=Paper=
{{Paper
|id=Vol-1927/paper2
|storemode=property
|title=Temporal Knowledge Extraction for Dataset Discovery
|pdfUrl=https://ceur-ws.org/Vol-1927/paper2.pdf
|volume=Vol-1927
|authors=Lama Saeeda, Petr Kŕemen
|dblpUrl=https://dblp.org/rec/conf/semweb/SaeedaK17
}}
==Temporal Knowledge Extraction for Dataset Discovery==
https://ceur-ws.org/Vol-1927/paper2.pdf
Temporal Knowledge Extraction for Dataset
Discovery
Lama Saeeda and Petr Křemen
Czech Technical University in Prague, Czech Republic,
{saeeda.lama, petr.kremen}@fel.cvut.cz
Abstract. Linked data datasets are usually created with different data
and metadata quality. This makes the exploration of these datasets a
quite difficult task for the users. In this paper, we focus on improving
discoverability of datasets based on their temporal characteristics. For
this purpose, we identify the typology of temporal knowledge that can be
observed inside data. We reuse existing temporal information extraction
techniques available, and employ them to create temporal search indices.
We present a particular use-case of dataset discovery based on more
detailed and completed temporal descriptions for each dataset in the
Czech LOD cloud based on the analyzing of the unstructured content in
the literals as well as the structured properties, taking into consideration
varying data and metadata quality.
Keywords: temporal information, text processing, linked data, dataset
discovery
1 Introduction
Amount of available semantic data in the Linked Open Data (LOD) cloud has
increased enormously in the recent years (since 2014 the number of datasets has
doubled). However, while the amount of data has been rising, quality of data
and metadata is still a serious problem [RMB16]. As a result, the current form
of LOD is not ready for direct exploring of the data inside. Datasets in LOD are
de-facto black boxes and their content – the actual data – is efficiently hidden
during dataset discovery. So, simply speaking – not only data are often hardly
interpretable – it is even hard to find them. Currently, real centralized search for
which one could offer filters or facets for temporal information does not exist.
The same applies for federated search in which e.g. VOiD [BLN11] descriptions
could be enriched with a temporal dimension. Thus, possible beneficiaries of our
approach involve people exploring LOD (developers, journalists, data analysts,
etc.). The aim is to explore datasets which contain potentially relevant data for
their applications without the need to search for particular data directly. Figure 1
shows the typical scenario of dataset exploration. Typically, users iterate multiple
times through this cycle in order to discover the desired datasets.
To achieve that, many tasks should be performed. Among these, finding
temporal context of dataset content and metadata is one of the prominent ones.
Fig. 1. Basic cycle in the exploration process
Temporal context is present in most data across various domains and it is im-
portant for both the data and the metadata about data/dataset management.
The ultimate goal of dataset exploration is to understand what the dataset is
about by minimal information extracted. In the context of this paper, we only
consider temporal information.
Varying metadata quality and relevance prevents dataset provision services
(CKAN1 catalogs, etc.) from supporting their users in sufficient search and ex-
ploration. For example DCAT [Wor14] start/end dates are often missing or im-
precise as we will discuss in the evaluation section. Temporal information is
tremendously important in search, question answering, information extraction,
data exploration, and more. Techniques for identifying and normalizing tempo-
ral expressions work well but they are still poorly used for datasets exploration
purposes. The problem is that context is often missing – knowledge evolves over
time, facts have associated validity intervals. Therefore, ontologies used for data
descriptions should include time as a first-class dimension [WZQ+ 10].
As structured data found in the LOD typically do not have explicitly defined
schema, the notion of dataset summaries emerged in the context of Linked Data.
This notion stems from the dataset exploration scenario in which datasets need
to be described in order to be discovered. In [BKK16], dataset descriptors have
the form of datasets describing other datasets, typically represented as metadata.
In this paper, we approach the problem of exploring datasets temporally where
we describe the datasets with their temporal information which allow to equip
various parts of the dataset with time instants/intervals that the actual content
of the dataset speaks about. A temporal value or set of temporal values, which we
call the time scope of the statements, is associated with each temporal statement
which describes its temporal extent. We compute the temporal coverage of the
dataset and formalize it using an integration between two ontologies, the Unified
Foundational Ontology (UFO-B) [Gui05] and the OWL Time Ontology [HP04],
which we call the Temporal Descriptor Ontology (TDO) and finally, we
order the temporal statements in the dataset chronologically on a linear timeline
to help revealing temporal relationships and reduce users effort while choosing
which datasets are aligned with their usage purpose. The main contributions of
this work are:
– analysis of temporal dimension of datasets utilizing their actual content;
– systematization of the temporal knowledge extracted with the Temporal
Descriptor Ontology (TDO).
