ABSTAT: Ontology-driven Linked Data Summaries with Pattern Minimalization Blerina Spahiu 0 Riccardo Porrini 0 Matteo Palmonari 0 Anisa Rula 0 Andrea Maurino 0 University of Milano-Bicocca

An increasing number of research and industrial initiatives have focused on publishing Linked Open Data, but little attention has been provided to help consumers to better understand existing data sets. In this paper we discuss how an ontology-driven data abstraction model supports the extraction and the representation of summaries of linked data sets. The proposed summarization model is the backbone of the ABSTAT framework, that aims at helping users understanding big and complex linked data sets. Our framework is evaluated by showing that it is capable of unveiling information that is not explicitly represented in underspeci ed ontologies and that is valuable to users, e.g., helping them in the formulation of SPARQL queries.

 data summarization knowledge patterns linked data Introduction

As of April 2014, 1014 data sets have been published in the Linked Open Data cloud, a number that is constantly increasing1. However, understanding the content of large and complex data sets is very challenging for users [ 10, 21, 6, 18, 14 ]. If a user wants to evaluate if a data set is useful for her or to formulate some queries, she needs rst to understand the content of the data set and its organization, by nding answers to questions such as: what types of resources are described in the data set? What properties are used to describe the resources? What types of resources are linked by a certain property and how frequently? How many resources have a certain type and how frequent is the use of a given property? Remarkably, di culties in answering those questions result in low adoption of many valuable but unknown data sets [ 16 ].

Linked data sets make use of ontologies to describe the semantics of their data. Ontologies may be large and underspeci ed. At the time of writing DBpedia uses 685 concepts and 2795 properties while the domain and range is not speci ed for 259 and 187 properties respectively.

Finally, the ontology does not tell how frequently a certain modelling pattern occurs in a data set. The above questions can be answered with explorative 1 http://linkeddatacatalog.dws.informatik.uni-mannheim.de/state/ queries, but at the price of a signi cant server overload for data publishers and high response time for data consumers.

ABSTAT is an ontology-driven linked data summarization model proposed to mitigate the data set understanding problem. In our view, a summary is aimed at providing a compact but complete representation of a data set. With complete representation we refer to the fact that every relation between concepts that is not in the summary can be inferred. One distinguishing feature of ABSTAT is to adopt a minimalization mechanism based on minimal type patterns. A minimal type pattern is a triple (C; P; D) that represents the occurrences of assertions <a,P,b> in RDF data, such that C is a minimal type of the subject a and D is a minimal type of the object b. Minimalization is based on a subtype graph introduced to represent the data ontology. By considering patterns that are based on minimal types we are able to exclude several redundant patterns from the summary. The ABSTAT2 framework supports users to query (via SPARQL), to search and to navigate the summaries through web interfaces. Other related work on data or ontology summarization have focused on complementary aspects of the summarization, such as the identi cation of salient subsets of knowledge bases using di erent criteria [ 21, 6, 18, 14 ], e.g., connectivity. Other approaches do not represent connections between instance types as our model does [ 7, 1, 8 ].

In this paper we make the following contributions: (i) we describe in detail the summarization model, focusing on the minimalization approach; (ii) we describe the summary extraction work ow; (iii) we provide an experimental evaluation of our approach from two di erent perspectives, evaluating the compactness and the informativeness of the summaries.

The paper is organized as follows. The summarization model is presented in Section 2. The implementation of the model in ABSTAT is given in Section 3. Experimental results are presented in Section 4. Related work is discussed in Section 5 while conclusions end the paper in Section 6. 2

 Summarization Model

We de ne a data set as a couple = (T ; A), where T is a set of terminological axioms, and A is a set of assertions. The domain vocabulary of a data set contains a set NC of types, where with type we refer to either a named class or a datatype, a set NP of named properties, a set of named individuals (resource identi ers) NI and a set of literals L. In this paper we use symbols like C, C0, ..., and D, D0, ..., to denote types, symbols P , Q to denote properties, and symbols a,b to denote named individuals or literals.

