The user expectations on the quality of results of web information searches are becoming more and more high. Very often, for business or personal needs, users require to retrieve, in a very fast way, all the available relevant information about a focused target entity, in order to take decisions, organize business work, plan future actions. A target entity is a keyword-based representation of a topic of interest, namely a real-world object/person, an event, a situation, or any similar subject that can be of interest for the user. To answer this kind of "entity"-driven user needs, a huge multiplicity of web resources is actually available, coming from the Social Web and related user-centered services (e.g., news publishing, social networks, microblogging systems), from the Semantic Web and related ontologies and knowledge repositories, and from the conventional Web of Documents. Each kind of web resource is differently structured according to a variety of formats, ranging from short, unstructured, and ready-to-consume news/posts, to well-structured, formal ontology, and each one can provide unique information for a given target entity. For example, only web resources coming from the Social Web are able to provide subjective information reflecting users opinions or preferences about Ph.D. thesis supervisor: Prof. Silvana Castano -Università degli Studi di Milano. the target entity, which complement in a useful way the more objective information provided by web resources coming from the other webs. To satisfy user expectations, a new generation of web information search techniques has to cope with different requirements: i) the capability to span across multiple webs, to properly consider the wide variety of available web resources and pieces of knowledge by properly assessing their information contribution nature; ii) the capability to anticipate the user needs by providing a focused but comprehensive set of web resources relevant for the target entity; iii) the capability to semantically organize all retrieved web resources into an intuitive and coherent structure for the given target entity.

With respect to this scenario, the Ph.D. thesis is devoted to define the notion of i-cloud and a semantic clouding approach for the construction of i-clouds that works over the Social Web, the Semantic Web, and the Web of Documents. i-clouds are built for a target entity of interest to organize all relevant web resources, modeled as web data items, into a graph, on the basis of their level of prominence and reciprocal closeness. Prominence captures the importance of a web resource within the i-cloud, by distinguishing, also in a visual way "a la tag-cloud", how much relevant web resources are with respect to the target entity. The level of closeness between web resources is evaluated using matching and clustering techniques, with the goal of determining how similar web resources are to each other and with respect to the target entity.

The research methodology followed for the Ph.D. activity is based on the following main phases: i) literature review with the aim of providing a critical comparison of the state of the art solutions for semantic data clouding, ii) conceptual design where requirements and foundational aspects related to the Ph.D. issues are formally addressed, iii) prototype implementation where a prototype tool is developed according to the defined architecture, and iv) evaluation of the proposed techniques on a number of real test cases.

Relevant research work with respect to the Ph.D. thesis regards Linked Data, instance matching, and data clouds. Linked Data. A new generation of web applications for the integration of both data and services is being emerging in the context of the Linked Data project [2]. Linked Data is mainly focused on the idea of improving interoperability and aggregation among large data collections already available on the web, such as for example DBLP1 , DBPedia2 , CiteSeer3 , IMDB 4 , and Freebase 5 , which are available as retrievable RDF datasets or SPARQL query endpoints. Linked Data is a step beyond the simple availability of data and syntactic compatibility, in that it promotes some important principles in making web data available and sharable to the Semantic Web community. Such principles are the following: i) all the web resources have to be referenced by a URI; ii) URIs have to be resolvable on the web to RDF descriptions; iii) RDF triples have to be consumed by a new generation of Semantic Web browsers and crawlers [15]. However, Linked Data does not take into account the web resources originated from user-generated contents like comments, posts and personal feeds, that are characterized by poor structure and rapid obsolescence. Moreover, Linked Data builds a flat graph structure of interconnected URIs, without distinguishing the prominence and closeness of web resources.

Instance matching. The same real-world object can be described multiple times in different knowledge repositories, possibly using different perspectives and by emphasizing different properties of interest. The capability of finding similar object descriptions assumes particular relevance in the field of Semantic Web, to promote effective web resource sharing on the global scale and to correctly interoperate/reuse individual knowledge chunks coming from disparate information repositories, disregarding their specific URIs. Such task is called instance matching, and consists in finding instances (i.e., object descriptions), coming from different sources, which describe the same realworld object in a different and heterogeneous way. Some contributions in this direction have been focused on defining techniques and approaches for the generation and management of identifiers at object-level, like, for example, the OKKAM project [3]. Other approaches have been proposed for the unification of different URIs associated to the same object [13]. Moreover, a problem related to instance matching is the one of finding object descriptions referring to similar objects. To this end, suitable matching techniques are required. Such techniques are mainly provided by the research work in the field of record linkage, which has been widely studied in the databases community [8]. More recently, some new techniques have been proposed to specifically match ontology instances [9] and to identify similar web resources [11]. However, none of the proposed approaches is able to compare different kinds of object.

Data clouds. In the recent years, the traditional World Wide Web based on "userconsuming" applications and informative web pages has changed into a more complex vision composed of a plurality of webs, where semantic-intensive applications as well as interactive "user-generated" platforms like microblogging, and news feeds are becoming more and more popular. In this scenario, the research efforts towards the development of solutions for organizing this huge amount of web resources according to semantic clouding or similar approaches is still at an initial stage [10]. Some interesting work has been done in the field of news aggregation, with the aim of providing techniques for their semantic organization and classification. Examples of proposed systems are NewsInEssence [14] and Relevant News [1], which automatically group news related to the same topic by exploiting hierarchical clustering algorithms and tag/keyword-based search functionalities. For what concerns microdata sources, like Twitter or Facebook, tools for semantic aggregation are still missing. In the same direction, structured and collaborative search engines are being emerging as a promising solution for presenting the query results in a sort of structured form. Examples in this field are Wolfram Alpha 6 and Google Wonder Wheel 7 . In particular, Wolfram Alpha is a computational knowledge engine based on data extraction from popular knowledge repositories, like Wikipedia. The goal of this engine is to provide answers to the user requests by returning a comprehensive picture of the available data retrieved about the given request. The same idea is enforced by Google Wonder Wheel, which provides also a graphical, cloud-oriented view of the query results based on terminological similarities among different web resources. However, all these proposed solutions still lack the integration between Social and Semantic Web resources, and provide a poor support of semantic matching techniques for identifying similar web resources.