1
http://ckan.org/
– provide a baseline formalization to support the representation of the dataset
chronologically on a linear timeline for easier dataset exploration process for
the future work.
2 State of the Art
Many temporal expression extractors were proposed in literature. Temporal
information extraction systems usually use rule-based NLP methods, se-
quence segmentation machine learning algorithms like Hidden Markov Models
(HMMs), or Conditional Random Fields (CRF).
[CDJJ14] presented a survey of the existing literature on temporal informa-
tion retrieval. Also, they categorized the relevant research, described the main
contributions, and compared different approaches. A similar survey was pre-
sented in [Lim16] which introduced existing methods for temporal information
extraction from input texts. Another consideration is the research about tempo-
ral information extraction based on the common patterns or knowledge bases. In
[KEB07] a supervised machine learning approach for solving temporal informa-
tion extraction problem (namely CRF) has been adapted. A hybrid system for
temporal information extraction from clinical text was proposed in [TWJ+ 13]
using both rule-based and machine learning based approaches to extract events,
temporal expressions, and temporal relations from hospital discharge summaries.
MedTime [LCB13] is also a system to extract temporal information from clini-
cal narratives. It comprises a cascade of rule-based and machine-learning pattern
recognition procedures and achieved a micro-averaged f-measure of 0.88. Heidel-
Time [SG10], an open source system for temporal expression extraction, is a rep-
resentative rule-based system that performed well in TempEval22 competition.
The Stanford temporal tagger (SUTime) [CM12] is one of best currently avail-
able temporal taggers with 90.32 F-measure score in TempEval-3 with English
Platinum Test set [ULD+ 13]. SUTime annotations are provided automatically
with the StanfordCoreNLP3 pipeline by including the Named Entity Recogni-
tion annotator. SUTime is a rule-based extractor, which means it can normally
be configured to use rules specified in external files that are convenient to the
data being analyzed.
There is considerable work in natural language processing on event extrac-
tion in narrative texts, [HKL13] presented a novel automated NLP pipeline for
semantic event extraction and annotation (EveSem). In [LCH+ 13], a method
is introduced to effectively identify events by considering the spreading effect
of event in the spatio-temporal space. Researchers in [KDF+ 13] developed and
evaluated a system to automatically extract temporal expressions and events
from clinical narratives. [DHHN16] address the problem of advanced event and
relationship extraction with an event and relationship attribute recognition sys-
tem. EVITA [SKVP05] is an application for recognizing events in natural lan-
guage texts. It recognizes events by applying linguistic rules encoded in the
2
http://www.timeml.org/tempeval2/
3
nlp.stanford.edu/software/corenlp.shtml
form of patterns. It considers Verb Phrases (NP), Noun Phrases (NP) and Ad-
jectival Phrases (ADJP) as most probable candidates for containing events. It
employs different strategies for extracting events from Verbal, Noun and Adjec-
tival Chunks. It make use of POS tagging, lemmatizing, chunking,lexical lookup
and contextual parsing to encode event extraction rules.
The markup language TimeML [PCI+ 03], including the TIMEX3 annotation,
has become an ISO standard and the most commonly used guideline for temporal
annotations, normalizing the temporal expressions to be convenient to handle
programmatically, and resolving relative temporal expressions with respect to a
reference date.
Enriching RDF graphs with missing temporal information has also been
given attention in publishing historical data as RDF. For instance, authors of
[RPNN+ 14] proposes a generic approach for inserting temporal information to
RDF data by gathering time intervals from the Web and knowledge bases. Au-
thors in [MPAGS17] extract temporal timestamps from legacy provenance infor-
mation. In [BKK16], researchers computed a summary of the used schema terms
from the datasets. Then, they enriched the summary by adding time instants
and intervals that are used in the dataset. The summary descriptor is reused
by other new content-based descriptors, one of which is the temporal descrip-
tor. Researchers focused only on the structured representation of the temporal
information to calculate the temporal descriptors.
The Semantic Web provides a suitable environment for representing the tem-
poral data. Web Ontology Language (OWL)4 through the annotation properties
from standard vocabularies, for example, DC5 gives the ability to incorporate
time entities into existing ontologies by representing temporal knowledge and
time-based information. Many ontologies were proposed to represent the tempo-
ral information in a structured way.
The OWL-Time ontology [HP04] is an OWL-2 DL ontology that provides
a vocabulary for expressing facts about topological relations among instants
and intervals, together with information about durations, and about temporal
position including date-time information. Time Intervals6 is an ontology that
extends the OWL-Time ontology. It includes government years and properly
handles the transition to the Gregorian calendar within the UK.
