Temporal data, i.e., time-annotated data in the RDF data model expressing the valid time of a Resource, Property or Relationship can be stored in various forms. However, the modeling of temporal data and how properties that describe them are named heavily depends on the creators' preferences. Thus, to utilize the temporal axis in SPARQL queries the user needs intimate knowledge of the data source and preferably of the underlying ontology. This paper covers the analysis of temporal data usage in various open RDF data sources, especially in the LOD cloud, and proposes a framework to automatically identify temporal data facts in unknown data sources by semantical and structural analysis. The temporal data is then made available through a temporal extension of the SPARQL query language which allows uniform temporal query patterns regardless of its specific underlying implementation.
CEUR ceur-ws.org
ral information is heavily dependent on the data model.
the RDF data model, one runs into its inherent limitaTemporal data enables researchers and practitioners to highly researched topic for every emerging data model. tions. The data model of standard RDF is limited to triples , , (s,p,o), or equivalently, binary regain valuable insights into the evolution of events, behav- lations where each describes the relationship between a iors, and processes over time. The interest in temporal data is so great that special temporal databases were developed to better facilitate the analysis of time-series data like temporal-spatial data.
But, the term temporal data can mean diferent things the temporal information it can be classified into several in the context of databases. Depending on the purpose of , , , distinct sub-categories. For example, one purpose is the tracking of transaction times, which concerns itself with the temporal logging of the diferent states a fact has resource and either a literal or another resource. It it is indeed possible to put a resource itself on the temporal axis by simply describing it with its properties, and annotating these property values and relationships with temporal information. However, this requires – at least– quadrupels or quintuples , , , ,
. Quadrupels are basically present in RDF when named data graphs are considered to represent snapshots or intervals. Then, a temporal dataset consists of several such graphs. Nevertheless, the most common way to compensate for these or had inside the database management system (DBMS). limitations is to use reification to represent temporally These cover temporal information such as creation, mod- annotated data in a single graph. Reification means, to ification and deletion timestamps. Another purpose is the valid time, which describes when a fact is considered true outside of the database. This could be for example the time interval when a person lived or the time point when an event happened. The processes and framework described in this paper are specifically concerned with temporally annotated data in the context of valid time.
While transaction times can be useful information for
database admins and for stream processing, semantical
queries usually consider valid time information. When
utilizing this type of information the user can query the
database facts with constraints on the temporal axis in
the context of the real world.
dard like RDF-star [
Besides the structural challenges of temporal valid data in RDF, there exist also semantical ones. The names of properties that describe the valid time of a resource can also be very diverse, because languages developed diferent words to describe the beginning and end of things depending on the semantical context. For example a person’s beginning is pretty likely to be described as their “birth” and the ending as their “death”, while a building their temporal values as is the case when a property has its “construction” and “demolition” respectively. Us- ranges over a time-expressing datatype (XSD or othing these words as property names help the human user erwise) or indirectly if they are further encapsulated to deduce the semantic meaning of its property value through a time-describing resource. This is often the after the discovery, but also limit the exploration of algo- case if the relevant time frame of the value is not a sinrithms that do not comprehend this human knowledge. gle time point, but a time interval. For simplicity, in the
The structural and semantic discrepancies of tempo- scope of this paper for all further mentions of time valral data between various RDF data sources requires a ues, it should be assumed that a mapping is used that new user to gain an intimate understanding of the in- converts the encapsulated indirect connections to direct ner workings of the data set before being able to query ones and that they are in a format that conforms to an temporal data eficiently. Data integration and queries XSD datatype. against several data sources, e.g., in LOD, even require to use diferent modelings in a single task. This is especially unpleasant because the RDF data model was designed to make data readily available in the Web and encourages the exploration of unknown data sets through links as is demonstrated through the Linked Open Data (LOD) initiative. To lighten the burden of data exploration from the user this paper covers a framework proposal that aims to automatically identifying temporal data in a new data set and allows uniform temporal query patterns regardless of its specific underlying implementation.
