Morph-CSV: Virtual Knowledge Graph Access for Tabular Data

Abstract

Virtual knowledge graph access has traditionally focused on providing ontology-based access to relational databases (RDB) proposing SPARQL-to-SQL query translation techniques and optimizations. With the advent of mapping languages or annotations such as RML or CSVW, these techniques have been applied over tabular data by considering each source as a single table that can be loaded into an RDB. However, such techniques do not take into account those characteristics that are normally present in real-world CSV les (e.g., normalization, constraints, joins). In this paper we present Morph-CSV, a framework for enhancing virtual knowledge graph access over a set of CSV les by using a combination of CSVW annotations and RML mappings with FnO transformation functions. Exploiting these inputs, the framework creates an enriched RDB representation of the CSV les together with the corresponding R2RML mappings, enabling the use of existing query translation (SPARQL-to-SQL) techniques and tools.

Keywords

• Knowledge Graphs
• CSV

-

RML

CSVW Semi-structured data formats, and particularly spreadsheets in the form of CSV or Excel les, are one of the most widely-used formats to publish data on the Web. There are several reasons why tabular formats are so popular for data publication. First, they are easy to generate by data providers. In many cases, they are even used as one of the main ways to manage data inside organizations. Second, they are easy to consume with common o ce tools (e.g., Excel, LibreO ce) and there are advanced tools that can be used to process them (e.g., OpenRe ne, Tableau). However, more advanced consumers (e.g., application developers, knowledge workers) often have to face some relevant challenges when consuming tabular data: there is no standard way to query data in them as it can be done with other types of data formats, such as RDB, JSON or XML; data Copyright c 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). are di cult to integrate since data constraints and relationships across di erent les are not explicit; data are often di cult to understand since column names are generally heterogeneous.

Some of these challenges may be dealt following a Semantic Web approach. Virtual knowledge graphs (VKG) provide a uni ed view and common access to a set of data sources based on ontologies and mappings, translating SPARQL queries into queries that are supported by the underlying source. Although current proposals [ 7 ] provide support for querying this kind of formats, they treat each source as if it was a single not-normalized RDB table with no keys or integrity constraints, important elements that are used by SPARQL-to-SQL engines for e cient querying. Several languages have been proposed to specify annotations to deal with the heterogeneity of tabular datasets such as CSVW [ 8 ] metadata and RML+FnO [ 5 ] mapping rules, but engines or systems have to take them into account in their VKG access pipeline.

In this demo we present Morph-CSV, an open source engine1 that extends the typical VKG work ow to enhance performance and query completeness over tabular datasets. Our approach exploits the information from CSVW annotations and RML+FnO mappings so as to obtain details on the underlying schema, required transformation functions, missing information, etc., pushing down their application directly over the tabular dataset. It generates and populates an enriched and normalized RDB schema from the CSV les, and translates RML+FnO to an equivalent function-free R2RML mapping document [ 4 ], so that existing SPARQL-to-SQL optimizations can be used to query them. Finally, we describe two real use cases from transport and biomedical domains where Morph-CSV is applied to enhance virtual KG access. 2

Tabular Annotations for VKG: RML+FnO and CSVW There are speci c challenges on querying tabular datasets using a VKG access approach that have not been tackled by existing techniques. The selection of the sources to answer a query, the normalization or heterogeneity of the dataset and the absence of indexes a ect the performance and completeness of SPARQL-toSQL engines. To deal with these challenges, RML [ 6 ] extends the R2RML W3C Recommendation to provide support beyond relational databases, such as XML, CSV, JSON, etc. Recently, RML has been integrated with the Function Ontology (FnO) to support other types of transformations [ 5 ]. Additionally, CSVW annotations [ 8 ] is a W3C Recommendation that provides metadata annotations for tabular data on the web. In Table 1 we summarized the properties of these two speci cations and its related challenge(s). The manual and ad-hoc preparation of a tabular dataset for VKG access is usually the most time-consuming and less reproducible task. Exploiting available standard and declarative annotations allows its generalization and automatization, as well as ensuring query completeness and improving performance of SPARQL-to-SQL techniques. 1 https://doi.org/10.5281/zenodo.3572132 The Morph-CSV2 open source engine exploits the typical inputs of a VKG process (query, metadata and mappings) to improve performance and query completeness over tabular sources, dealing with their identi ed challenges. More in detail, it extends the starting phase of a typical VKG access work ow to select the relevant sources from an input query, extract implicit constraints from RML+FnO [ 5 ] mappings and CSVW [ 8 ] metadata, pushing down their application directly to the selected sources and nally, it generates enriched inputs for a SPARQL-to-SQL process (R2RML mappings and an RDB instance). The architecture of Morph-CSV is shown in Figure 1, where we present the steps to exploit declarative annotations for enhancing SPARQL query translations over tabular data: i) Source Selection: Using the SPARQL query and the mapping rules, the engine selects only the relevant sources (and columns inside each source) that are relevant to answer the input query. ii) Normalization: Two functions for performing data normalization were implemented. The rst one is the treatment of multi-values in columns while the second one is the treatment of multiple entities in the same source. iii) Data Preparation: In this step, three di erent functions are executed. First, it performs all the substitutions such as default values, NULL values and date formats, then, it creates a new column in the speci c source applying the transformation function de ned in RML+FnO and nally, the engine removes all duplicates in the raw data. iv) Mapping Translation: The mapping rules are translated accordingly to the generated