Assertions in A are of two kinds: typing assertions of form C(a), and relational assertions of form P (a; b), where a is a named individual and b is either a named individual or a literal. We denote the sets of typing and relational assertions by AC and AP respectively. Assertions can be extracted directly from RDF data (even in absence of an input terminology). Typing assertions occur in 2 http://abstat.disco.unimib.it a data set as RDF triples < x; rdf:type; C > where x and C are URIs, or can be derived from triples < x; P; y^^C > where y is a literal (in this case y is a typed literal), with C being its datatype. Without loss of generality, we say that x is an instance of a type C, denoted by C(x), either x is a named individual or x is a typed literal. Every resource identi er that has no type is considered to be of type owl:Thing and every literal that has no type is considered to be of type rdfs:Literal. Observe that a literal occurring in a triple can have at most one type and at most one type assertion can be extracted for each triple. Conversely, an instance can be the subject of several typing assertions. A relational assertion P (x; y) is any triple < x; P; y > such that P 6= Q , where Q is either rdf:type, or one of the properties used to model a terminology (e.g. rdfs:subClassOf).

Abstract Knowledge Patterns (AKPs) are abstract representations of Knowledge Patterns, i.e., constraints over a piece of domain knowledge de ned by axioms of a logical language, in the vein of Ontology Design Patterns [ 17 ]. For sake of clarity, we will use the term pattern to refer to an AKP in the rest of the paper. A pattern is a triple (C; P; D) such that C and D are types and P is a property. Intuitively, an AKP states that there are instances of type C that are linked to instances of a type D by a property P . In ABSTAT we represent a set of AKP occurring in the data set, which pro les the usage of the terminology. However, instead of representing every AKP occurring in the data set, ABSTAT summaries include only a base of minimal type patterns, i.e., a subset of the patterns such that every other pattern can be derived using a subtype graph. In the following we better de ne these concepts and the ABSTAT principles.

Pattern Occurrence. A pattern (C; P; D) occurs in a set of assertions A i there exist some instances x and y such that fC(x); D(y); P (x; y)g A. Patterns will be also denoted by the symbol .

For data sets that publish the transitive closure of type inference (e.g., DBpedia), the set of all patterns occurring in an assertion set may be very large and include several redundant patterns. To reduce the number of patterns we use the observation that many patterns can be derived from other patterns if we use a Subtype Graph that represents types and their subtypes.

Subtype Graph. A subtype graph is a graph G = (NC ; ), where NC is a set of type names (either concept or datatype names) and is a relation over NC .

We always include two type names in NC , namely owl:Thing and rdfs:Literal, such that every concept is subtype of owl:Thing and every datatype is subtype of rdfs:Literal. One type can be subtype of none, one or more than one type.

Minimal Type Pattern. A pattern (C; P; D) is a minimal type pattern for a relational assertion P (a; b) 2 A and a terminology graph G i (C; P; D) occurs in A and there does not exist a type C0 such that C0(a) 2 A and C0 G C or a type D0 such that D0(b) 2 A and D0 G D.

Minimal Type Pattern Base. A minimal type pattern base for a set of assertions A under a subtype graph G is a set of patterns b A;G such that 2 b A;G i is a minimal type pattern for some relation assertion in A. = types = named individuals

.

A subclassOf subclassOf

= literals

B subclassOf

E type type

C c type

F type type subclassOf

D type a P b

type R

Q “s” type T

Patterns (E, Q, F) (A, Q, D) (C, P, A) (C, Q, A) (B, Q, D) (C, P, F) (C, Q, F) (B, Q, A) (A, P, A) (E, Q, D) (B, Q, F) (A, P, F) (C, Q, D) (A, Q, A) (B, R, T) (A, P, D) (E, Q, A) (A, Q, F) (A, R, T) (C, P, D) (E, Q, D) (C, Q, D) (E, R, T)

(C, R, T) Minimal Type Pattern Base

Observe that di erent minimal type patterns (C; P; D) can be de ned for an assertion P (a; b) if a and/or b have more than one minimal type. However, the minimal type pattern base excludes many patterns that can be inferred following the subtype relations and that are not minimal type for any assertion. In the graph represented in Figure 1 considering the assertion set A = fP (a; b); C(a); A(a); F (b); D(b); A(b)g, there are six patterns occurring in A, i.e., (C; P; D), (C; P; F ), (C; P; A), (A; P; D), (A; P; F ), (A; P; A). The minimal type pattern base for the data set includes the patterns (E; Q; D), (E; R; T ), (C; Q; D), (C; R; T ) and (C; P; D) since E and C are minimal types of the instance c, while excluding patterns like (B; Q; D) or even (A; Q; A) since not B nor A are minimal types of any instance.

Data Summary. A summary of a data set = (A; T ) is a triple A;T = (G; ; S) such that: G is Subtype Graph, b A;G is a Minimal Type Pattern Base for A under G, and S is a set of statistics about the elements of G and .

