In this paper we present Mastro, a Java tool for ontologybased data access (OBDA) developed at Sapienza Universita di Roma. Mastro manages OBDA systems in which the ontology is speci ed in a logic of the DL-Lite family of Description Logics speci cally tailored to ontology-based data access, and is connected to external data management systems through semantic mappings that associate SQL queries over the external data to the elements of the ontology. Advanced forms of integrity constraints, which turned out to be very useful in practical applications, are also enabled over the ontologies. Optimized algorithms for answering expressive queries are provided, as well as features for intensional reasoning and consistency checking. Mastro has been successfully used in several projects carried out in collaboration with important organizations, on which we brie y comment in this paper.
In this paper we present the current version of Mastro, a system for ontologybased data access (OBDA) developed at Sapienza Universita di Roma. Mastro allows users for accessing external data sources through an ontology expressed in a fragment of the W3C Web Ontology Language (OWL).
As in data integration systems [11], mappings are used in OBDA to specify the semantic correspondence between a uni ed view of the domain (called global schema in data integration terminology) and the data stored at the sources. The distinguishing feature of the OBDA approach, however, is the fact that the global uni ed view is speci ed using an ontology language, which typically allows to provide a rather rich conceptualization of the domain of interest, that is independent from the representation adopted for the data stored at the sources. This choice provides several advantages: it allows for a declarative approach to data access and integration and provides a speci cation of the domain that is independent from the data layer; it realizes logical/physical independence of the information system, which is therefore more accessible to non-experts of the underlying databases; the conceptual approach to data access does not impose to fully integrate the data sources at once, as it often happens in data integration mediator-based system, but the design can be carried out in an incremental way; the conceptual model available on the top of the system provides a common ground for the documentation of the data stores and can be seen as a formal speci cation for mediator design.
Mastro has solid theoretical basis [
Answering unions of conjunctive queries in OBDA systems managed by
Mastro can be done through a very e cient technique that reduces this task to
standard SQL query evaluation. Indeed, conjunctive query answering has been
shown to be in LogSpace (in fact in AC0) w.r.t. data complexity [
Mastro is developed in Java and can be connected to any data management system allowing for a JDBC connection, e.g., a relational DBMS. In those cases in which several, possibly non-relational, sources need to be accessed, Mastro can be coupled with a relational data federation tool1, which wraps sources and represents them as a single (virtual) relational database.
The rest of the paper is organized as follows. In Section 2, we brie y describe the framework of ontology-based data access. In Section 3, we describe the query answering algorithm of the Mastro system. In Section 4, we report on some real world information integration applications where Mastro has been successfully trialed. In Section 5, we conclude the paper by discussing related work. 2
In OBDA, the aim is to give users access to a data source or a collection thereof,
by means of a high-level conceptual view speci ed as an ontology. The ontology
is usually formalized in Description Logics (DLs) [
A DL ontology is a pair hT ; Ai [
The semantics of an ontology is given in terms of rst-order
interpretations [
Among the extensional reasoning tasks w.r.t. a given ontology hT ; Ai, the most relevant ones are ontology satis ability and query answering.
In particular, we are interested in the class of conjunctive queries (CQ). A CQ q over an ontology O (resp. TBox T ) is an expression of the form q(x) 9y.conj(x; y) where x are the so-called distinguished variables, y are existentially quanti ed variables called the non-distinguished variables, and conj(x; y) is a conjunction of atoms of the form A(z), P (z; z0), U (z; z0) where A is a concept name, P is a role name and U is an attribute name, and z, z0 are either variables in x or in y or constants. The arity of q is the arity of x. A CQ of arity 0 is called a boolean conjunctive query. A union of conjunctive queries (UCQ) is a query of the form q(x) Wi 9yi.conj(x; yi):
Given a query q(x) (either a conjunctive query or an union of conjunctive queries) and an ontology O, the certain answers to q(x) over O is the set cert(q; O) of all tuples t of constants appearing in O, such that, when substituted for the variables x in q(x), we have that O j= q(t), meaning that tI 2 qI for every I 2 M od(O). Notice that the answer to a boolean query is either the empty tuple, considered as true, or the empty set, considered as f alse.
In OBDA, the extensional level is not represented directly by an ABox, but rather by a database that is connected to the TBox by means of suitable mapping assertions2. Such mapping assertions have the form ; , where , called the body of the assertion, is an arbitrary SQL query over the underlying database, and , called the head, is a CQ over the TBox T . Intuitively, a mapping assertion speci es that the tuples returned by the SQL query are used to generate the facts that instantiate the concepts, roles, and attributes in .
