engineers define annotations for this entity, and they reflect the semantics which the engineer need.

ALOWS analyzes semantics in a sensor ontology by two steps based on structural subsumption reasoning algorithms. Firstly, it parses axioms and constructors, and rewrites entities into normal forms (as shown in Formula 3) to make the implied semantics to be read-off easily (see [6], [7], [10], [11], [21],[22] for more details). Secondly, it compares the syntactical structure of the derived normal forms of entities to infer association links.

Entities from different ontologies are created by various knowledge engineers, and such engineers sometimes ignore some elements or give different ranges by using restrictions when defining entities, so the difference between entities from various ontologies often exists. We import three thresholdsα,βandγto reflect the allowable degree of discrepancy between entities when comparing the syntactical structure of the normal forms of entities.

When displaying the implied semantic information into the normal form of an entity,the Comparing Normal Forms modules (including Concepts and Properties) compute relaitons between entities by comparing the syntactical structure of their normal forms derived above. The process in the Comparing Concept Normal Forms module is similar to the one in the Comparing Property Normal Forms module. We take as an example the Comparing Concept Normal Forms module. Given the conceptsCandDfrom different ontologies and their normal forms as shown below:

For alli,there existssuch thatWe callif the conditions below holdand

In Condition 1, when determining if the element fromis the subsumption of the element from(namely), the two elements, which are (negated) primitive concepts, are first replaced with the representation of their images and then the relations between them are calculated based onMCFwith the help of reasoners which are shown in the dotted line in Fig. 1. Such as, the relation between (negated) primitive conceptsAandBis replaced with the relation between ?IM(A) andIM(B) if it exists. We use reasoners over the knowledge base, which is composed ofMCF,to infer the relation between ?IM(A) andIM(B).

We importα,βandγto adjust the allowable degree of discrepancy between concepts from different ontologies. The thresholdα,which varies in [0, 1], reflects the degree of ignoring some elements inandThe greaterαis, the more elements inandare ignored. Whenα=0, it reflects that we do not consider relations between any two elements respectively fromandWhenα=1, it reflects that we consider relations between any two elements respectively fromandThe thresholdβreflects the similar case toα, butβis for properties. The thresholdγreflects the allowable degree of discrepancy between ranges w.r.t theat-leastandat-mostrestrictions. The smallerγis, the greater the discrepancy between ranges is. When, it means that discrepancy between ranges is not allowable.

According to the steps above for inferring subsumption between concepts, equivalence between them can be determined based on rules below:

The process for aligning properties in the Comparing Property Normal Forms module is similar to the one for aligning concepts. This section generates a set of association links.Table 1 partially shows association links when aligningSensor Ontology leftandSensor Ontology right.

The construction stage converts the derived association links from the previous stage as the property assertions of the SKOS mapping properties, and then utilizes the property assertions to establish the integration sensor ontology in the SKOS model. This process is described in Section 4.

IV. CONSTRUCTING THE INTEGRATION SENSOR ONTOLOGY IN THE SKOS MODEL

After identifying association links between entities from different ontologies, we establish an interconnection between ontologies, basedon a set of the SKOS mapping properties, to form the integration sensor ontology accessed via the IoT.

Table I A partial view of association links

There are three steps in the construction stage. Firstly, it establishesthe base ontologywhich includes the vocabularies (such asskos:Concept, skos:ConceptSchemeandskos:broadMatch) in the SKOS model as shown in Section 2. The base ontology also includes the concept assertions of theskos:ConceptSchemewhich represent the aligned ontologies as shown by Line 4-5 in Fig. 3. We use the SKOS Protégé plug-in2http://code.google.com/p/skoseditorand SKOS API3http://skosapi.sourceforge.netto establish the base ontology represented in the OWL-Full form as partially shown by Line 1-5 in Fig. 3.

Secondly, the construction stage converts the association links obtained from the previous stage as the property assertions of the SKOS properties as shown in Section 4.1 and then insert them into the base ontology to construct the integration sensor ontology in the SKOS model as shown in Section 4.2.

Finally, the construction stage expresses the SKOS integrity conditions by the OWL Full syntaxes to ensure the consistency and integrity of the integration sensor ontology in the SKOS model as shown in Section 4.3.

4.1 Converting association links

The Inserting Property Assertions of SKOS Mapping Properties module will insert the property assertions of the SKOS mapping properties, which reflects the association links derived from the previous stage, into the base ontology as the part of the integration sensor ontology. So we need to convert association links as the property assertions of the SKOS mapping properties.