Statistics describe the occurrences of types, properties and patterns. They show how many instances have C as minimal type, how many relational assertions use a property P and how many instances that have C as minimal type are linked to instances that have D as minimal type by a property P . 3

 Summary Extraction

Our summarization process, depicted in Figure 2, takes in input an assertion set A and a terminology T and produces a summary A;T . First, the typing assertion set AC is isolated from the relational assertion set AP , while the subtype graph G is extracted from T . Then, AC is processed and the set of minimal types for each named individual is computed. Finally, AP is processed in order to compute the minimal type patterns that will form the minimal pattern base b A;G. During each phase we keep track of the occurrence of types, properties and patterns, which will be included as statistics in the summary.

Summary Extraction. The subtype graph G is extracted by traversing all the subproperty and subtype relations in T . The subtype graph will be further enriched with types from external ontologies asserted in AC while we compute minimal types of named individuals (i.e., external types).

Given a named individual x, we compute the set Mx of minimal types with respect to G. We rst select all the typing assertions C(x) 2 AC and form the set AxC of typing assertions about x. We then iteratively process AxC . At each iteration we select a type C and remove from Mx all the supertypes of C according to G. Then, if Mx does not contain any C0 such that C0 G C, we add C to Mx. Notice that one preliminary step of the algorithm is to include C in G if it was not included during the subtype graph extraction phase. If a type C is not de ned in the input terminology, is automatically considered as a minimal type for the individual x. This approach allows us to handle the types of individuals that are not included in the original terminology.

For each relational assertion P (x; y) 2 AP , we get the minimal types sets Mx and My. For all C; D 2 Mx; My we add a pattern (C; P; D) to the minimal type pattern base. If y is a literal value we consider its explicit type if present, rdfs:Literal otherwise.

Summary Storage and Presentation. Every summary is stored, indexed and made accessible through two user interfaces, i.e., ABSTATBrowse and ABSTATSearch, and a SPARQL endpoint. SPARQL based access and ABSTATBrowse3 are described in our previous demo paper [ 11 ]. ABSTATSearch4 is a novel interface that implements a full-text search functionality over a set of summaries. Types, properties and patterns are represented by means of their local names (e.g., Person, birthPlace or Person birthPlace Country), conveniently tokenized, stemmed and indexed, and retrieved using Lucene Score as ranking model. 3 http://abstat.disco.unimib.it and http://abstat.disco.unimib.it/sparql 4 http://abstat.disco.unimib.it/search We evaluate our summaries from di erent, orthogonal perspectives. We measure the compactness of ABSTAT summaries and compare the number of their patterns to the number of patterns extracted by Loupe [ 10 ], an approach similar to ours that does not use minimalization. The informativeness of our summaries are evaluated with two experiments. In the rst experiment we show that our summaries provide useful insights about the semantics of properties, based on their usage within a data set. In the second experiment, we conduct a preliminary user study to evaluate if the exploration of the summaries can help users in query formulation tasks. In our evaluation we use the summaries extracted from three linked data sets: DBpedia Core 2014 (db2014-core)5, DBpedia 3.9 (db3.9-infobox)6 and Linked Brainz (lb). db2014-core and db3.9-infobox data sets are based on the DBpedia ontology while the lb data set is based on the Music Ontology. DBpedia and LinkedBrainz have complementary features and contain real and large data. For this reason they have been used, for example, in the evaluation of QA systems [ 9 ]. 5 The DBpedia 2014 version with mapping based property only 6 The DBpedia Core 3.9 version plus automatically extracted properties Insights about the semantics of the properties. Our summaries convey valuable information on the semantics of properties for which the terminology does not provide any domain and/or range restrictions. Table 2 provides an overview of the total number of unspeci ed properties from the data sets. For example, around 18% of properties from db2014-core data set have no domain restrictions while 13% have no range restrictions. Observe that this data set is the most curated subset of DBpedia as it includes only triples generated by user validated mappings to Wikipedia templates. In contrast for db3.9-infobox data set which includes also triples generated by information extraction algorithms, most of the properties (i.e., the ones from the dbpepdia.org/property namespace) are not speci ed within the terminology.