2 https://morph.oeg.fi.upm.es/tool/morph-csv

Morph-CSV

Normalization

&

Data Preparation

RDB Schema Creation &

Load Mapping

Translation Extracting Constraints

Source Selection SPARQLSQL

Engine + RML+FnO data from RML+FnO to a standard R2RML document [ 4 ]. v) Schema Creation and Load: An optimized SQL schema is generated applying integrity constraints (PK-FK), and the selected data sources are loaded. 4

Use Cases

In this demo we run Morph-CSV over two real use cases: 1. Transport National Access Points (NAP). Since 2019, most European countries are required to public transport data in accessible open query points called National Access Points or NAP3. The main issues related to access to transport data across Europe, will be how to deal with the heterogeneity of these access points and data formats, and how to e ciently query them. Using the de-facto standard for publishing open transport data, GTFS4, which is composed by a set of tabular sources, our engine will exploit RML+FnO and CSVW annotations to enable e cient and complete access to GTFS data through SPARQL queries. 2. Virtual KG over Bio2RDF. Bio2RDF [ 1 ] is one of the most popular projects that integrates and publishes biomedical datasets using Semantic Web technologies. Although its community has actively contributed to the generation of these datasets, they perform the integration using ad-hoc programming scripts, which negatively a ects the maintainability of the project, therefore, SPARQL queries may return outdated results. Selecting the tabular original sources of Bio2RDF, Morph-CSV constructs a virtual KG over them following a declarative approach, hence, improving the maintainability of the project and ensuring up to date results from the SPARQL queries. 3 https://ec.europa.eu/transport/themes/its/road/action_plan/nap_en 4 https://developers.google.com/transit/gtfs

The obtained results are shown in a landing page5 and in a video6. Besides the real use cases, we present the results obtained in terms of performance and completeness with Morph-CSV, using two virtual knowledge graph benchmarks from the state of the art (BSBM [ 2 ] and GTFS-Madrid-Bench [ 3 ]), and two well known open source SPARQL-to-SQL engines (Morph-RDB and Ontop). 5

Conclusions and Future Work

Morph-CSV enhances virtual knowledge graph access over heterogeneous CSV les. It takes as input a set of CSV les, CSVW annotations, and an RML+FnO mapping, and generates as output an enriched RDB instance with data from the CSV les together with R2RML mappings, so that they can be used by any state-of-the-art R2RML-compliant OBDA engine. As part of our future work, we will improve its performance with new optimizations in the query-translation process. We will also extend it for other types of data (e.g., XML, JSON). Acknowledgements. The work presented in this paper is supported by the Spanish Ministerio de Econom a, Industria y Competitividad and EU FEDER funds under the DATOS 4.0: RETOS Y SOLUCIONES - UPM Spanish national project (TIN2016-78011-C4-4-R) and by an FPI grant (BES-2017-082511)

5 https://morph.oeg.fi.upm.es/demo/morph-csv

6 https://youtu.be/yzskzFSAMzA

References

1. Belleau , F. , Nolin , M.A. , Tourigny , N. , Rigault , P. , Morissette , J.: Bio2RDF: towards a mashup to build bioinformatics knowledge systems . Journal of biomedical informatics 41(5) , 706 { 716 ( 2008 )

2. Bizer , C. , Schultz , A. : The Berlin SPARQL Benchmark. International Journal on Semantic Web and Information Systems (IJSWIS) 5 ( 2 ), 1 { 24 ( 2009 )

3. Chaves-Fraga , D. , Priyatna , F. , Cimmino , A. , Toledo , J. , Ruckhaus , E. , Corcho , O. : GTFS-Madrid-Bench: A Benchmark for Virtual Knowledge Graph Access in the Transport Domain . Journal of Web Semantics 65 ( 2020 )

4. Corcho , O. , Priyatna , F. , Chaves-Fraga , D. : Towards a new generation of ontology based data access . Semantic Web 11 , 153 { 160 ( 2020 )

5. De Meester , B. , Maroy , W. , Dimou , A. , Verborgh , R. , Mannens , E.: Declarative data transformations for Linked Data generation: the case of DBpedia . In: European Semantic Web Conference. pp. 33 { 48 . Springer ( 2017 )

6. Dimou , A. , Vander Sande , M. , Colpaert , P. , Verborgh , R. , Mannens , E., Van de Walle, R.: RML: a generic language for integrated RDF mappings of heterogeneous data . In: LDOW ( 2014 )

7. Priyatna , F. , Corcho , O. , Sequeda , J.: Formalisation and experiences of R2RMLbased SPARQL to SQL query translation using morph . In: Proceedings of the 23rd international conference on World wide web . pp. 479 { 490 ( 2014 )

8. Tennison , J. , Kellogg , G. , Herman , I. : Model for tabular data and metadata on the web . W3C recommendation. World Wide Web Consortium (W3C) ( 2015 )