Regarding the temporal information understanding in the current state of
the art, there are missing connection between the existing techniques of extrac-
tion and the reality of the data in the LOD. Data in LOD use particular formats
and knowledge. Extracting the temporal information should not only be about
applying NLP techniques but also taking into consideration the RDF knowl-
edge,the connections and relations that already exist, and the typology and th
structure of the data within LOD. In this work we approach this matter by con-
textualizing the current extraction techniques and directing them to be aligned
with the ontology.
4
https://www.w3.org/TR/owl-ref/
5
http://purl.org/dc/terms/
6
http://reference.data.gov.uk/def/intervals
Fig. 2. Temporal data-properties in the vocabularies
3 Temporal data analysis
3.1 Temporal data analysis in the Linked Open Vocabularies (LOV)
Vocabularies provide the semantic glue enabling Data to become meaningful
Data. Linked Open Vocabularies7 (LOV) offers several hundred of such vocabu-
laries8 frequently used in LOD, together with their mutual interconnections. A
vocabulary in LOV gathers definitions of a set of classes and properties.
To offer a general solution, extensible towards most properties used across
the LOD, we experimented with the most commonly used vocabularies in the
datasets according to LOV focusing mainly on the vocabularies that are reused
in the Czech Linked Open Data. In order to perform the temporal analysis and to
distinguish the temporal properties describing the actual content of the datasets,
two annotation properties were created:
is-temporal denotes data properties with explicitly temporal data type ex-
pressing time or date, for example but not limited to xsd:date, xsd:time,
xsd:gYear, etc..
has-temporal-potential to indicate data properties that potentially contain
temporal knowledge in an unstructured form, usually expressed using natural
language, for example but not limited to dcterms:title9 , dcterms:description10 ,
etc..
7
http://lov.okfn.org/dataset/lov
8
currently 603 vocabularies, last accessed 27.07.2017.
9
http://purl.org/dc/terms
10
http://purl.org/dc/elements/1.1/description
Table 1. Temporal knowledge in the Czech LOD (Cz LOD)
is-temporal has-temporal-potential
number of number of triples number of number of triples
properties properties
dcterms 12 7955431 2 7734043
schema.org 8 512781 4 149339
goodRelations 6 3110226 0 0
prov-o 2 146115 0 0
CzLOD 36 15268994 8 2597668
We experimented with the following vocabularies as they are most frequently
reused in the Czech LOD: Schema11 , dcterms12 , prov-o13 , and goodRelations14 .
An ontology combining these vocabularies augmented with is-temporal and has-
temporal-potential annotaion properties is available15 for further details. The
graph in figure 2 shows the overall number of annotated properties that describe
temporal knowledge in the selected common vocabularies. We notice that some
properties accept multiple data type ranges, schema:temporalCoverage for in-
stance, which accepts either DateTime, Text, or URL. Another observation is
that dcterms doesn’t have any explicitly temporal data type. Even when the
property is aimed to express only time or specific date, it is modeled to accept
general literal16 ranges.
3.2 Analysis of the nature of temporal data in Czech LOD
Czech Linked Open Data cloud (Czech LOD)17 contains dozens of datasets with
approx. 1 billion of records. In this work, as an attempt to investigate the nature
of the temporal knowledge, we manually experimented with a test set of 10
datasets found in Czech Linked Data Cloud involving datasets published by the
OpenData.cz initiative, to reveal the nature of the temporal knowledge taking
into consideration the variety of the topics and the contexts of the datasets.
Exploration of the temporal structure in the graph is done through a set of
SPARQL queries. We build the graph out of data properties used within the
datasets in a structured temporal form or the temporal knowledge found in
the strings. After analyzing the nature of the properties used in each dataset,
we recognize, in the same manner as the LOV exploration process, two types
11
http://schema.org/
12
http://purl.org/dc/terms/
13
http://www.w3.org/ns/prov\#
14
http://purl.org/goodrelations/v1
15
available at https://kbss.felk.cvut.cz/ontologies/dataset-descriptor/
temporal-properties-model.ttl
16
http://www.w3.org/2000/01/rdf-schema\#Literal
17
http://linked.opendata.cz/
of temporal knowledge, is-temporal denotes the properties that represent time
explicitly through a temporal data property, and has-temporal-potential which
denotes the properties that contains temporal knowledge in the form of string
literals. The latter type is where we will perform the analysis to reveal the hidden
temporal knowledge in the datasets. Table 1 shows the number of triples that use
these properties inside the datasets. We notice that several temporal properties
in the vocabularies are reused across the cloud.
The data inside the Czech LOD are heterogeneous. Triples contain string
literals in English as well as in Czech which makes the available temporal analysis
tools not efficient enough. Also, temporal information is varying from one dataset
to another. In one dataset, usually only structured temporal information or only
unstructured temporal information can be found. However, most of the datasets
have both types of temporal information.