In general, underspeci cation may be the result of precise modelling choices, e.g., the property dc:date from the lb data set. This property is intentionally not speci ed in order to favor its reuse, being the Dublin Core Elements (i.e., dc) a general purpose vocabulary. Another example is the dbo:timeInSpace property from the db2014-core data set, whose domain is not speci ed in the corresponding terminology. However, this property is used in a speci c way as demonstrated by patterns (dbo:Astronaut, dbo:timeInSpace, xsd:double) and (dbo:SpaceShuttle dbo:timeInSpace, xsd:double). Gaining such understanding of the semantics of the dbo:timeInSpace property by looking only at the terminology axioms is not possible.

We can push our analysis further to a more ne grained level. Figure 3 provides an overview of the number of di erent minimal types that constitute the domain and range of unspeci ed properties extracted from the summary of the db2014-core data set. The left part of the plot shows those properties whose semantics is less \clear", in the sense that their domain and range cover a higher number of di erent minimal types e.g., the dbo:type property. Surprisingly, the dbo:religion property is among them: its semantics is not as clear as one might think, as its range covers 54 disparate minimal types, such as dbo:Organization, dbo:Sport or dbo:EthnicGroup. Conversely, the property dbo:variantOf, whose semantics is intuitively harder to guess, is used within the data set with a very speci c meaning, as its domain and range covers only 2 minimal types: dbo:Automobile and dbo:Colour.

Small-scale user study. Formulating SPARQL queries is a task that requires prior knowledge about the data set. ABSTAT could support users that lack such knowledge by providing valuable information about the content of the data set. We designed a user study based on the assignment of cognitive tasks related to query formulation. We selected a set of queries from the Questions and Answering in Linked Open Data benchmark7 [ 19 ] to the db3.9-infobox data set. The selected queries were taken from logs of the PowerAqua QA system and are believed to be representative of realistic information needs [ 9 ], although we cannot guarantee that they cover every possible information need. We provided the participants the query in natural language and a \template" of the corresponding SPARQL query, with spaces intentionally left blank for properties and/or concepts. For example, given the natural language speci cation Give me all people that were born in Vienna and died in Berlin, we asked participants to ll in the blank spaces: SELECT DISTINCT ?uri WHERE f ?uri ... <Vienna> . ?uri ... <Berlin> . g We selected ve queries of increasing length, de ned in terms of the number of triple patterns within the WHERE clause; one query of length one, two of length two and two of length three. Intuitively, the higher the query length, the more di cult it is to be completed. We could use a limited number of queries because the tasks are time-consuming and fatigue-bias should be reduced [ 13 ]. Overall 20 participants with no prior knowledge about the ABSTAT framework were selected and split into 2 groups: abstat and control. We pro led all the participants in terms of knowledge about SPARQL, data modelling, DBpedia dataset and ontology, so as to create two homogeneous groups. We trained for about 20 minutes on how to use ABSTAT only the participants from the rst group. Both groups execute SPARQL queries against the db3.9-infobox data set through the same interface and were asked to submit the results they considered correct for each query. We measured the time spent to complete each query and the correcteness of the answers. The correcteness of the answers is calculated as the ratio between the number of correct answers to the given query agains the total number of answers. Table 3 provides the results of the performance of the 7 http://greententacle.techfak.uni-bielefeld.de/~cunger/qald/ users on the query completion task8. The time needed to perform the 5 queries from all partecipiants in average is 38.6m, while the minimum and the maximum time is 18.4m and 59.2m respectively. The independent t-test, showed that the time needed to correctly answer Q5, the most di cult query, was statistically signi cant for two groups. There was a signi cant e ect between two groups, t(16) = 10.32, p < .005, with mean time for answering correctly to Q5 being signi cantly higher (+336s) for the control group than for abstat group. Using 5 queries is coherent with other related work which suggest that the user study would have 20-60 participants, who are given 10-30 minutes of training, followed by all participants doing the same 2-20 tasks, during a 1-3 hour session [ 13 ].

