=Paper=
{{Paper
|id=Vol-2456/paper12
|storemode=property
|title=T2WML: A Cell-Based Language to Map Tables into Wikidata Records
|pdfUrl=https://ceur-ws.org/Vol-2456/paper12.pdf
|volume=Vol-2456
|authors=Pedro Szekely,Daniel Garijo,Jay Pujara,Divij Bhatia,Jiasheng Wu
|dblpUrl=https://dblp.org/rec/conf/semweb/SzekelyGPBW19
}}
==T2WML: A Cell-Based Language to Map Tables into Wikidata Records==
https://ceur-ws.org/Vol-2456/paper12.pdf
T2WML: A Cell-Based Language To Map Tables
Into Wikidata Records
Pedro Szekely, Daniel Garijo, Jay Pujara, Divij Bhatia, and Jiasheng Wu
Information Sciences Institute
University of Southern California
4676 Admiralty Way, Marina del Rey, CA 90292, USA
{pszekely, dgarijo, jpujara}@isi.edu, {divijbha, jiashenw}@usc.edu
Abstract. The web contains millions of useful spreadsheets and CSV
files, but these files are difficult to use in applications because they use
a wide variety of data layouts and terminology. We present Table To
Wikidata Mapping Language (T2WML), a language that makes it easy
to map and link arbitrary spreadsheets and CSV files to the Wikidata
data model. The output of T2WML consists of Wikidata statements
that can be loaded in the public Wikidata, or loaded in a Wikidata
clone, creating an augmented Wikidata knowledge graph that application
developers can query using SPARQL.1
Keywords: Knowledge Graphs, RDF, Entity Linking, Wikidata
1 Introduction
The web contains millions of useful spreadsheets and CSV files, including data
from many government and international organizations. Organizations that offer
data often have web sites where users can search, browse and download data
on a large number of topics. Most institutions offer their data in Excel and
CSV formats. The downloaded data is seldom directly usable because, unlike
databases, which use one column per variable, spreadsheets often arrange the
data in different layouts.
Fig. 1 illustrates the problem using data about homicide rates in different
countries, downloaded from the United Nations web site2 . We truncated and
colored the files for ease of presentation. The cells with the homicide numbers
are highlighted in green, the cells that provide contextual information for the
value are highlighted in blue, and header cells are highlighted in dark blue.
Fig. 1a shows the layout of the data provided in the UN website; Fig. 1b shows
a more compact representation using multi-level headers; Fig. 1c shows a layout
1
Copyright (c) 2019 for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0). This material is based
upon work supported by United States Air Force under Contract No. FA8650-17-C-
7715.
2
https://dataunodc.un.org/crime/intentional-homicide-victims
Fig. 1. Intentional Homicide Data (Excel file downloaded from dataunodc.un.org)
that could be used to store the data in a database, and that can be used directly
in tools such as Pandas; Fig. 1d illustrates a common convention for arranging
data by topic, by creating stacked tables that share common headings. All tables
present the same homicide data. The interpretation of each value is defined by
four cells (country, year, population and source) that identify the context for a
value. In each table, the context cells are located in different parts of the data.
Only in Fig. 1c (Database) the context cells are in the same row as the value; in
the other tables, context cells appear in different rows, in header rows (examples
a and b), or in visually distinct rows within the table (example d).
Existing languages for mapping structured data to RDF, including R2RML3 ,
RML [1], Karma [3] and CSV2RDF [2] process tabular data row by row, requiring
tabular data to be in database format (Fig. 1c). RML supports non-tabular
formats (JSON and XML) and Karma provides folding and unfolding operators
to rearrange data for row-based processing. None support complex layouts such
as those in examples b or d.
T2WML is a mapping language designed to meet three objectives: 1) Identify
and map data and their context qualifiers in arbitrary data layouts found in Excel
and CSV files without the need of complex preprocessing steps to transform
tables into a canonical "Database" representation; 2) Enable users who are not
familiar with RDF to map spreadsheets and CSV files to knowledge graphs
so that they can augment knowledge graphs with data useful in downstream
applications. 3) Integrate mapping and linking so that the resulting output is
linked to a reference knowledge graph, avoiding the need for a separate linking
process that is often neglected in the current workflows.