Related Work. Implementing and querying temporal
information in RDF data has been a long-time research
topic [8]. In [9], Tappolet et al. introduce -SPARQL
based on named graphs to store temporal data and
temporal wildcards [?s, ?e] that bind to the respective
underlying named graph(s) to bind the temporal context
of triples. Grandi [10] proposed a variation of temporal
SPARQL named T-SPARQL that aims to adapt features
from the TSQL2 [11] temporal query language for
relational databases. The query language works on an
underlying multi-temporal RDF database model consisting
2. Preliminaries of multiple time domains. The triples in the WHERE
clause are extended with an optional fourth position
In addition to the structural and semantic dificulties of ?s ?p ?o | ?t where ?t binds the complex timestamps.
temporal data, time values in RDF datasets can also be With the temporal SPARQL implementation in [12]
Roexpressed in various forms. The standardized way is to batjazi highlights the advantages of hiding the intricate
use the RDF-compatible XSD datatypes defined in XML n-ary relationship structures from the user by
providSchema [
The identification of the semantical meanings/senses manual identification. A circumstance we aim to address of a word is the core topic in the field of word-sense in- with this work. Table 1 contains our findings. duction. The general consensus in linguistics is that the context of a word is tremendously important to under- Table 1 stand its meaning [14, 15]). Arora et al. [16] have shown Occurrences of time related datatypes, classes and properties that word embeddings produced even from simple mod- in open RDF datasets els like word2vec [17] capture the meanings of words as long as the context was used to train them. Further XSD datatypes Frequency in % research results [18, 19] indicate that the inclusion of additional sense information for example from BERT [20] sense embeddings or semantic networks like WordNet [21] improve disambiguation tasks. dateTime date gYear gYearMonth gDay / gMonth
TIME ontology classes
Frequency in % Instant
Interval
TemporalEntity DateTimeDescription
DateTimeInterval To verify the results of [13] and to check the current status of temporal data usage in open RDF sources, we conducted our own analysis. Given that the structure of triples is not suficient to clearly identify them as temporal data, we look for further characterizing features. The obvious solution would be to search and employ the spe- DCMI time properties Frequency in % cialized properties and classes in standardized ontologies dcterms:date 55 that denote some type of temporal annotation. Thus we dc:date 22 focus on the adoption rate of these standardized notions dcterms:temporal 7.4 to determine whether they are suficiently utilized to be dcterms:valid 6.5 reliable in the identification process. Again, we were first of all interested how valid times are described and stored in a wide variety of topics. Reification Method Frequency in %
For that the analysis utilizes first and foremost the rdf:Statement 8.3 data sources listed in the LOD cloud [22], which cover diferent disciplines from geographical and governmen- From the results of the analysis it can be projected that tal data to life science and music. In addition to that a a significant portion of open RDF datasets utilize at least selection of smaller and bigger open RDF data sources the more general time properties of DCMI and standard found by search engines were also included. To focus XSD datatypes. Furthermore the OWL-Time ontology on more recent and active sources and in compliance was adopted at least partially in more datasets as one with the requirements of high quality open RDF data, might suspect given its niche application. In contrast, we specifically searched for data sources that feature a the more specialized datatypes of singular time units and functioning SPARQL service. At the time of the analysis properties to explicitly describe valid time are rarely seen. the LOD-cloud featured 1300 RDF datasets with over 450 One needs to keep in mind that not all disciplines have registered SPARQL endpoints. Unfortunately of those a use for valid times or even time values in general, but only just over 100 were able to answer a simple query. it seems likely that most datasets containing valid time This prompted the inclusion of further RDF dataset repos- facts apply their own custom property names to describe itories for example Govdata.de [23] and data.europa.eu them. In addition to that we found that only 8.3% of [24] for governmental data. the datasets utilize the rdf:Statement, which matches the
We tested the datasets on the occurrence of XSD time- findings of [ 13] and the conclusion that reifications are related datatypes, OWL-Time ontology classes and DCMI done more likely informally as n-ary relationships. time-related properties. Additionally we tested the usage As a result we determined that we can not rely on these of the rdf:Statement class to check if reifications were standardized notions for a comprehensive identification done strictly formally. Of course it would also be inter- of temporal data. The temporal annotation of data in RDF esting to analyse how often custom n-ary relationships is still a highly varied procedure that heavily depends on the personal preferences of its creator. Therefore a suitable automatic identification framework must cater to these custom characteristics as well as possible. 4. Proposed Framework prop, value] statement as shown in figure 2. These could be for example the temporal annotations P585 (point in time), P580 (start time) and P582 (end time). It should be noted that in addition to the reification there exists almost always a direct property wdt:prop to the same value as the path p:prop / ps:prop.