Fig. 2 Mapping between the SKOS mapping properties and association relations

The SKOS mapping properties state that two concepts from different ontologies have comparable meanings, and specify that how these meanings compare. They includeskos:narrowMatch,skos:closeMatch, skos:exactMatch,skos:broadMatchandskos:related-Match. The relations in the association links(called association relations) between entities from two different ontologies, derived from the previous stage, have three types:equivalent,beIncludedandinclude.

Fig. 2 illustrates the hierarchical structure of the SKOS mapping properties, and the association relations defined in the alignment stage. It also defines mapping between the association relations and SKOS mapping properties. Entities from different ontologies are linked to each other through converting association links into the SKOS mapping properties by applying these mappings.

We map theincludeto theskos:broader-Match, but not theskos:broaderas shown in Fig. 2. The reason is that theskos:broaderis used to assert the direct hierarchical links between different SKOS concepts, and theskos:broaderMatchis used to state the mapping (i.e. alignment) links between SKOS concepts from different schemes [37]. Mapping thebeIncludeto theskos:narrowMatchalso follows the same reason.

To define similarity between two entities,theskos:exactMatchandskos:closeMatchare considered. According to the W3C SKOS specification [39], theskos:closeMatchis used to link two concepts that are similar enough and they can be used interchangeably in some information retrieval IOT service; theskos:exactMatchindicates that concepts with high reliability and confidence level can be used interchangeably across a broader range of information retrieval IoT service.

We select theskos:closeMatchandskos:exactlMachto specify an equality relation in association links by employing a threshold.If a similarity value between two entities in an equality association link is higher than the threshold, this equality association link is specified using theskos:exactMatch, OTHERWISE utilizing theskos:closeMatch.

By utilizing the filtered candidates, we can get the semantic information in a sensor ontology and string similarity by the similarity value computed between the two entitiesAandBexpressed asS(A, B). The purpose of getting sum of three kinds of similarity measures is: lexical similarity expressed asLS(A,B), semantic similarity expressed asSS(A,B),and string similarity expressed asSM(A,B).The formal definition is as follows.

At first, obtain all the parent concepts ofAandB. Then, obtain the properties ofAandA’s parent concepts (denoted asPS(A)), the same process is repeated forBand then the set of properties forBis expressed asPS(B).If two properties, one fromPS(A)and one fromPS(B), occur together as an association link, the counter adds 1. The process inspects any two properties separately fromPS(A)andPS(B)in the same way. At last, we can get the value of thess(A,B)by calculating the size ofPS(A)to the size ofPS(B) and then dividing the counter.

According to the mapping between the SKOS mapping properties and association relations, we convert association links (a four-tuple) as the property assertions of the SKOS mapping properties (a triple-tuple) as defined below:

＜C, D, SKOS mapping properties＞

The following list is the examples of converting association links as the property assertions of the SKOS mapping properties when aligningSensor Ontology leftandSensor Ontology right.

4.2 Inserting property assertions into the base ontology

The SKOS property assertions, derived from the previous step, are the basis for constructing the integration sensor ontology. The Inserting Property Assertions of SKOS Mapping Properties module in Fig. 1 converts every association link as the property assertion of the SKOS mapping properties, and then inserts them into the base ontology to construct the integration sensor ontology in SKOS model.

We take as an example Item 1 above to specify the process. We first define the entitiesleft: Bookandright: Bookas the concept assertions of theskos:Conceptin the base ontology and identify that they are respectively fromSensor Ontology leftandSensor Ontology right. We then define the property assertion of theskos:exactMatchbetweenleft: Bookandright: Book. The inserted property assertion for Item 1 above is shown in Fig. 3, which is the snapshot of the RDF file for the integration sensor ontology in the SKOS model when integratingSensor Ontology leftandSensor Ontology right. Line 1-3 show some vocabularies in the SKOS model. Line 4-5 show thatOntology 301andSensor Ontology rightis the concept assertions of theskos:ConceptSchemeand also show thatSensor Ontology rightis aligned withOntology left. Line 6-8 specify thatright: Bookis the concept from theSensor Ontology right. Line 9-11 specify thatleft:Bookis the concept from theOntology left.Line 12-15 mean the relation between two concepts.

Fig. 3 Snapshot of the integration sensor ontology in RDF format

Algorithm 1 Constructing the Integration sensor ontology in the SKOS model

The process for constructing the integration sensor ontology is shown in Algorithm 1. However, it is not all of property assertions are inserted into the integration sensor ontology. Line 6 checks if the integration sensor ontology is consistent and satisfies the integrity conditions in the SKOS model as shown in Section 4.3. We convert the selecting of property assertions as combination problems.Line 3 shows that the functionGetCombinationFromPA(i, SKOSPropertyAssertionSet)returns a set, each of which is consisted of theielements chosen in a set of the property assertions of the SKOS mapping properties derived from the step before. The algorithm promises that the most association links are adopted and the integration sensor ontology does not break the consistency and integrity.