Based on the analysis of the temporal information in the datasets, we could
recognize several forms. Time is mentioned explicitly or implicitly. Explicit time
mentions have two possibilities, an instant, which is a precise time point and
an interval, which is a time period with distinct start and end points. However,
instant form of temporal information can be understood as an interval with the
same start and end time. Implicit time information found in the datasets has the
form of mentions with a specific common well-known implicit temporal property
(Christmas, Mother’s Day, Independent Day). This information can still differ
according to the spatial information though.
3.3 Applying TIE tools on the unstructured temporal knowledge in
Czech LOD datasets
Temporal information extractors (namely SUTime and Heidel Time) are used
in order to extract temporal knowledge from the string literals in the datasets.
In the string literals, which are expressed in the natural language form, temporal
information does not follow any specific structure. For example, the temporal
expression ”the 20 century” is expressed in many other variants even within
the same dataset. We could find it in the written forms, ”Twenty century”, ”20
c”, and also ”20th-century”. Also, the date usually doesn’t follow one rule. For
example, dates are mentioned in different formats, ”DD MONTH YYYY”, ”DD
MM YY”, ”MM DD YY”, and so on. The granularity of the temporal information
found in the literals is varying from the very general information ”Century” and
going down to describe the fractions in specific ”minutes”. For example, ”during
the 20th century” and ”7,00 till 8,30 and from 10,30 till 15,30”. However, the
granularity of time information is poorly represented in the structured time
schema properties.
Another observation while analyzing the data is the problem of the relative
temporal information. For example, in the temporal mention found in one of
the string literals describing some event last year, the temporal data here has
uncertain implicit information, ”Which year”. It may be related to the year that
this dataset piece of information is speaking about, or it can be referring to
the creation date of the dataset itself. This information can only be clarified by
understanding the whole context of the dataset and making use of the structured
data and the data that was already detected in its graph.
4 Improving temporal knowledge representation of
datasets
4.1 Temporal information extraction and event extraction
In [BKK16], researchers describe an extension of the basic summary descriptor
that they call a temporal descriptor. It extracts temporal references from the
dataset and incorporate them into the s-p-o summary. They used only part of
the temporal knowledge extracted from some structured temporal data types.
They created a simple temporal descriptor that extracts xsd:date literals from
the dataset. A huge amount of other structured as well as unstructured temporal
information can be found in the cloud. If this information is not considered,
accuracy of temporal representation of the dataset will be reduced.
Though SUTime and Heidel time tools have many advantages and can detect
wide spectrum of temporal knowledge, we observed many lacks applying these
tools on the string literals. These lacks stem directly from the nature of the
temporal data within the datasets, as well as lacking the understanding of the
context around. Also, the data heterogeneity regarding the language used inside
the datasets. Samples of these lacks are mentioned in Table 2.
Table 2. Samples of the temporal information extraction lacking
Temporal Data Samples Timex3 value Lacking
from 1991 till 2006 1991&2006 the range
during 24 h No temporal expressions no detection
in the period of 17th - 20th 20th centuries the range
centuries
since its establishment in 2000 the range
2000 until today
30. September 2002 September 2002 day granular-
ity loss
2010 and further 2010 the range
in the 90s of the 20th c No temporal expressions no detection
until 1945 1945 the range
We overcome some of these limits by extending the rules in the rule files of the
extraction tools with more definitions to detect the temporal expressions that
were not possible to be detected by the default settings of the tools. This way, we
were able to increase the recall for the retrieved temporal knowledge. Following,
respectively, are samples of the rules that we created to spot expressions like {
05/11/1999 },{ 05-11-1999 }, and even the special cases to extract the year from
the data like {253/1992 Sb}.
{ruleT ype : ”time”, pattern : /dd/M M /yyyy/}
{ruleT ype : ”time”, pattern : /dd − M M − yyyy/}
{ruleT ype : ”time”, pattern : /[0 − 9]{1, 4}.?yyyy 0 Sb0 /}
The analysis of the data inside the datasets showed which features are needed
to describe the nature of the time information. For example, if this temporal
information is connected to a person, it means that it has a relationship, for
example (date of birth, date of death, etc.). This information is needed to un-
derstand the nature of the temporal knowledge, so not only detecting abstract
time intervals and instances is needed, but also the context of this temporal
information and what is the meaning of it which is an important information
to consider. This approach is designed to provide a better representation and
hence better exploration experience of the datasets temporally. To provide the
context to the temporal information extracted, we applied a knowledge-driven
event extraction from the sentences that contain temporal knowledge following
the approach in [SKVP05]. Finally, we formalize this extraction as a pair (e, t),
whereas e refers to the event extracted from the string literals and te is the
corresponding temporal information related to this event.