Employing the advances in programmatic processing of word senses, the foundation of the proposed framework 4.1. Structural analysis for identifying temporal data in unknown sources is a The main purpose of the structural analysis is to reduce two-step analysis. First a structural analysis on graph the search space for temporal facts before applying the patterns is done to create a general set of potential tem- semantical analysis step. In addition to that, it allows to poral data candidates. Then a semantical analysis of the gain a higher-level understanding of the dataset in case property names being used in the candidates to perform an OWL ontology is not provided and may even uncover the actual classification follows. For this crucial second relations between properties that are not covered by it. step, a neural network is employed that is trained on a The common consensus is that there are four diferent corpus of English words that appear in the context of ways to annotate triples with temporal information in temporal values. The goal of the supervised learning standard RDF. They range from more straightforward apapproach is to create a classifier that can diferentiate proaches like singleton properties (construct new properwords that are typically used to describe temporal data. ties) to more abstract approaches like standard reification In the beginning we focus on property names that denote (construct new auxiliary resources), n-ary relationships the temporal sense of valid times, but the process can (introduce new resources, properties and classes) and be extended to also include other temporal senses, for even named graphs (construct new RDF graphs). example active times. Regardless of their specific tem- However, we want to utilize a higher level view on the poral sense, properties that are thus classified by their structure of temporal data that focuses more on the purname are in the following generally referred to as time pose of the temporal annotation, especially what is being properties. annotated. This view centers around what a user would
After running the identification process on a dataset, it want to express in queries and illustrates the detachment is planned to use the classified temporal data fragments from the specific underlying structural implementation. to create a dictionary which stores the relevant time Additionally, it introduces the necessary flexibility to acproperties for each resource. Through further heuristical commodate new structural extensions to RDF like RDFexamination of the findings and its surrounding graph star later on by adding appropriate transformations to the patterns, it is aimed to generalize the dictionary from unifying view. In the confines of this paper it should be specific resources to general classes. Supplementing an assumed that there already exist suitable transformations OWL ontology for this step can help with the generaliza- for every mentioned standard structure for temporal antion efort. Ultimately this dictionary can then in turn be notations even though only the ones found in Wikidata deployed as the backbone for a temporal SPARQL imple- will be represented in the examples. Thus we determined mentation that allows the user to query the dataset with- the following three diferent kinds of temporal data. out knowing its specific time properties by providing a lookup opportunity for the returned variable bindings. A typical sketch of the consolidation of the SPARQL query 4.1.1. Temporal Resources is shown in Figure 1. The first and simplest form of temporal data are temporal
To illustrate the details of each process, they will be resources. These express the validity of the resource itself, explained on the basis of one of the most prominent open so for example the birth and death dates of a person or RDF data sources, Wikidata [25]; more specifically on the time point an event happened. As a standard graph the resource that describes the Entity “Germany”1. To pattern they are not diferent from any other property maintain readability the common Wikidata prefixes 2 will that is being used to describe the resource. Thus any be used. Notably the wdt: prefix for properties with di- pattern rect →− connections and the p:, ps: & pq: prefix used for reification. Where p:prop is the connec- < r e s o u r c e > ? t i m e − p r o p e r t y ? t i m e tion from the resource to the auxiliary reified resource, ps:prop the simple key value and pq:prop any number of qualifier properties that further describe the [ resource, where ?time is a temporal value, is a potential candidate.
There should be only a single property for each resource which denotes its valid time if it is a time point, or a pair of properties for intervals, denoting their begin and end.