All the notions given above can be easily generalized to OBDA systems, where a TBox T is connected to an external database D through mappings M, denoted hT ; M; Di. In particular, the models of hT ; M; Di are those interpretations of T that satisfy the assertions in T and that are consistent with the tuples retrieved by M from D (see [13] for the formal details). Satis ability amounts to checking whether hT ; M; Di admits at least one model, while answering a query Q amounts to computing the tuples that are in the evaluation of Q in every model of hT ; M; Di.
Mastro is able to deal with DL TBoxes that are expressed in
DLLiteA;id;den , a member of the DL-Lite family of lightweight DLs [
Basic DL-LiteA;id;den expressions are de ned as follows: B ! C ! V !
A j 9Q j (U ) B j :B U j :U
Q ! R !
P j P Q j :Q
E ! F ! (U ) T1 j j Tn where, A, P , and P denote an atomic concept, an atomic role, and the inverse of an atomic role respectively; (U ) (resp. (U )) denotes the domain (resp. the range) of an attribute U , i.e., the set of objects (resp. values) that U relates to values (resp. objects); T1; : : : ; Tn are unbounded pairwise disjoint prede ned value-domains; B is called basic concept.
A DL-LiteA;id;den TBox is a nite set of the following assertions: B v C Q v R U v V E v F (funct Q) (funct U ) (id B 1; : : : ; n) 8y.conj(t) ! ? (concept inclusion assertion) (role inclusion assertion) (attribute inclusion assertion) (value-domain inclusion assertion ) (role functionality assertion) (attribute functionality assertion ) (identi cation assertion) (denial assertion) In identi cation assertions [5], i is a path, i.e., an expression built according to the following syntax: ! S j D? j 1 2, where S denotes an atomic role, the inverse of an atomic role, an attribute, or the inverse of an attribute, 1 2 denotes the composition of paths 1 and 2, and D?, called test relation, represents the identity relation on instances of D, which can be a basic concept or a value-domain. Test relations are used to impose that a path involves instances of a certain concept or value-domain. In DL-LiteA;id;den , identi cation assertions are local, i.e., at least one i 2 f 1; :::; ng has length 1, i.e., it is an atomic role, the inverse of an atomic role, or an attribute. Intuitively, an identi cation assertion of the above form asserts that for any two di erent instances o, o0 of B, there is at least one i such that o and o0 di er in the set of their i- llers, that is the set of objects that are reachable from o by means of i.
In denial assertions [10], conj(y) is de ned as for boolean CQs. Intuitively, a denial assertion of the above form states that there must not exist any tuple y satisfying conj(y), i.e., that the answer to the boolean query q() 9y.conj(y) must be empty.
Finally, in a DL-LiteA;id;den TBox T , the following condition must hold: each role or attribute that either is functional in T or appears (in either direct or inverse direction) in a path of an identi cation assertion in T is not specialized, i.e., it does not appear in the right-hand side of assertions of the form Q v Q0 or U v U 0.
Mapping assertions handled by Mastro are assertions of the form ; , where is an arbitrary SQL query over the underlying database, and is a conjunction of atoms whose predicates are the concepts, roles, and attributes of the TBox. Notice that, due to the fact that is a conjunction of atoms (as opposed to a query, possibly with existentially quanti ed variables), such mappings can be considered as a special form global-as-view (GAV) mappings [11].
In order to overcome the so-called impedance mismatch between the database, storing values, and the TBox, to be interpreted over a domain of objects, the mapping assertions are used in Mastro to specify how to construct abstract objects from the tuples of values retrieved from the database. This is done by allowing one to use function symbols in the atoms in : together with the values retrieved by , such function symbols generate so called object terms, which serve as object identi ers for individuals in the ontology. We notice that the semantics we adopt in Mastro establishes that di erent terms denote di erent objects (unique name assumption), so that di erent terms never need to be equated during reasoning, which is coherent with the assumption of not having existentially quanti ed variables in the body of mappings.