Observe that the two used strategies to answer the queries by participants from the control group were: to directly access the public web page describing the DBpedia named individuals mentioned in the query and very few of them submitted explorative SPARQL queries to the endpoint. Most of the users searched on Google for some entity in the query, then consulted DBpedia web pages to nd the correct answer. DBpedia is arguably the best searchable dataset, which is why this explorative approach was successful for relatively simple queries. However, this explorative approach does not work with other non-indexed datasets (e.g., LinkedBrainz) and for complex queries. Instead, participants of the abstat group took advantage of the summary, obtaining huge bene ts in terms of average completion time, accuracy, or both. Moreover, they achieved increasing accuracy over queries at increasing di culty, still performing the tasks faster. We interpret the latter trend as a classical cognitive pattern, as the participants became more familiar with ABSTATBrowse and ABSTAT8 The raw data can be found at http://abstat.disco.unimib.it/downloads/ user-study Search web interfaces. The noticeable exception is query 3. In particular, participants from the abstat group completed the query in about twice the time of participants from control group. This is due to the fact that the individual Surfing (which is used as object of the property dbo:occupation) is classi ed with no type other than owl:Thing. As a consequence, participants from the abstat group went trough a more time consuming trial and error process in order to guess the right type and property. Participants from the abstat group nally came to the right answer, but after a longer time. This issue might be solved by applying state-of-the-art approaches for type inference on source RDF data [ 12 ] and suggest possible improvements of ABSTAT for example including values for concepts that are de ned by closed and relatively small instance sets. 5

Related Work

We compare our work to approaches explicitly proposed to summarize Linked Data and ontologies, and to extract statistics about the data set.

A rst body of work has focused on summarization models aimed at identifying subsets of data sets or ontologies that are considered to be more relevant. Authors in [ 21 ] rank the axioms of an ontology based on their salience to present to the user a view about the ontology. RDF Digest [ 18 ] identi es the most salient subset of a knowledge base including the distribution of instances in order to e ciently create summaries. Di erently from these approaches, ours aims at providing a complete summary with respect to the data set.

A second body of work has focused on approaches to describe linked data sets by reporting statistics about the usage of the vocabulary in the data. The most similar approach to ABSTAT is Loupe [ 10 ], a framework to summarize and inspect Linked Data sets. Loupe extracts types, properties and namespaces, along with a rich set of statistics. Similarly to ABSTAT, Loupe o ers a triple inspection functionality, which provides information about triple patterns that appear in the data set and their frequency. Triple patterns have the form <subjectType, property, objectType> and are equivalent to our patterns. However, Loupe does not apply any mimimalization technique: as shown in Section 4.1, summaries computed by our model are signi catively more compact.

In [ 2 ], authors consider vocabulary usage in the summarization process of an RDF graph and use information similar to patterns. A similar approach is also used in MashQL [ 4 ], a system proposed to query graph-based data (e.g., RDF) without prior knowledge about the structure of a data set. Our model excludes several redundant patterns from the summary through minimalization, thus producing more compact summaries. Knowledge pattern extraction from RDF data is also discussed in [ 15 ], but in the context of domain speci c experiments and not with the purpose of de ning a general linked data summarization framework. Our summarization model can be applied to any data set that uses a reference ontology and focuses on the representation of the summary.

Other approaches proposed to describe data sets do not extract connections between types but provide several statistics. SchemeEx extracts interesting theoretic measures for large data sets, by considering the co-occurrence of types and properties [ 7 ]. A data analysis approach on RDF data based on an warehousestyle analytic is proposed in [ 3 ]. This approach focuses on the e ciency of processing analytical queries which poses additional challenges due to their special characteristics such as complexity, evaluated on typically very large data sets, and long runtime. However, this approach di erently from ours requires the design of a data warehouse specially for a graph-structured RDF data. Linked Open Vocabularies9, RDFStats [ 8 ] and LODStats [ 1 ] provide several statistics about the usage of vocabularies, types and properties but they do not represent the connections between types.

The approach in [ 20 ] induces a schema from data and their axioms represent stronger patterns compared to the patterns extracted by our approach. ABSTAT aims to represent every possible connections existing among types while EL axioms aims to mine stronger constraints.

The authors in [ 5 ] have a goal even more di erent than ours. They provide lossless compression of RDF data using inference obtaining thus a reduction rate of 0.5 in best cases. Our approach loses information about instances because aims at representing schema-level patterns, but achieves a reduction rate of 0.002. 6

Conclusion and Future Work

Getting an understanding of the shape and nature of the data from large Linked Data sets is a complex and a challenging task. In this paper, we proposed a minimalization-based summarization model to support data set understanding. Based on the experimentation we show that our summarization framework is able to provide both compact and informative summaries for a given data set. We showed that using ABSTAT framework the summaries are more compact than the ones generated from other models and they also help the user to gain insights about the semantics of underspeci ed properties in the ontology. The results of our preliminary experiment showed that ABSTAT help users formulating SPARQL queries both in terms of time and accuracy.