In the case of Wikidata’s Germany this would match the properties schema:dateModified , wdt:P571 and < r e s o u r c e > ? p r o p e r t y [ wdt:P1249 as temporal data candidates. The subsequent ? t i m e − p r o p e r t y ? t i m e . semantic algorithm would then classify wdt:P571 as the ? v a l u e − p r o p e r t y ? v a l u e ] correct time property. A representation of the triple is with ?time a temporal value and ?value the literal value shown in Figure 3. For some entities it can be possible of the original property before annotation. The names that multiple entries of its time property exist. This can of both ?property or ?value-property are likely to stand make sense for more conceptual entities that can stop in relation to the orignal property name, but are not existing for a while or change their nature, which is the guaranteed to be. In addition the reification could be case for the entity Germany. This forces the introduc- done as a blank node or as an auxiliary resource with tion of reified resources representing each period of its or without a proper class type that may give further valid time. Luckily Wikidata uses the same property semantical hints for its purpose. names for the reified resources, but with diferent pre- Naturally this simple graph pattern would match a ifxes. The property path p:P571/ps:571 then points to the huge number of data fragments that are no temporal relevant time values. wdt:P571 can be seen more as a data. Further heuristical filtering must be applied to direct shortcut to the most recent value for the property reduce the number of generated candidates. For examP571 (inception). In most datasets this shortcut to the ple a very efective condition would be that no other most recent valid time will not exist, but nevertheless an resources reference the reification. An obstacle so far interesting consideration must be made when querying is the identification of the correct value-property if the the valid time of resources with multiple valid intervals. reification features multiple properties for even more deDoes the user want to consider every valid interval or tailed descriptions. It could be argued that every property only the most recent one? grouped by the reification is temporally annotated by the time-property, but the user is likely to be more interested countries to organizations. Germany is a member of the in the value of the original property. European Union wd:Q458 with “start time” P580 ’1957
In Wikidata the “population” property is called P1082. 03-25’. While there exist temporal relationships from the Each occurrence of p:P1082 points to a reified resource EU to Germany (“has part(s)” and “contains the adminwith an internal name, which in turn groups among oth- istrative territorial entity”), they do not share the same ers the property ps:P1082 describing the value of the temporal reification. Nevertheless the possible existence population and the property for “point in time” pq:P585 of inverse properties means that the condition outlined describing the valid time of the value. Figure 4 shows a to reduce the number of false temporal candidates needs representation of these data facts for Germany. to be adjusted.
Again, just like with temporal literals, there might exist a shortcut property which connects the two resources without the temporal information, which in this example is wdt:P463 as shown in Figure 5. The same benefits apply for temporal relationships too.
Furthermore if there are multiple time-annotated instances of the original property, for example multiple population values for each year a population census occurred, then the dataset might also feature a shortcut property to the most recent value. Most datasets do not include these direct properties like wdt:P1082, but if they exist, they can be beneficial in various ways. First of all identifying them would allow to create a mapping so the user can utilize either in the queries. A heuristical approach for this could not rely on the thoroughly planned naming scheme of Wikidata, but would need to check for property pairs that most of the time appear together with one pointing to a (set of) temporal reification(s) and the other pointing to a literal that is also the value occurring in the newest instance of those reifications. This would also help to indicate which property is the correct value-property inside the temporal reification.
but allowing the value of the temporally annotated property to be a resource. This opens up new possibilities how the temporal reification is modeled. A fully connected temporal reification would feature properties both from and to the resources which the relationship describes. 4.2.1. Network development This is rarely the case so that only one of the directions is present for each resource. The most common structure is similar to temporal literals with one resource pointing to the temporal reification which itself points towards the related resource as its value. An example for this is the “member of” P463 property in Wikidata relating
the construction of a simple network that should just classify whether an English word appears in the context of valid time. This network is then expanded and tweaked until a satisfactory accuracy is achieved to serve as a proof of concept. 4.2. Semantic temporal words The structural analysis of the dataset through graph patterns can only generate temporal data candidates. Any optimization done on the patterns and further heuristic analysis on structural information may reduce the number of generated false positives, but does not remove them entirely. There will always be data fragments that match the pattern, but are no true temporal data. A human on the other hand can discern which properties are used to describe valid time by interpreting the meaning of their name in the context of the resources’ real-world nature.
The main building block of the proposed framework aims to harness this semantical knowledge to obtain an accurate classifier for the produced candidates. To do this, a machine learning approach has been chosen that is trained with manually labelled property names extracted from a wide variety of open RDF datasets. The goal is that the network will be able to label unknown words even from topics it was not trained upon.