For the logics of the DL-Lite family it has been shown that for unions of
conjunctive queries (UCQs), under the unique name assumption, query
answering can be carried out e ciently in the size of the data, by reducing it to SQL
query evaluation over the ABox seen as a database [
As an example, consider the OBDA system hT ; M; Di, where the TBox T is constituted by the following set of intensional assertions: fN ationalF light v F light, InternationalF light v F lightg, D is a database constituted by a set of relations with the following signature: FL TB[fl num:string, departure:integer, arrival:integer], AIRPORT TB[airpt code:integer, name:string, country:string], and M contains the following mapping assertions: SELECT fl num FROM FL TB,AIRPORT TB A1,AIRPORT TB A2 WHERE departure = A1.airpt code and arrival = A2.airpt code and A1.country = 'IT' and A2.country = 'IT' SELECT fl num FROM FL TB,AIRPORT TB A1,AIRPORT TB A2 WHERE departure = A1.airpt code and arrival = A2.airpt code and (A1.country != 'IT' or A2.country != 'IT') ; NationalFlight( (fl num)) ; InternationalFlight( (fl num)) which specify how to construct instances of the ontology concepts N ationalF light and InternationalF light starting from the database relations FL TB and AIRPORT TB.
In this section we describe the query rewriting process of the Mastro system. The technique is purely intensional and is performed in three steps (see Figure 1): 1. TBox rewriting: The rst step rewrites the input UCQ according to the knowledge expressed by the TBox. The rewriting, performed using the Presto algorithm [15], produces as output a non-recursive Datalog program, which encodes the knowledge expressed by the TBox and the user query. The output Datalog program contains the de nition of auxiliary predicates, not belonging to the alphabet of the ontology. 2. Datalog Unfolding: The output of the rst step is then unfolded into a new UCQ by means of the Datalog Unfolding algorithm. It consists of a classic rule unfolding technique which eliminates all the auxiliary predicate symbols introduced by the Presto algorithm and produces a nal UCQ expressed in terms of ontology concepts, roles, and attributes. 3. Mapping Unfolding: The last step takes the unfolded UCQ and the mapping assertions as input and produces an SQL query which can be directly evaluated over the data sources. In particular, the mapping assertions are rst split into assertions of a simpler form, in which the head of every mapping assertion contains only a single ontology predicate; then, the nal reformulation is produced through a mapping unfolding step, as described in [13].
More speci cally, the Presto algorithm is an optimization of the well-known
PerfectRef [
The usefulness of OBDA and the e ciency of the Mastro system were proved
by several real world applications in which it has been experimented. In the
following, we report on the experiments carried out with Banca Monte dei Paschi
di Siena (MPS), the Italian Ministry of Economy and Finance (MEF) and the
Telecom Italia, the main Italian telephone company. Other experiments have
been recently carried out with SELEX Sistemi Integrati (SELEX-SI), and
Accenture [
Monte dei Paschi di Siena. Within a joint project with Banca Monte dei Paschi di Siena (MPS)3, Free University of Bozen-Bolzano, and Sapienza Universita di Roma, we used Mastro for accessing a set of data sources from the actual MPS data repository by means of an ontology [16]. In particular, we focused on the data exploited by MPS personnel for risk estimation in the process of granting credit to bank customers. A 15 million tuple database, stored in 12 relational tables managed by the IBM DB2 RDBMS, has been used as data source collection in the experimentation. Such source data are managed by a dedicated application, which is in charge of guaranteeing data integrity (in fact, the underlying database does not force constraints on data). Not only the application performs various updates, but data is updated on a daily basis to identify connections between customers that are relevant for the credit rating estimation.
The main challenge that we tackled within the experimentation was the ontology and mapping design. This was a seven man-months process that required to both inspect the data source and interview domain experts, and was complicated by the fact that the source was managed by a speci c application. The resulting OBDA system is de ned in terms of approximately 600 DL-LiteA;id assertions over 79 concepts, 33 roles and 37 attributes, and 200 mapping assertions.
The experimentation showed that the usefulness of the Mastro system goes beyond data integration applications and embraces data quality management. In particular, it con rmed the importance of several distinguished features of our system, namely, identi cation constraints and denial constraints, which have been used extensively to model important business rules. Notably, checking that such rules were satis ed by data retrieved from the sources through mappings led to highlight unexpected incompleteness and inconsistency in the data sources.
Our work has also pointed out the importance of the ontology itself, as a precious documentation tool for the organization. Indeed, the ontology developed in our project is adopted in MPS as a speci cation of the relevant concepts in the organization. At present we are still working with MPS in order to extend the work to cover the core domain of the MPS information system, with the idea that the ontology-based approach could result in a basic step for the future IT architecture evolution. 3 MPS is one of the main banks, and the head company of the third banking group in
Italy (see http://english.mps.it/).