4.2 Temporal Descriptors Ontology
Based on the analysis in the previous section, in order to describe datasets based
on the temporal knowledge and achieve better representation of its temporal
dimension, we want to create an integrating ontology on the top of the common
vocabularies. To grasp the reality, Temporal Descriptor Ontology (TDO) is built
on the top of the Unified Foundational Ontology (UFO) [Gui05], which is one of
the top-level ontologies that has a good modeling language and it is supported by
several useful tools. UFO presents a level of abstraction that forms a perfect point
to start building the TDO ontology. Namely, UFO-B, which is used to model
the extracted events whether they are atomic or complex, the participants of the
events, and the time span the events occur within. UFO-B suggests that since
events happen in time, they are framed by a Time Interval [RdAFBG14] but
the provided representation of temporal knowledge is very limited.
At this point, we extend the UFO-B with the captured temporal data by
the OWL Time Ontology and connect it to the events. This provides a generic
framework for extracting the temporal information from the datasets. The full
definition of OWL-Time Ontology can be found in [HP04]. Here we present those
parts that are essential for capturing temporal information found in the datasets.
The basic structure of the ontology is based on an algebra of binary relations
on intervals developed by Allen and Ferguson [AF97]. Temporal knowledge is
represented by the class TemporalEntity. This class has only two subclasses,
:Instant and :Interval which axiomatize fundamental intuitions about timepoints
Fig. 3. Temporal Descriptor Ontology
(Instants) and time intervals (Intervals). Core parts of the TDO ontology are
shown in fig3.
(∀T )[T emporalEntity(T ) ≡ Interval(T ) ∨ Instant(T )] (1)
Time ontology also offers the class :DateTimeInterval which may be ex-
pressed using a single DateTimeDescription class. This class with its related
properties are perfect to define a higher level of granularity. For eg. day of the
week, or a year, using the TemporalUnit class. ProperInterval class is used to
represent proper intervals, which are intervals whose extremes are different. Ex-
ample in fig 4 shows the representation of an extracted event and it is temporal
interval for the year 1996 using the properties hasBeginning and hasEnd.
Fig. 4. Temporal Descriptor Ontology
5 Evaluation
Comparison of our approach to DCAT metadata and to the temporal
descriptors
DCAT vocabulary provides the property dct:temporal to annotate datasets with
temporal coverage. According to DCAT specifications, it describes The temporal
period that the dataset covers. We used this property, also the temporal descrip-
tors presented in [BKK16] to investigate the quality of the temporal metadata
based on comparison with the temporal information extracted from the actual
content of the datasets.
Dataset DCAT startDate DCAT endDate TD minDate TD maxDate AC minDate AC maxDate
ds:coi.cz/kontroly 2012-01-01 2014-12-31 2012-01-02 2014-12-31 2012-01-02 2014-12-31
ds:coi.cz/sankce 2012-01-01 2014-12-31 2012-01-01 2014-12-31 2012-01-01 2014-12-31
ds:currency 1999-01-04 2016-04-29 1999-01-04 2016-04-29
ds:drugbank 0002-02-26 2030-12-31 1633-01-01 2030-12-31
ds:legislation/psp.cz 0014-01-01 2099-01-01 0014-01-01 2099-01-01
ds:political-parties-cz 1928-04-12 2013-12-18 1928-04-12 2013-12-18
ds:seznam.gov.cz/agendy 2011-11-16 2030-12-31 1939-01-01 2030-12-31
seznam:objednavky 2014-01-01 2015-02-26 0201-10-07 2020-08-19 2000-01-01 2020-08-19
seznam:plneni 2013-03-26 2015-05-07 2013-03-26 2016-03-31
seznam:smlouvy 2014-01-01 2015-02-26 1992-08-25 2015-05-22 1945-01-01 2019-12-31
ds:vavai/evaluation/2009 2009-01-01 2009-12-31 1034-01-01 2030-12-31
ds:vavai/programmes 2015-03-30 2015-03-30 2003-02-12 2015-01-27 1959-01-01 2022-12-31
ds:vavai/research-plans 2015-03-30 2015-03-30 2003-01-01 2014-06-30 1890-01-01 2014-06-30
ds:vavai/tenders 2015-03-30 2015-03-30 1995-10-02 2016-04-01 1995-10-02 2022-12-31
ds:pravni-vztahy-listiny 2015-02-26 2015-02-26 1993-01-01 2015-07-16 1950-12-31 2015-07-16
Table 3. Comparison of temporal coverage by DCAT metadata, temporal descriptor
(TD), and the actual content (AC ) temporal representation computed using our ap-
proach. Missing temporal DCAT metadata are indicated by empty cells within 2nd and
3rd column. Empty cells in the 4th and 5th column might indicate either missing data
or incomplete descriptor computation procedure. Complete computation of the temporal
coverage can be found in the 6th and 7th column.