4.3 Consistency and integrity

When the OWL Full is seen as a data modeling (i.e., SKOS) language, the concept of inconsistency is useful, therefore, we can determine whether some given data is appropriate for the given data model. When the data are out of accord with respect to the data model,that means the data does not fit. The SKOS data model uses the notion ofintegrityto express this view.These integrity conditions are included to promote interoperability by defining the circumstances under which data are not consistent with respect to the SKOS data model.

The integration ontology derived from the previous step should be consistent to ensure the data fit the SKOS data model, because the interoperability of a given class of IoT services hinges on data conforming to a common data model. The Checking Consistency and Integrity module in Fig. 1 implements this function by checking if the integration ontology is consistent and whether it satisfies the SKOS integrity conditions as shown by Line 6 in Algorithm 1, with the assistance of reasoners shown in the dotted line in Fig. 1.

The SKOS data model includes a finite number of statements that are intended as integrity conditions5www.w3.org/TR/skos-reference/#L1228as partially shown below.

? skos:exactMatch is disjoint with skos:broadMatch;

? skos:exactMatch is disjoint with skos:relatedMatch;

? skos:ConceptScheme is disjoint with skos:Concept;

The integration ontology in the SKOS model is represented in the OWL-Full form as shown in Fig. 3, so ALOWS expresses the integrity conditions defined in the SKOS model by the constructers and axioms of concepts and properties in the OWL-Full language. We take as an example the first SKOS integrity condition above to specify the process. We use the owl:propertyDisjointWith to express it,and a triple is defined as below,

? ＜ skos:exactMatch＞ owl:propertyDisjoint with ＜ skos:broadMatch＞

If we insert two SKOS properties assertions below into the integration ontology, reasoners will alert that the integration ontology is inconsistent.

? ＜A,C,skos:exactMatch ＞ and ＜A,C,skos:broadMatch ＞

V. EXPERIMENT

This section provides an evaluation of the implemented prototype, namely ALOWS. It is implemented by using OWL-API6http://owlapi.sourceforge.net, Pellet API7http://clarkparsia.com/pellet,Alignment API8http://alignapi.gforge.inria.fr, the MIT Java Wordnet Interface9http://projects.csail.mit.edu/jwi, Java WordNet Library API10http://sourceforge.net/apps/mediawiki/t, SKOSEd and SKOS API to process ontologies, compare normal forms, interact with WordNet and construct the integration ontology in the SKOS model.

5.1 Test dataset

We selected four real and in use ontologies as target ontologies for evaluations. They are taken from Folder 301 to 304 in the OAEI benchmark dataset in the domain of bibliography. We choose an ontology in Folder left as a source ontology. The reference alignment of each aligning task is also offered on the OAEI Web site.

In summary, each ontology in this dataset contains about 72 properties ,37 concepts and 108 axioms.

5.2 Evaluation metrics

Precision, F-Measure, fallout, overall and recall are usually used to evaluating derived alignments.

We assume thatArepresents the size of association links in a alignment reference resources,Crepresents the size of association links in a found alignment, otherBin overlapping area characterize the size of association links in a discovered and accurate alignment.Precision(Prec.) is defined asPrecision=B/Cand is the percentage of correctly found association links among total found association links. It is in the rang of 0 and1. The larger the value of Prec. is, the smaller set of found and wrong association links there will be (i.e. false positives).Fallout(Fall.) is defined asFallout=(C-B)/Cand is the percentage of wrongly found association links among total found association links. It is clear thatFall.+Prec. =1.Recall(Rec.) is defined asRecall=B/Awhich is the percentage of accurately found association links among a reference association links.It is in the rang of 0 and1. With the increace of the value of Rec., it is not easy to find set of right and not found association links (i.e.true positives).Overall(Over.) is defined asOverall(P)=Recall(P)× (2-(1/Precision(P))).F-Measure(Fmeas.) is defined asF?Measure=2×Prec.×Rec./ (Prec+Rec). Both reflect the combination of Rec. and Prec. Fmeas.represents the weighted harmonic mean of Prec. and Rec. when they are evenly weighted.Over. focuses more on difference betweenBandCwhen Prec. is greater than 0.5. Under normal conditions, we usually adopt harmonic mean (H-mean for short) to deal with a group of test data. The standard formula for H-mean is:H-mean=n/ (1/A1+1/A2…+1/An)whereAirepresents the related value of ontology alignment taski.