Italian Ministry of Economy and Finance. Mastro has been used within a joint project between Sapienza Universita di Roma and the Italian Ministry of Economy and Finance (MEF). The main objectives of the project have been: the design and speci cation in DL-LiteA of an ontology for the domain of the Italian public debt; the realization of the mapping between the ontology and relational data sources that are part of the management accounting system currently in use at the ministry; the de nition and execution of queries over the ontology aimed at extracting data of core interest for MEF users. In particular, the information returned by such queries relates to sales of bonds issued by the Italian government, maturities of bonds, monitoring of various nancial products, etc., and are at the basis of various reports on the overall trend of the national public debt.
The Italian public dept ontology is over an alphabet containing 164 atomic concepts, 47 atomic roles, 86 attributes, and comprises around 1440 DL-LiteA assertions. The 300 mapping assertions involve around 60 relational tables managed by Microsoft SQLServer. We tested a very high number of queries and produced through Mastro several reports of interest for the ministry. We point out that around 80% of the queries we tested could be executed only thanks to a series of further optimizations introduced in the system that, due to lack of space, we cannot describe here.
Telecom Italia. We nally describe a project we are carrying out in the domain of network inventory systems, together with Telecom Italia, the main Italian company for telecommunication services, which is also a world leading company in this eld. The main objectives of the project are (i) the speci cation of an ontology that formalizes the entire telecommunication network owned by Telecom Italia and (ii) the analysis through the ontology of the information systems that are currently used for network management. The ontology we are going to develop can be partitioned into four layers: Infrastructures and territory layer, which represents main infrastructures used to realize the network, the way in which network elements (e.g., cables, apparatus, connection points) are localized into such infrastructures, and how both infrastructures and network elements are localized with respect to the territory; network topology layer, which represents how connections are realized into the network, essentially representing it as a graph in which edges represent elementary connections among apparatus, and nodes represent apparatus which realizes signal permutations between elementary connections; service layer, which represents all telecommunication services that are deployed on the network and o ered to customer (e.g., voice communication, ADSL, voip); data layer, which represents the actual data exchanged on the network (e.g., data on telephone calls, internet access). In each such layer, the ontology provides the means to precisely represent the current state of the world, and, when considered of interest, also captures past situations, for example to provide tracks to all changes to which certain in the network have undergone. The use of identi cation assertions and epistemic constraints turned out to be crucial for faithful representation of such aspects.
Accessing (possibly disperse) data through a virtual global schema has been deeply investigated in the last two decades in the eld of data integration [11, 8]. From the modeling perspective, however, the main systems produced by this research su er from some weakness, mainly due to the limited expressive power of the languages provided to model the global schema of the integration system. In this respect, Mastro aims at overcoming this limitation by providing the best expressive power allowed while preserving tractability of conjunctive query answering and of the integration tasks. As for the mappings, Mastro adopts a powerful form of the so-called Global-As-View (GAV) mappings [11], and provides optimized algorithms for rewriting global queries with respect their speci cation.
To the best of our knowledge, the only existing system designed for the same aims of Mastro is Quest [14], which has indeed common roots with our tool. Quest is a system for query answering over DL-LiteA ontologies, which can work in both \classical" (i.e., with a local ABox) and \virtual" mode (i.e., as an OBDA system). Quest implements speci c optimizations for query answering, which in particular exploit completeness of the ABox with respect to the TBox. Although rst experiments show e ectiveness of Quest in the classical scenario [14], its usage in the virtual mode is in a still preliminary stage. In particular, we tried to compare Quest with Mastro in the OBDA scenario of the Italian Ministry of Economy and Finance described described in the previous section. Unfortunately, we have not been able to perform such experiments for two reasons: (i) the data source of this application is an SQL Server database; since Quest does not support this DBMS, we could not compare query answering in the two systems; (ii) Quest was not actually able to compute the TBox rewriting of the 23 queries used in our experiments, which are very long conjunctions of atoms, so we could not even compare the query rewriting performances of the two systems.
Nyaya [6] is a novel system which allows for query answering over
ontologies speci ed into linear Datalog , a language that essentially corresponds to
DLR-Lite [
Other DL-Lite-based approaches and reasoners have been developed, which, however, are not able to deal with full OBDA scenarios. In [9] an alternative approach to query answering is presented. Besides a (less complex) query reformulation step, such an approach requires to suitably \extend" the ABox (managed by a RDBMS) with the aim of reducing the amount of rewritten queries produced by the reformulation step. The experimental results support well this approach (notice that in Mastro the size of the reformulation may be exponential in the size of the input query). However, the ABox manipulation that it requires makes it extremely di cult to apply this approach in an OBDA scenario.