We compute the temporal representation of the datasets in the Czech cloud
using the approach we presented in the previous section. We compute the tempo-
ral scope of the actual content of the triples in the datasets. Table 318 shows the
overall temporal coverage datasets in the Czech cloud that have both temporal
descriptor representation as presented in [BKK16], computed by our approach
and compared to the temporal coverage computed by DCAT and the temporal
descriptors. Each line represents a dataset in the Czech cloud. Second and third
column represent temporal coverage defined by DCAT property dct:temporal19 .
Next two columns represent minimal and maximal computed date by temporal
descriptor. Last two columns represent minimal and maximal temporal extrac-
tion that we computed from the actual content of the dataset.
18
Each dataset in the table prefixed with “ds” representing URL http://linked.
opendata.cz/resource/dataset/; Each dataset in the table prefixed with “sez-
nam” representing URL http://linked.opendata.cz/resource/dataset/seznam.
gov.cz/rejstriky/.
19
http://purl.org/dc/terms/temporal
We can notice that for this test pad of 15 datasets, computed temporal
representation using our approach is compatible with the temporal descriptors
for 46.6% of the cases. For 46.6% of the cases, they differ due to the fact that our
approach take the unstructured temporal information into consideration, while
the temporal descriptors cares only about the temporal meta-data in the dataset.
For the same reason, the dataset ds:vavai/evaluation/2009 doesn’t have any
temporal descriptor representation while we are able to compute the temporal
coverage for this dataset.
Next, We extend the experiments and we utilize our approach to compute the
temporal description of all the datasets available in the SPARQL endpoint of the
Czech cloud which contains 76 datasets20 . We are able to augment the temporal
representation for 57.89% of the datasets. The rest of the datasets doesnt have
any temporal knowledge in their resources to be extracted. Table 421 shows the
computed temporal coverage (minimum and maximum date extracted) of each
dataset in the cloud compared to DCAT temporal coverage when available.
Dataset DCAT startDate DCAT endDate AC minDate AC maxDate
ds:ic 2015-01-01 2015-02-26 1972-01-01 2013-12-20
ds:mfcr/ciselniky 2010-01-01 2014-12-31 1900-01-01 9999-12-31
ds:coi.cz/zakazy 2012-01-01 2014-12-31 1986-01-01 2001-01-01
ds:vavai/evaluation/2013 2013-01-01 2013-12-31 1277-01-01 2016-12-31
ds:sukl/drug-prices 2012-01-01 2015-07-29 1990-01-01 2016-04-20
ds:cenia.cz/irz 2015-02-01 2015-03-31 2004-01-01 2012-12-31
ds:vavai/funding-providers 2015-03-30 2015-03-30 1996-12-31 2007-06-01
ds:vavai/evaluation/2011 2011-01-01 2011-12-31 1100-01-01 2050-12-31
ds:vavai/evaluation/2010 2010-01-01 2010-12-31 0900-01-01 2050-12-31
ds:check-actions-law 1945-10-27 2013-12-31
ds:vavai/results 2015-02-26 2015-02-26 1970-01-01 2020-12-31
ds-external:pomocne-ciselniky 1988-01-01 2010-12-31
ds:vavai/evaluation/2012 2012-01-01 2012-12-31 1100-01-01 2050-12-31
ds:vavai/projects 2015-03-30 2015-03-30 1100-01-01 2050-12-31
ds-external:check-actions 1992-01-01 2016-08-01
ds:it/aifa/drug-prices 2012-01-01 2015-07-31 2015-07-15 2015-07-15
ds:/court/cz 2013-06-17 2013-06-17
ds:nci-thesaurus 2013-03-25 2013-05-15
ds:obce-okresy-kraje 2012-09-05 2012-09-05
ds:cpv-2008 2008-01-01 2008-01-01
ds:spc/ai-interactions 2013-05-15 2013-05-15
ds-external:nuts2008/ 2008-01-01 2011-12-31
ds-external:geovoc-nuts 2013-01-04 2013-01-04
ds:dataset/fda/spl 2013-15-05 2013-15-05
ds:vavai/cep 1141-01-01 2020-12-31
ds:regions/momc 2014-02-24 2014-02-24
ds:buyer-profiles/contracts/cz 2000-01-01 2024-12-31
ds:legislation/nsoud.cz 2004-01-27 2014-02-27
ds-external:souhrnn-typy-ovm 1969-01-01 2012-12-31
Table 4. The complementary comparison of the temporal coverage by DCAT meta-
data, to the actual content temporal representation computed using our approach for
the rest of the datasets in the Czech cloud. Missing temporal DCAT metadata are
indicated by empty cells within 2nd and 3rd column.
20
last access 2017-08-05
21
All the dates are normalized to a day granularity.