5.3 Evaluation results

We evaluate ALOWS against thirtheen ontology alignment methods, that is aroma, DSSim,RiMOM, Lily, ASMOV, MapPSO, GeRoMe,SOBOM aflood, Kosimap, TaxoMap, amaker,as well as a simple edit distance algorithm on labels called edna, which have participated in the OAEI [38].

It is not difficult to find that ALOWS performs better over the four real ontologies than other methods as shown in Fig. 4.

Fig. 4 shows Rec., Rrec., Fmeas., Fall. and Over. for all fourtheen methods on average.As we can seen from the column chart, the improvement of ALOWS in Fmeas. is of 28%over edna, 6% over Lily, 10% over DSSim,7% over aroma, 7% over ASMOV, 19% over SOMBOM, 7% over RiMOM, 3% over aflood,51% over MapPSO, 3% over amaker, 24%over GeRoMe, 29% over kosimap and 44%over TaxoMap as shown by the organge trend line. The improvement of ALOWS in Rec. is respectively of 7%, 8%, 22%, 11%, 7%, 34%,7%, 8%, 60%, 10%, 29%, 49%, 68%. The improvement of ALOWS in Over. is respectvely of 86%, 10%, 12%, 11%, 13%, 25%, 12%,3%, 71%, 3%, 43%, 45%, 53%. The over.of edna is the worst (-0.11), this is because its precision is lower than 0.5. However, it is clear that ALOWS is in the top group.

Fig. 4 also illustrates that both DSSim and SOBOM have high Prec., but lower Rec., for they produce about 50 association links only at a time, and others produce around 70 ones; but edna is opposite, it has high Rec., low Prec.,because it finds too many association links which include non-accurate ones. The Rec. of ALOWS is not the best one, but its Fmeas.,which reflects the combination of Prec. and Rec., is better than the ones of other methods.

In Fig. 7, the shorter the distance from the black point for a reference alignment (called refalign) to the one for some method is, the better its accuracy is. It is clearly seen that the distance for ALOWS represented as the blue point is the shortest; edna, GeRoMe, kosimap and TaxoMap are longest; the ones for Lily,RiMOM, aroma and amaker are close to each other, because they have similar accuracy.

In contrast to the methods which utilize reasoners to obtain semantics in an ontology, ALOWS extends structural subsumption reasoning algorithms to enable the implied semantics to be read-off easily. This character leads to exploit semantics more explicitly and deeply. So ALOWS has better accuracy and produces some association links which do not often appear in other methods. We take as an example a partial of association links as shown below when aligningSensor Ontology leftandSensor Ontology right, which do not often appear in other methods, to specify the benefit from such character.

1.＜left:Manual, right:Publication, be-Included＞;

2.＜left: Academic, right: Publication,beIncluded ＞;

3.＜left: Deliverable, right: Publication,beIncluded ＞;

Fig. 4 Evaluation results

4.＜left: Informal, right: Publication,beIncluded ＞;

5.＜left: Report, right: Publication, be-Included ＞;

6.＜left: LectureNotes, right: Publication, ＞;

7.＜left: Collection, right: Book, beIncluded ＞;

8.＜left: Monograph, right: Book, beIncluded ＞;

Under the condition of classic information retrieval scenarios, recall is not as higher as better as it will lead to lower precision. The green line in Fig. 5 shows the difference for the methods between Prec. and Rec.. The blue and red lines respectively represent the trends of the Prec. and Rec. of the methods. It is seen that ALOWS, ASMOV and RiMOM is more balanced between Rec. and Prec.; DSSim,SOBOM and TaxoMap have bigger difference.

Alignments are cut under a threshold necessary for achievingn% recall and compute the corresponding precision [40] in Fig. 6. It responds a tradeoff between Rec. and Prec..When Rec. varies in [0%, 20%], ALOWS,edna, ASMOV, Lily, amaker, and RiMOM have better Prec.. When Rec. changes from 20% to 50%, the Prec. Of ASMOV can be better. At the time of Rec. varies in [50%, 60%],Lily and amaker have better Prec. When Rec.changes from 60% to 90%, ALOWS has better Prec. than others. When Rec. varies in [90%,100%], they are similar. In summary, the closer to the top line, which represents the reference alignment (named refalign), the curve for some solution is, the better the solution is. It can seen that the black doteed curve for ALOWS, the blue dashed one for ASMOV,the pink one for RiMOM, the olive one for aflood, and the aqua one for Lily are closer to the black one for refalign than others.