The Requiem reasoner [12] implements a rewriting algorithm which reduces the number of queries in the nal reformulation, still being purely intensional like Mastro. However, it currently supports none of the Mastro advanced features, such as identi cation or EQL constraint management, nor mappings to external databases.
The OWLGres prototype [19], which allows for TBox speci cation in DLLite, uses the PostgreSQL DBMS for the storage of the ABox, and provides conjunctive query processing. The algorithm for query answering implemented in OWLGres, however, is not complete with respect to the computation of the certain answers to user queries.
Mastro can also be compared with ontology reasoners which support DLs
di erent from DL-Lite, and in particular with their query answering
capabilities. In this respect, well-known DL reasoners such as RacerPro [7], Pellet [18],
Fact++ [20], and HermiT [17] provide only limited forms of query answering, i.e.,
instance checking/retrieval or grounded conjunctive query answering (c.f. [
Acknowledgments. This research has been partially supported by the EU under FP7 project ACSI { Artifact-Centric Service Interoperation (grant n. FP7257593), and by Regione Lazio under the project \Integrazione semantica di dati e servizi per le aziende in rete". 5. D. Calvanese, G. De Giacomo, D. Lembo, M. Lenzerini, and R. Rosati. Path-based identi cation constraints in description logics. In Proc. of KR 2008, pages 231{241, 2008. 6. R. de Virgilio, G. Orsi, L. Tanca, and R. Torlone. Semantic data markets: a exible environment for knowledge management. In Proc. of CIKM 2011, pages 1559{1564, 2011. 7. V. Haarslev, R. Moller, and M. Wessel. Description logic inference technology: Lessions learned in the trenches. In Proc. of DL 2005, volume 147 of CEUR, ceur-ws.org, 2005. 8. A. Y. Halevy, A. Rajaraman, and J. Ordille. Data integration: The teenage years.
In Proc. of VLDB 2006, pages 9{16, 2006. 9. R. Kontchakov, C. Lutz, D. Toman, F. Wolter, and M. Zakharyaschev. The combined approach to query answering in DL-Lite. In Proc. of KR 2010, pages 247{257, 2010. 10. D. Lembo, M. Lenzerini, R. Rosati, M. Ruzzi, and D. F. Savo. Inconsistencytolerant rst-order rewritability of dl-lite with identi cation and denial assertions.
In Proc. of DL 2012, volume 846 of CEUR, ceur-ws.org, 2012. 11. M. Lenzerini. Data integration: A theoretical perspective. In Proc. of PODS 2002, pages 233{246, 2002. 12. H. Perez-Urbina, B. Motik, and I. Horrocks. A comparison of query rewriting techniques for DL-lite. In Proc. of DL 2009, volume 477 of CEUR, ceur-ws.org, 2009. 13. A. Poggi, D. Lembo, D. Calvanese, G. De Giacomo, M. Lenzerini, and R. Rosati.
Linking data to ontologies. J. on Data Semantics, X:133{173, 2008. 14. M. Rodriguez-Muro and D. Calvanese. High performance query answering over dl-lite ontologies. In Proc. of KR 2012, 2012. To appear. 15. R. Rosati and A. Almatelli. Improving query answering over DL-Lite ontologies.
In Proc. of KR 2010, pages 290{300, 2010. 16. D. F. Savo, D. Lembo, M. Lenzerini, A. Poggi, M. Rodr guez-Muro, V. Romagnoli, M. Ruzzi, and G. Stella. Mastro at work: Experiences on ontology-based data access. In Proc. of DL 2010, volume 573 of CEUR, ceur-ws.org, pages 20{31, 2010. 17. R. Shearer, B. Motik, and I. Horrocks. HermiT: A highly-e cient OWL reasoner.
In Proc. of OWLED 2008, volume 432 of CEUR, ceur-ws.org, 2008. 18. E. Sirin and B. Parsia. Pellet: An OWL DL reasoner. In Proc. of DL 2004, volume 104 of CEUR, ceur-ws.org, 2004. 19. M. Stocker and M. Smith. Owlgres: A scalable OWL reasoner. In Proc. of
OWLED 2008, volume 432 of CEUR, ceur-ws.org, 2008. 20. D. Tsarkov and I. Horrocks. FaCT++ description logic reasoner: System description. In Proc. of IJCAR 2006, pages 292{297, 2006.