The dataset ds:vavai/evaluation/2011 contains data about events that
started during the 12th century (eg. Pilgrimage element in crusades with Czech
participation in the twelfth century). For that the actual data coverage starts
at 1100-01-01. On the other hand, the dataset ds:vavai/projects maximum
coverage date is 2050-12-31 (Prediction processing of systems utilizing renewable
energy sources in Czech republic till 2050).
6 Conclusion and Future Work
In this work, we proposed an approach to improve the temporal knowledge repre-
sentation of the datasets. We first experimented with the most commonly reused
vocabularies in the Linked Open Data Vocabulary (LOV), where we proposed the
notions is-temporal and has-temporal-potential, then we used the same notions
to discover the nature of the temporal data in the Czech linked open data cloud
as a use case. We proposed to include the temporal knowledge extracted from
the textual representation of some properties in describing datasets temporally.
Then, Temporal Descriptor Ontology was built on the top of UFO ontology and
OWL Time Ontology, utilizing the notions proposed and the temporally anno-
tated properties to provide a better representation for datasets on the temporal
dimension. This approach allows representing datasets based on their temporal
knowledge time-line. It also improves the metadata quality, allowing the auto-
matic usage of datasets for searching and integrating tasks.
As a future work, we are planning to experiment with more vocabularies in
LOV to widen the choices of the temporal properties aligned with the Temporal
Descriptor ontology.
Also, timeline visualization is an important tool for sensemaking. It allows
analysts to examine information in chronological order and to identify temporal
patterns and relationships [NXWW14]. Timelines also reduce the chances of
missing information and they facilitate spotting anomalies. Currently, we are
working on utilizing the computed indices of each dataset aligned with the TDO
representation to represent the datasets’ triples chronological on a timeline for
easy exploration of the datasets temporally. We assume a discrete representation
of time with a single day being the atomic granularity. This single granularity can
be grouped into larger units like weeks, months, years, decades, and centuries.
The base timeline can be defined on any other atomic time interval, such as an
hour (of a day), a specific minute, or even seconds. The representation depends
on the characteristics of the triples collection in the dataset to be time annotated
and do not have effect on the generality of the proposed approach thanks to the
TDO representation.
At its core, this visualization is a typical interactive slide-timeline presenta-
tion uses single-frame interactivity, meaning that interaction manipulates items
within a single-frame without taking the user to new visual scenes. It is aug-
mented by two important features. First, it allows the user to determine the
pace of the presentation by using the provided progress bar. Second, it allows
the user to interact with the presentation by hovering areas of interest and by
using the slider to explore different time windows.
Acknowledgments. This work was partially supported by grants No. GA 16-
09713S Efficient Exploration of Linked Data Cloud of the Grant Agency of the
Czech Republic and No.SGS16/229/OHK3/3T/13 Supporting ontological data
quality in information systems of the Czech Technical University in Prague.
References
[AF97] James F. Allen and George Ferguson. Actions and events in interval
temporal logic. Spatial and Temporal Reasoning, page 205245, 1997.
[BKK16] Miroslav Blaško, Bogdan Kostov, and Petr Křemen. Ontology-based
Dataset Exploration – A Temporal Ontology Use-Case. In INTELLI-
GENT EXPLORATION OF SEMANTIC DATA (IESD 2016), Kode,
2016.
[BLN11] Christoph Bohm, Johannes Lorey, and Felix Naumann. Creating voiD
descriptions for Web-scale data. Journal of Web Semantics, 9(3):339–345,
2011.
[CDJJ14] Ricardo Campos, Gaël Dias, Alı́pio M Jorge, and Adam Jatowt. Sur-
vey of Temporal Information Retrieval and Related Applications. ACM
Comput. Surv., 47(2):15:1—-15:41, 2014.
[CM12] Angel X Chang and Christopher D Manning. SUTime: A library for
recognizing and normalizing time expressions. Lrec, (iii):3735–3740, 2012.
[DHHN16] Peter David, Timothy Hawes, Nichole Hansen, and James J Nolan. Con-
sidering context: reliable entity networks through contextual relationship
extraction. In SPIE Defense+ Security, pages 985107–985107. Interna-
tional Society for Optics and Photonics, 2016.
[Gui05] Giancarlo Guizzardi. Ontological Foundations for Structural Conceptual
Model, volume 015. 2005.
[HKL13] Siaw Nyuk Hiong, Narayanan Kulathuramaiyer, and Jane Labadin. NAT-
URAL LANGUAGE SEMANTIC EVENT EXTRACTION PIPELINE.
(063):333–339, 2013.
[HP04] Jerry R Hobbs and Feng Pan. An Ontology of Time for the Semantic
Web. ACM Transactions on Asian Language Information Processing,
3(1):66–85, 2004.