Then, the class of “valid time” property names is split a combination of multiple words for example startTime into more refined categories that denote the occurrence or productionDate. Therefore it was crucial to choose a of a time point, the beginning of a time interval, or the model that is able to produce vector embeddings for conending of a time interval. Formally it could be argued catenated words like fastText [27], developed from the that any time point is a time interval of minimum du- Facebook AI Research (FAIR) lab. It provides models for ration and thus a time point property could also be the either skipgram or cbow and can produce embeddings beginning (or end) of an interval. Semantically, and for even for unknown words through the use of subword inthe actual meaning of the words to be detected, the dif- formation. Utilizing subword information should enable ference between events and time intervals is crucial and fastText to produce meaningful embeddings especially helpful, e.g., the events of the signature and the notice of for concatenated words by factoring in every single comtermination of a rental contract, and the actual duration. ponent word. However, for getting realistic results the Also, for the execution of temporal queries it is of vital property names need to be words that at least resemble importance to estimate the boundaries of the resources’ correct english dictionary words. As in the Wikidata valid time, including intervals that have a beginning, but case, this is sometimes not directly given. There, to creno end (so far). Semantically, finding out which time ate a unifying model across all languages, every resource point and/or time frame properties relate to each other and property is tokenised into an alphanumerical code, should help in the data exploration. In case multiple time like the above P1082 for the “population” property. The values are present it is not always prudent to take the proper property names are then stored as an rdfs:label for minimum and maximum values as the start and end time each language. For cases like these, the learning process respectively (cf. the rental contract case). uses the corresponding label, preferrably in the English
Because these words rely so heavily on the context language (@en), instead of the tokenised property name. they are used in, the ultimate goal would be that the network can be expanded so far that it not only takes the words, but also for example the classes of the resource 5. Conclusion into consideration. It is suspected that training relations like ’birth for a person is like construction for a building’ would boost the performance significantly. More intricate is the case of a rental contract, which is a temporal object whose existence starts with the signature, and it describes the property “being rented” for another (maybe ongoing) interval, and its existence may bind the contractors even longer. For this, we extracted so far over 6000 property names with classes from the associated context that were found to be related to time values. We manually labelled them accordingly for [valid timepoint] [valid time begin] [valid time end] to perform the initial training of the neural network.
In this paper we have discussed issues of modeling temporal data in RDF. We have analyzed how temporal data is represented structurally in real-world RDF data, and we have provided an approach for classifying properties describing the valid time of data facts by the semantical meaning of their names. With this, temporal data can be detected in unknown data sources.
The second part of the project will exploit the lifting of temporal data to a conceptual view on its modeling for designing a temporal SPARQL query extension which provides generic temporal constructs instead of querying the exact (temporal) RDF triples. With this approach the user will be able to query the specific purpose of the temporal annotation, beginning with valid times but extensible for further purposes like active times.
be first converted into numeric vector representations.
To realistically achieve any of the previously mentioned goals it is paramount that the words are embedded to vectors in such a way that the semantical meaning of them is conserved. We are particularly interested in an embedding model that keeps the relationship between words intact, so that for example the vector pointing from “human” to “birth” might relate to the vector from “building” to “construction”, and from “contract” to “signature”.
Much research has been done on this topic, which is often covered under the generic term “Word2Vec”.
Some well known approaches include skip-grams [26] and CBOWs (continuous bag of words). One characteristic of real-world property names is that they are often [5] J. F. Allen, G. Ferguson, Actions and Events estimation of word representations in vector space, in Interval Temporal Logic, Springer Nether- 2013. arXiv:1301.3781. lands, Dordrecht, 1997, pp. 205–245. URL: [18] S. Mizuki, N. Okazaki, Semantic specialization https://doi.org/10.1007/978-0-585-28322-7_7. for knowledge-based word sense disambiguation, doi:10.1007/978- 0- 585- 28322- 7_7. arXiv preprint arXiv:2304.11340 (2023). [6] J. F. Allen, Towards a general theory of ac- [19] A. Saha, A. Gittens, B. Yener, Word sense induction and time, Artificial Intelligence 23 (1984) tion with knowledge distillation from bert, arXiv 123–154. URL: https://www.sciencedirect.com/ preprint arXiv:2304.10642 (2023). science/article/pii/0004370284900080. doi:https: [20] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, //doi.org/10.1016/0004- 3702(84)90008- 0. Bert: Pre-training of deep bidirectional transform[7] Dublin Core DCMI, Dublin Core: Dcmi ers for language understanding, arXiv preprint metadata terms, 2020. Www.dublin- arXiv:1810.04805 (2018).