We plan to run the experiment in large scale, thus including more users with di erent background characteristics in order to analyse in details which is the target group of users for which ABSTAT is more useful. Several are the future research directions. We plan to complement our coverage-oriented approach with relevance-oriented summarization methods based on connectivity analysis. Another interesting direction was highlighted by our user study, that is the inference of speci c types for untyped instances found in the data set. We are also planning to consider the inheritance of properties to produce even more compact summaries. Finally, we envision a complete analysis of the most important data set available in the LOD cloud. 9 http://lov.okfn.org/

 1. S. Auer , J. Demter , M. Martin , and J. Lehmann. LODStats - An Extensible Framework for High-Performance Dataset Analytics . In EKAW (2) , 2012 . 2. S. Campinas , T. E. Perry , D. Ceccarelli , R. Delbru , and G. Tummarello. Introducing RDF Graph Summary with Application to Assisted SPARQL Formulation . In DEXA, 2012 . 3. D. Colazzo , F. Goasdoue , I. Manolescu, and A. Roatis . RDF Analytics: Lenses over Semantic Graphs . In WWW , 2014 . 4. M. Jarrar and M. Dikaiakos . A Query Formulation Language for the Data Web . IEEE Trans. Knowl. Data Eng , 24 ( 5 ), 2012 . 5. A. K. Joshi , P. Hitzler , and G. Dong . Logical linked data compression . In The Semantic Web: Semantics and Big Data , pages 170 { 184 . Springer, 2013 . 6. S. Khatchadourian and M. P. Consens . ExpLOD: Summary-Based Exploration of Interlinking and RDF Usage in the Linked Open Data Cloud . In ESWC (2) , 2010 . 7. M. Konrath , T. Gottron , S. Staab , and A. Scherp. SchemEX - E cient construction of a data catalogue by stream-based indexing of linked data . J. Web Sem., 16 , 2012 . 8. A. Langegger and W. Wo . RDFStats - An Extensible RDF Statistics Generator and Library . In DEXA, 2009 . 9. V. Lopez , C. Unger , P. Cimiano , and E. Motta . Evaluating question answering over linked data . Web Semantics: Science, Services and Agents on the World Wide Web , 21 :3{ 13 , 2013 . 10. N. Mihindukulasooriya , M. Poveda Villalon , R. Garcia-Castro , and A. GomezPerez. Loupe - An Online Tool for Inspecting Datasets in the Linked Data Cloud . In ISWC Posters & Demonstrations , 2015 . 11. M. Palmonari , A. Rula , R. Porrini , A. Maurino , B. Spahiu , and V. Ferme . ABSTAT: Linked Data Summaries with ABstraction and STATistics . In ESWC Posters & Demonstrations , 2015 . 12. H. Paulheim and C. Bizer . Type Inference on Noisy RDF Data . In ISWC , 2013 . 13. A. Perer and B. Shneiderman . Integrating statistics and visualization: case studies of gaining clarity during exploratory data analysis . In Proceedings of the SIGCHI conference o Human Factors in computing systems , pages 265 { 274 . ACM, 2008 . 14. S. Peroni , E. Motta, and M. d'Aquin. Identifying Key Concepts in an Ontology, through the Integration of Cognitive Principles with Statistical and Topological Measures . In ASWC , 2008 . 15. V. Presutti , L. Aroyo , A. Adamou , B. A. C. Schopman , A. Gangemi , and G. Schreiber. Extracting Core Knowledge from Linked Data . In COLD2011 , 2011 . 16. M. Schmachtenberg , C. Bizer , and H. Paulheim . Adoption of the Linked Data Best Practices in Di erent Topical Domains . In ISWC, 2014 . 17. S. Staab and R. Studer . Handbook on ontologies. Springer Science & Business Media , 2010 . 18. G. Troullinou, H. Kondylakis , E. Daskalaki, and D. Plexousakis . RDF Digest: E cient Summarization of RDF/S KBs . In ESWC, 2015 . 19. C. Unger , C. Forascu , V. Lopez , A. N. Ngomo , E. Cabrio, P. Cimiano , and S. Walter . Question Answering over Linked Data (QALD-4) . In CLEF , 2014 . 20. J. Volker and M. Niepert . Statistical schema induction . In The Semantic Web: Research and Applications , pages 124 { 138 . Springer, 2011 . 21. X. Zhang , G. Cheng, and Y. Qu . Ontology summarization based on rdf sentence graph . In WWW , 2007 .