5.4 Ontology integration results

The construction stage uses the ontology alignment results to create the integration sensor ontology in the SKOS model. Fig. 8 demonstrates the part of the integration sensor ontology for the sample ontologies, including the SKOS mapping properties and association links between the concepts from different ontologies. The black left-right arrow line represents theskos:exactMatch. On the red line for theskos:narrowMatchorskos:broad-Match, the arrow directions point to related super concepts. The gray one means theskos:RelatedMatch.

Such alignment, conforming to a common data model, can be accessed via the IoT. For example, a information retrieval IoT service can found the conceptsleft: bookandright:-bookin the integration sensor ontology at the same time. It also knows that the two concepts can be consideration as the same to each other.So this IoT service can interoperate the data from the two different sources via the IoT.

Fig. 5 The difference between Prec. and Rec

Fig. 6 Precision/recall graphs

Fig. 7 Each point expresses the position of a system with regard to Prec. and Rec.

Fig. 8 The snapshot of the specified entities in the integration sensor ontology

5.5 Analysis and discussion

For showing the strength and weakness of ALOWS and how it works, we analyze and discuss the alignment for the sample ontologies. The Prec. and Rec. of ALOWS is respectively of 93% and of 90% when aligning them.There are still some generated association links by ALOWS, which do not come along in the reference alignment. There are also some association links in the reference alignment,which are not found by ALOWS.

The association links in the reference alignment are not discovered as shown below:

1. ＜left: Unpublished, right: Resource, be-Included＞;

2. ＜left: proceedings, right: booktitle,equivalent＞;

3. ＜left: URL, right: softCopyURI, equivalent＞;

4. ＜left: isPartOf, right: booktitle, equivalent ＞;

The association links are discoveried, but do not appear in the reference alignment as shown below:

5. ＜left:Chapter, right:Publication, beIncluded ＞;

6. ＜left:MotionPicture,right:Publicaiton,beIncluded＞;

7. ＜left:Unpublished,right:Publicaiton,be-Included＞;

Actually Item 1 is included in the original candidates, because of the filtering process, we remove it. But if no filtering is used, what we get will be association links not emerge in the reference alignmen.

For Item 2 and 4, the Tokenizer module does not have enough capability to process“booktitle”into the two words＜book, title＞.Also, because ALOWS tries to ensure consistency, we can only find some separate association which do not links together. For instance,＜left:proceedings, right:booktitle, equivalent＞,＜left:isPartOf, right:booktitle, equivalent＞and＜left:book, right:booktitle, equivalent＞cannot be found together, because this will lead to＜left:isPartOf≡left:proceedings≡left:book＞, and it is a conflict with the axiom inOntology left:＜left:isPartOf, left:proceedings, include＞.The Synset Finder module cannot find the “url” in WordNet, so Item 3 is missing.

The Synset Finder find the unsuitable synsets for ?unpublish? in WordNet, so we discovery Item 7. The Synset Extension module broadens the extended synset set too much,so ALOWS establishes the wrong relaiton for Item 6. ALOWS infers Item 5 which seems to be correct.

In addition, because we make use of lexical analysis, we can get ＜left: date, right: publishedOn, equivalent equivalent＞ which do not appear in most of other systems. Because of making use of semantic analysis, we can get＜left:Collection, right:Book, beIncluded＞and＜left:Monograph, right:Book, beIncluded＞which do not appear in most of other systems,either.

VI. CONCLUSION AND FUTURE WORK

This paper discusses that using lexical and semantic analysis infers association links between entities from different sensor ontologies, presents that utilizing the mapping rules and thresholds determines the SKOS mapping properties between entities, demonstrates the integration sensor ontology which can be accessed via the IoT, finally checks the consistence of the integration sensor ontology to ensure that it satisfies the integrity conditions in the SKOS model.

Future work will focus on utilizing individual information as another criterion for aligning ontologies. We also plan to apply the SKOS-based ontology integration method to ontology learning mechanisms. This will provide the seamless integration of automatically constructed data to existing knowledge representation and organization systems.

ACKNOWLEDGEMENT

Financially Supported by National Natural Science Foundation of China (No.61601039);financially supported by the State Key Research Development Program of China (Grant No. 2016YFC0801407); financially supported by the Natural Science Foundation of Beijing Information Science&Technology University(No.1625008); financially supported by the Opening Project of Beijing Key Laboratory of Internet Culture and Digital Dissemination Research (NO. ICDD201607); Open Foundation of State key Laboratory of Networking and Switching Technology (Beijing University of Posts and Telecommunications) (NO. SKLNST-2016-2-08); financially supported by the Importation and Development of High-Caliber Talents Project of Beijing Municipal Institutions (Grant No. CIT&TCD201504056).