[KDF+ 13] Aleksandar Kova, Azad Dehghan, Michele Filannino, John A Keane, and
Goran Nenadic. Combining rules and machine learning for extraction of
temporal expressions and events from clinical narratives. pages 859–866,
2013.
[KEB07] Anup Kumar Kolya, Asif Ekbal, and Sivaji Bandyopadhyay. A Super-
vised Machine Learning Approach for Temporal Information Extraction
3 Conditional Random Field Based Approach. pages 447–454, 2007.
[LCB13] Yu Kai Lin, Hsinchun Chen, and Randall A. Brown. MedTime: A tem-
poral information extraction system for clinical narratives. Journal of
Biomedical Informatics, 2013.
[LCH+ 13] Xuefei Li, Hongyun Cai, Zi Huang, Yang Yang, and Xiaofang Zhou.
Spatio-temporal Event Modeling and Ranking, pages 361–374. Springer
Berlin Heidelberg, Berlin, Heidelberg, 2013.
[Lim16] Chae Gyun Lim. A survey of temporal information extraction and lan-
guage independent features. 2016 International Conference on Big Data
and Smart Computing, BigComp 2016, pages 447–449, 2016.
[MPAGS17] Albert Meroño-Peñuela, Ashkan Ashkpour, Christophe Guéret, and Ste-
fan Schlobach. Cedar: the dutch historical censuses as linked open data.
Semantic Web, 8(2):297–310, 2017.
[NXWW14] Phong H Nguyen, Kai Xu, Rick Walker, and BL William Wong. Schema-
line: timeline visualization for sensemaking. In Information Visualisation
(IV), 2014 18th International Conference on, pages 225–233. IEEE, 2014.
[PCI+ 03] James Pustejovsky, Jose Castano, Robert Ingria, Roser Sauri, Robert
Gaizauskas, Andrea Setzer, Graham Katz, and Dragomir Radev.
TimeML: Robust Specification of Event and Temporal Expressions in
Text. New directions in question answering, 3:28–34, 2003.
[RdAFBG14] Fabiano Borges Ruy, Ricardo de Almeida Falbo, Monalessa Perini Bar-
cellos, and Giancarlo Guizzardi. An ontological analysis of the iso/iec
24744 metamodel. In FOIS, pages 330–343, 2014.
[RMB16] Anisa Rula, Andrea Maurino, and Carlo Batini. Data Quality Issues in
Linked Open Data, page 87112. Springer International Publishing, Cham,
2016.
[RPNN+ 14] Anisa Rula, Matteo Palmonari, Axel-Cyrille Ngonga Ngomo, Daniel Ger-
ber, Jens Lehmann, and Lorenz Bühmann. Hybrid Acquisition of Tem-
poral Scopes for RDF Data, pages 488–503. Springer International Pub-
lishing, Cham, 2014.
[SG10] Jannik Strötgen and Michael Gertz. HeidelTime: High quality rule-based
extraction and normalization of temporal expressions. Proceedings of
the 5th International Workshop on Semantic Evaluation, (July):321–324,
2010.
[SKM11] Jason Switzer, Latifur Khan, and Fahad Bin Muhaya. Subjectivity clas-
sification and analysis of the asrs corpus. In IRI, pages 160–165. IEEE
Systems, Man, and Cybernetics Society, 2011.
[SKVP05] Roser Saurı́, Robert Knippen, Marc Verhagen, and James Pustejovsky.
Evita: A robust event recognizer for qa systems. In Proceedings of the
Conference on Human Language Technology and Empirical Methods in
Natural Language Processing, HLT ’05, pages 700–707, Stroudsburg, PA,
USA, 2005. Association for Computational Linguistics.
[TWJ+ 13] Buzhou Tang, Yonghui Wu, Min Jiang, Yukun Chen, Joshua C Denny,
and Hua Xu. A hybrid system for temporal information extraction from
clinical text. Journal of the American Medical Informatics Association :
JAMIA, 20(5):828–35, 2013.
[ULD+ 13] Naushad UzZaman, Hector Llorens, Leon Derczynski, Marc Verhagen,
James Allen, and James Pustejovsky. Semeval-2013 task 1: {T}empeval-
3: Evaluating time expressions, events, and temporal relations. Sec-
ond joint conference on lexical and computational semantics (* SEM),
2(SemEval):1–9, 2013.
[Wor14] World Wide Web Consortium. Data Catalog Vocabulary (DCAT). W3C,
(January), 2014.
[WZQ+ 10] Yafang Wang, Mingjie Zhu, Lizhen Qu, Marc Spaniol, and Gerhard
Weikum. Timely yago: harvesting, querying, and visualizing temporal
knowledge from wikipedia. In Proceedings of the 13th International Con-
ference on Extending Database Technology, pages 697–700. ACM, 2010.