core.org/specifications/dublin-core/dcmi-terms. [21] WordNet, Princeton university ”about word[8] C. Gutierrez, C. A. Hurtado, A. Vaisman, In- net.” wordnet. princeton university., 2010. Wordtroducing time into rdf, IEEE Transactions on net.princeton.edu.
Knowledge and Data Engineering 19 (2007) 207–218. [22] LOD-cloud Website, LOD cloud: The linked open doi:10.1109/TKDE.2007.34. data cloud monitors and provides links to datasets [9] J. Tappolet, A. Bernstein, Applied temporal rdf: Ef- available as open linked data, 2019. Lod-cloud.net. ifcient temporal querying of rdf data with sparql, [23] Das Datenportal für Deutschland, Open Governin: L. Aroyo, P. Traverso, F. Ciravegna, P. Cimi- ment : Verwaltungsdaten transparent, ofen und ano, T. Heath, E. Hyvönen, R. Mizoguchi, E. Oren, frei nutzbar, 2023. Www.govdata.de. M. Sabou, E. Simperl (Eds.), The Semantic Web: [24] data.europe.eu Website, data.europa.eu : The ofiResearch and Applications, Springer Berlin Heidel- cial portal for european data, 2023. Data.europa.eu. berg, Berlin, Heidelberg, 2009, pp. 308–322. [25] Wikidata Website, Wikidata: multilingual knowl[10] F. Grandi, T-sparql: A tsql2-like temporal query edge graph hosted by the wikimedia foundation, language for rdf., in: ADBIS (local proceedings), 2023. Www.wikidata.org.
2010, pp. 21–30. [26] D. Guthrie, B. Allison, W. Liu, L. Guthrie, Y. Wilks, [11] R. T. Snodgrass, The TSQL2 temporal query lan- A closer look at skip-gram modelling, in: LREC, guage, volume 330, Springer Science & Business volume 6, 2006, pp. 1222–1225.
Media, 2012. [27] T. Mikolov, E. Grave, P. Bojanowski, C. Puhrsch, [12] M. RobatJazi, M. Z. Reformat, W. Pedrycz, A. Joulin, Advances in pre-training distributed P. Musilek, Lori: Linguistically oriented rdf inter- word representations, in: Proceedings of the Inface for querying fuzzy temporal data, in: Flexible ternational Conference on Language Resources and Query Answering Systems 2015: Proceedings of Evaluation (LREC 2018), 2018. the 11th International Conference FQAS 2015, Cracow, Poland, October 26-28, 2015, Springer, 2015, pp. 337–352. [13] A. Rula, M. Palmonari, A. Harth, S. Stadtmüller,
A. Maurino, On the diversity and availability of temporal information in linked open data, in: P. CudréMauroux, J. Heflin, E. Sirin, T. Tudorache, J. Euzenat, M. Hauswirth, J. X. Parreira, J. Hendler, G. Schreiber, A. Bernstein, E. Blomqvist (Eds.), The Semantic Web – ISWC 2012, Springer Berlin Heidelberg, Berlin,
Heidelberg, 2012, pp. 492–507. [14] R. C. Anderson, Z. Shifrin, The meaning of words in context, Theoretical issues in reading comprehension (2017) 331–348. [15] N. S. Dash, Context and contextual word meaning.,
SKASE Journal of Theoretical Linguistics (2008). [16] S. Arora, Y. Li, Y. Liang, T. Ma, A. Risteski, Linear algebraic structure of word senses, with applications to polysemy, Transactions of the Association for
Computational Linguistics 6 (2018) 483–495. [17] T. Mikolov, K. Chen, G. Corrado, J. Dean, Eficient