[1] C. Perera, A. Zaslavsky, C. H. Liu, M. Compton, P.Christen and D. Georgakopoulos, Sensor Search Techniques for Sensing as a Service Architecture for the Internet of Things,IEEE Sensors Journal,vol. 14, no. 2, pp. 406-420, Feb. 2014.

[2] A. K. Akanbi, and M. Masinde, Towards Semantic Integration of Heterogeneous Sensor Data with Indigenous Knowledge for Drought Forecasting.Proceedings of the Doctoral Symposium of the 16th International Middleware Conference. ACM, 2016.

[3] A. Whitmore, A. Agarwal, and L. D. Xu, “The Internet of Things—A survey of topics and trends,”Information Systems Frontiers,vol. 17,no. 2, pp. 261-274, 2015.

[4] P. Shvaiko and J. Euzenat, Ontology Matching:State of the Art and Future Challenges,IEEE Transactions on Knowledge and Data Engineering, vol. 25, no. 1, pp. 158-176, Jan. 2013.

[5] Y. Kalfoglou, B. Hu, D. Reynolds, and N. Shadbolt, “Capturing, Representing and Operationalising Semantic Integration (CROSI) project -final report,”ECS, University of Southampton,Tech. Rep., October 2005.

[6] B. Hollunder, and W. Nutt, “Subsumption Algorithms for Concept Languages.”Deutsches Forschungszentrum fur Kunstliche Intelligenz GmbH(DFKI), Tech. Rep., 1990.

[7] S. Brandt, and A. Y. Turhan, “Approximating ALCN -Concept Descriptions.”Proc. of the 2002 Description Logic Workshop (DL 2002),2002.

[8] D. J??r?′me, G. Fabrice, and B. Henri, “Matching directories and OWL ontologies with AROMA,”Proceedings of the 15th ACM international conference on Information and knowledge management, pp. 830-831, 2006.

[9] M. Nagy, M. Vargas-Vera, and E. Motta, “DSSim-ontology mapping with uncertainty.”Proceedings of the 1st International Workshop on Ontology Matching (OM-2006), pp. 115–123,2006.

[10] N. Fanizzi, and C. D’Amato, “A Similarity Measure for the ALN Description Logic.”Proceedings of CILC 2006 – Italian conference on Computational Logic, pp. 26–27, 2006.

[11] C. D’Amato, N. Fanizzi, and F. Esposito, “A dissimilarity measure for ALC concept descriptions.”SAC, pp. 1695-1699. 2006.

[12] N. Choi, I. Y. Song, and H. Han, “A survey on ontology mapping,”Acm Sigmod Record,vol. 35,no. 3, pp. 34-41, 2006.

[13] P. Shvaiko, and J. Euzenat, “Ten Challenges for Ontology Matching.”Proceedings of the OTM 2008 Confederated International Conferences,CoopIS, DOA, GADA, IS, and ODBASE 2008. Part II on On the Move to Meaningful Internet Systems, ser. LNCS, no. 5332, pp. 300-313, 2008.

[14] P. Shvaiko, and J. M. Euzenat, “A Survey of Sche-ma-based Matching Approaches,”Lecture Notes in Computer Science,vol. 4, no. 1, pp. 146--171,2010.

[15] F. Giunchiglia, and P. Shvaiko, “Semantic matching,”Knowledge Engineering Review,vol. 18, no.3, pp. 265-280, September 2003.

[16] P. Maio, and N. Silva, “An extensible argument-based ontology matching negotiation approach,”Science of Computer Programming,vol. 95, no. 95, pp. 3-25, 2014.

[17] R. C. Chen, C. T. Bau, and C. J. Yeh, “Merging domain ontologies based on the WordNet system and Fuzzy Formal Concept Analysis techniques,”Applied Soft Computing,vol. 11, no. 2, pp. 1908-1923, 2011.

[18] X. Xue, Y. Wang, and A. Ren, “Optimizing ontology alignment through Memetic Algorithm based on Partial Reference Alignment,”Expert Systems with Applications,vol. 41, no. 7, pp.3213–3222, 2014.

[19] A. Algergawy, S. Massmann, and E. Rahm, “A clustering-based approach for large-scale ontology matching, ”Advances in Databases and Information Systems, ser. Lecture Notes in Computer Science, J. Eder, M. Bielikova, and A. Tjoa,Eds. Springer Berlin / Heidelberg, vol. 6909, pp.415–428, 2011.