3
https://www.w3.org/2001/sw/wiki/R2RML
Fig. 2. T2WML user interface to map the spreadsheet in Fig. 1b to Wikidata
2 T2WML
T2WML uses two concepts to model spreadsheets and CSV files: data blocks,
possibly non-contiguous blocks of cells that contain values of interest (green cells
in Fig. 1), and templates to construct knowledge graph statements. T2WML
uses the Wikidata item/statement data model [4] where items are described
using statements consisting of a property, a value, qualifiers and references4 .
For ease of use, T2WML specifications are defined in YAML (see Fig. 2), a
popular language for writing structured data. Another difference with existing
mapping tools is that T2WML adopts a “link first” strategy where cells that
identify entities in the target knowledge graph (Wikidata in our implementation)
must be linked to the corresponding entities before the input data is processed to
produce statements in the knowledge graph. Multiple tools exist to automatically
link cells in tables to knowledge graphs5 . The standard format for recording the
linking results developed for this challenge, enables our T2WML tool to use the
results from multiple linking tools, in addition to our own.
The T2WML user interface (Fig. 2) colors cells with clickable links in orange.
Users can use the T2WML interface to curate (modify, add or delete) links before
invoking our T2WML processor to produce statements in the target knowledge
graph. The YAML Editor panel in Fig. 2 shows a T2WML specification to
map the data in Fig. 1b to Wikidata. The region section identifies cell blocks
containing the values of a statement (i.e., homicide counts, highlighted in green
4
The Wikidata data model can be translated into RDF and queried using SPARQL
5
www.aicrowd.com/challenges/iswc-2019-cell-entity-annotation-cea-challenge
in the Table Viewer). T2WML defines blocks in terms of the edges that surround
the data. In the example, the edges are defined using constants, but it is possible
to define blocks using predicates (e.g., "two empty rows" or "row with a value
in column A, but otherwise empty".)6
The template section defines the mapping of cell values to elements of a
statement. The T2WML tool instantiates the template once for every cell defined
in the region section, binding the variables $col and $row to the coordinates of
the cell being processed. To facilitate understanding of the template instantiation
procedure, users can click on a data cell in the Table Viewer to see how it is
mapped. The interface shows the values of $col and $row, highlights the cell
containing the item (subject) of the statement, and the cells containing the
qualifiers, and shows the resulting statement in an Output panel.
Users define the relationships between a value cell and other parts of a state-
ment using a simple expression language. Fig. 2 illustrates several use cases: line
8 defines the item for the value in $col /$row as the item in column A and the
same row (A/$row); line 13 identifies the “point in time” qualifier as the value
in the current column in row 4 ($col/4); line 17, illustrates a more complex
expression to define gender as the values of the "applies to people" qualifier.
The value is located in row 3, in the closest non-empty column on the left. This
is specified using the expression $col-n/3. The “-n” part directs the processor
to substitute n with values 0, 1, ... until it reaches a non-empty cell. Simpler
arithmetic expressions such as $col+2/$row-4 are also supported.
3 Discussion
We tested the expressivity of T2WML using 19 variants of the homicides table
(Fig. 1 shows 4 variants). We also created T2WML specifications for several
country indicator files from the World Bank7 , generated Wikidata-compliant
RDF and loaded it on a Wikidata clone. We are now working to enhance T2WML
to support definition of new entities and properties, and constructing a corpus
of mappings for use in research to automate the generation of mappings.
References
1. Dimou, A., Vander Sande, M., Colpaert, P., Verborgh, R., Mannens, E.,
and Van de Walle, R. Rml: A generic language for integrated rdf mappings of
heterogeneous data.
2. Ermilov, I., Auer, S., and Stadler, C. Csv2rdf: User-driven csv to rdf mass
conversion framework. In Proceedings of the ISEM (2013), vol. 13, pp. 04–06.
3. Gupta, S., Szekely, P., Knoblock, C. A., Goel, A., Taheriyan, M., and
Muslea, M. Karma: A system for mapping structured sources into the semantic
web. In Extended Semantic Web Conference (2012), Springer, pp. 430–434.
4. Vrandečić, D., and Krötzsch, M. Wikidata: A free collaborative knowledge-
base. Commun. ACM 57, 10 (Sept. 2014), 78–85.
6
T2WML provides a rich expression language documented in http://anonymopus.
7
For example, http://api.worldbank.org/v2/en/indicator/NY.GDP.PCAP.CN