With respect to the state of the art, the contributions of the Ph.D. thesis are mainly the following.

-Definition of a cross-web approach considering the different kinds of available web resources (e.g., tagged resources, microdata resources, Semantic Web resources), and considering both objective and subjective information. As far as we know, our semantic clouding approach represents a first attempt to bridge the gap between Semantic Web resources (typically managed in Linked Data) and other kinds of web resource, such as, for example, tagged and microdata resources. -Definition of i-cloud as a new data structure for organizing relevant web resources for a given target entity on the basis of their prominence and closeness. -Definition of matching techniques for comparing different kinds of web resources.

In particular, in Table 1, the differences between Linked Data and i-clouds are summarized.

In Figure 1, we show the semantic clouding approach developed for i-cloud construction. The approach is articulated in three phases: i) modeling of web resources, ii) classification of web resources, and iii) clouding of web resources. Classification of web resources. Our semantic clouding approach is based on the capability of grouping the web data items on the basis of their closeness. The closeness between two web data items wdi i and wdi j captures the level of similarity/semantic relation holding between them and it is represented by a closeness coefficient cc(wdi i , wdi j ) ∈ [0, 1], calculated by comparing wdi i and wdi j . Such closeness coefficient cc(wdi i , wdi j ) is calculated for each possible pair of web data items stored in the WDI repository using appropriate matching techniques, and the corresponding values are then used by a hierarchical clustering procedure in order to produce a closeness tree where each leaf corresponds to a web data item, and inner nodes denote the closeness coefficient values. To choose the matching techniques to use, we take into account the nature and the different complexity that can characterize the different web resources, and consequently, the corresponding web data items. In [4], we address the problem of matching Semantic Web resources; in [6], we analyze the problem of classifying and comparing microdata; in [7], we provide specific methods and techniques for organizing and matching tags extracted from the Social Web. Moreover, in [5,12], we present a system for integrating Social and Semantic knowledge in a P2P environment.

Clouding of web resources. The clouding phase is based on the results of the classification activity and aims at constructing the appropriate i-cloud organization for a given target entity by prominence and closeness levels. An i-cloud is formally defined as an undirected weighted graph IC e = (N, E) associated with a target entity e. A node n i ∈ N represents a web data item wdi i relevant for e, while an edge (n i , n j ) ∈ E between two nodes n i and n j represents the level of closeness between wdi i and wdi j . IC e is equipped with a labeling function ρ : N → [0, 1], that associates each node n i ∈ N with a value p(n i ) ∈ [0, 1], and a labeling function σ : E → [0, 1], that associates each edge (n i , n j ) ∈ E with a value c(n i , n j ) ∈ [0, 1]. A value p(n i ) denotes the level of prominence of the web data item wdi i in IC e . A high value of p(n i ) denotes that the web resource corresponding to wdi i is very relevant for e. Different techniques are possible for the evaluation of the prominence in an i-cloud and these techniques can be used alone or in combination. We devise three main categories of techniques for prominence evaluation, namely provenance-base, target-based, and popularity-based techniques. A value c(n i , n j ) denotes the level of closeness between the web data items wdi i and wdi j in IC e . In particular, c(n i , n j ) is equal to the closeness coefficient cc(wdi i , wdi j ) calculated in the previous phase.

An example of i-cloud is shown in Figure 2, collecting web resources related to the target entity "Star Wars". We can observe that web resources in the i-cloud are not only those directly related to this popular movie, such as the titles of the six movies of the Star Wars saga, but also resources that are close to the movie saga even if not directly matching the target, such as some of the most important characters in the movies. The dimension of each node in the i-cloud is proportional to the prominence of the corresponding web resource for "Star Wars" and the edges connecting the nodes are labeled with their closeness degree.

We have presented the thesis work we are undergoing for semantic data clouding. Ongoing and future work will be devoted to formally define the properties of i-clouds and the operations that can be applied between different i-clouds (e.g., selection, projection, join). Furthermore, some preliminary evaluation of our semantic clouding approach has been performed using data extracted from Delicious8 , Twitter9 , and Freebase10 . i-clouds are evaluated on the basis of their level of accuracy and by analyzing the dependency between their size (i.e., the number of web data items) and their cohesion (i.e., the average level of closeness between web data items). The accuracy of an i-cloud is defined as its capability to collect web resources which are really relevant with respect to the given target entity, and it depends on the matching techniques that are used for clustering web data items. In order to evaluate the quality of our matching techniques, we exploited the IIMB 2010 dataset 11 and related tools, that are used also for the international instance matching evaluation contest of the Ontology Alignment Evaluation Initiative (OAEI) 12 . The obtained results show that the accuracy of our matching tool HMatch 2.0 is significantly higher than the one of a simple string matching algorithm. The effective applicability of the semantic clouding approach in real application contexts and how it is affected by the number of web data items stored in the WDI repository is also under study.