[20] C. E. Yap, and M. H. Kim, “Instance-Based Ontology Matching with Rough Set Features Selection.”IT Convergence and Security (ICITCS), 2013 International Conference,pp. 1–4, Dec 2013.

[21] F.Baader, D. Calvanese, D. McGuinness, “The Description Logic Handbook: Theory, Implementation and Applications,”Cambridge University Press, ch. Basic Description Logics, pp.46–99, 2003.

[22] I. Horrocks, “Description logic reasoning”University of Manchester. 2005.

[23] P. Xu, H. Tao, T. Zang, and Y. Wang, “Alignment Results of SOBOM for OAEI 2009,”International Workshop on Ontology Matching, 2009.

[24] Daniel W S. Olmedilla. (2007, January) Introduction to the semantic Web. L3s Research Center,Hannover University. [Online]. Available: http://www.l3s.de/_olmedilla/events/2007/TENCWS/20070124 TENC WS.pdf.

[25] M. Houshmand, and E. Khorram, “Similarity aggregation in ontology matching based on reliability maximization.”Electrical Engineering(ICEE),pp. 1-6, May 2011.

[26] O. Gerbé, and B. Kerhervé, “A Model-Driven Approach to SKOS Implementation.”Internet and Web Applications and Services, International Conference, pp. 484-488.

[27] H. Mousselly-Sergieh, and R. Unland, “IROM:information retrieval-based ontology matching.”Semantic Multimedia, ser. Lecture Notes in Computer Science, T. Declerck, M. Granitzer, M.Grzegorzek, M. Romanelli, S. Ruger, and M. Sintek, Eds. Springer Berlin / Heidelberg, vol. 6725,pp. 127–142, 2011.

[28] G. Acampora, V. Loia, and A. Vitiello, “Enhancing ontology alignment through a memetic aggregation of similarity measures,”Information Sciences,vol. 250, no. 11, pp. 1-20, 2013.

[29] X. Ma, E. J. M. Carranza, C. Wu, F. D. V. D. Meer,and G. Liu, “A SKOS-based multilingual thesaurus of geological time scale for interoperability of online geological maps,”Computers & Geosciences,vol. 37, no. 10, pp. 1602-1615, 2011.

[30] G. Pirró, and D. Talia, “UFOme: An ontology mapping system with strategy prediction capabilities,”Data & Knowledge Engineering,vol. 69,no. 5, pp. 444-471, 2010.

[31] N. Jian, W. Hu, G. Cheng, “Falcon-ao: Aligning ontologies with falcon”,K-Cap 2005 Workshop on Integrating Ontologies., pp. 87–93, 2005.

[32] Y. Qu, W. Hu, and G. Cheng, “Constructing virtual documents for ontology matching,”International Conference on World Wide Web, WWW 2006, Edinburgh, Scotland, Uk, May, pp. 23-31,2006.

[33] Y. Kalfoglou, and M. Schorlemmer, “IF-Map: An Ontology-Mapping Method Based on Information-Flow Theory,” vol. 1, pp. 98-127, October 2003.

[34] W.Peng, “ Research on the key issues in ontology mapping,”Ph.D. dissertation, Dept. Electron.Eng., SouthEast Univ., NanJing, China, 2009.

[35] A. Buccella, A. Cechich, D. Gendarmi, F. Lanubile, G. Semeraro, and A. Colagrossi, “Building a global normalized ontology for integrating geographic data sources,”Computers & Geosciences,vol. 37, no. 7, pp. 893-916, 2011.

[36] A. Doan, J. Madhavan, P. Domingos, at al.. “Ontology matching: A machine learning approach,”Handbook on Ontologies in Information Systems.Springer, pp. 397–416, 2003.

[37] S. Khan, and M. Safyan, “Semantic matching in hierarchical ontologies,”Journal of King Saud University - Computer and Information Sciences,vol. 26, no. 3, pp. 247-257, 2014.

[38] M. Kolli, and Z. Boufaida, “A description logics formalization for the ontology matching,”Procedia Computer Science,vol. 3, no. 1, pp. 29-35,2011.

[39] A. Miles, and S. Bechhofer, “SKOS Simple Knowledge Organization System Reference,”World Wide Web Consortium, 2009.

[40] C. Caracciolo, J. Euzenat, L.Hollink, “esults of the ontology alignment evaluation initiative 2009”[Online]. Available: http://oaei.ontologymatching.org/2009/

[41] M. Haspelmath, “The indeterminacy of word segmentation and the nature of morphology and syntax”Folia Linguistica, vol. 45, no. 1,pp.31–80, Mar. 2011.

[42] S. Albagli, R. Ben-Eliyahu-Zohary, and S. E. Shimony, “Markov network based ontology matching.”Journal of Computer and System Sciences,pp. 1884-1889, 2011.

[43] Y. R. Jean-Mary, E. P. Shironoshita, and M. R.Kabuka, “Ontology matching with semantic verification,”Web Semantics Science Services &Agents on the World Wide Web,vol. 7, no. 3, pp.235-251, 2009.

[44] J. Tang, J. Li, B. Liang, X. Huang, Y. Li, and K.Wang, “Using Bayesian decision for ontology mapping,”Web Semantics Science Services &Agents on the World Wide Web,vol. 4, no. 4, pp.243-262, 2006.

[45] J. Li, J. Tang, Y. Li and Q. Luo, RiMOM: A Dynamic Multistrategy Ontology Alignment Framework, inIEEE Transactions on Knowledge and Data Engineering, vol. 21, no. 8, pp. 1218-1232,Aug. 2009.

[46] O. Udrea, L. Getoor, R. Miller, and e. J, “Leveraging data and structure in ontology integration.”Proceedings of the 2007 ACM SIGMOD international conference on Management of data. New York, NY, USA: ACM, pp. 449-460, 2007.

[47] B. M, D. A, M. A, and S. L, “Context matching for electronic marketplaces: a case study,”Knowledge Engineering Review,vol. 18, no. 4, pp. 317-328, 2003.

[48] Q. Reul, and J. Z. Pan, “KOSIMap: Use of Description Logic Reasoning to Align Heterogeneous Ontologies.”Proceedings of the 23rd International Workshop on Description Logics, ser.CEUR Workshop,vol. 573. 2010.

[49] Y. Jiang, X. Wang, and H. T. Zheng, “A semantic similarity measure based on information distance for ontology alignment,”Information Sciences,vol. 278, no. 10, pp. 76-87, 2014.

[50] X. Liu, “Research on the identity-related services and key technology application in internet of things,”Metallurgical & Mining Industry, 2015.

[51] Zhou Ming and Ma Yan Institute of Network Technology Beijing University of Posts and Telecommunications Beijing, China. Study on the Ontology-base Context-aware and Reasoning Model of IOT[A]. IEEE Beijing Section、Beijing University of Technology. Proceedings of 2012 IEEE 3rd International Conference on Emergency Management and Management Sciences(ICEMMS2012)[C]. IEEE Beijing Section, Beijing University of Technology, 2012:5.

[52] M. Tao, K. Ota, and M. Dong, “Ontology-based data semantic management and application in IoT- and cloud-enabled smart homes,”Future Generation Computer Systems, 2016.

[53] W. Gao, Y. Gao, and L. Zhu, “Ranking based ontology learning algorithm for similarity measuring and ontology mapping using representation theory,”Journal of Information & Optimization Sciences,vol. 37, no. 2, pp. 303-320, 2016.

[54] M. Ehrig, and Y. Sure, “Ontology Mapping – An Integrated Approach,”Lecture Notes in Computer Science, vol. 3053, pp. 76-91, 2004.

[55] R. Idoudi, K. S. Ettabaa, K. Hamrouni and B. Solaiman, An evidence based approach for multipe similarity measures combining for ontology mapping,International Image Processing, Applications and Systems Conference, Sfax, pp. 1-6,2014.

[56] H. Wang, C. C. Tan and Q. Li, Snoogle: A Search Engine for Pervasive Environments, inIEEE Transactions on Parallel and Distributed Systems,vol. 21, no. 8, pp. 1188-1202, Aug. 2010.

[57] C. Frank, P. Bolliger, F. Mattern, and W. Kellerer,“The sensor internet at work: Locating everyday items using mobile phones,”Pervasive & Mobile Computing,vol. 4, no. 3, pp. 421-447, 2008.

[58] K. Aberer, M. Hauswirth, and A. Salehi, “Infrastructure for Data Processing in Large-Scale Interconnected Sensor Networks.”Proc Int Conf Mobile Data Manage, Mannheim, Germany, pp.198-205, 2007.

[59] A. Kansal, S. Nath, J. Liu and F. Zhao, “SenseWeb:An Infrastructure for Shared Sensing,” inIEEE MultiMedia, vol. 14, no. 4, pp. 8-13, Oct.-Dec.2007.

[60] M. Ehrig, S. Staab, and Y. Sure, “Bootstrapping Ontology Alignment Methods with APFEL.” pp.186-200.

[61] W. Wong, W. Liu, M. Bennamoun, “Ontology Learning and Knowledge Discovery Using the Web: Challenges and Recent Advances[J]” 2011.