OEKG: The Open Event Knowledge Graph

Abstract

Accessing and understanding contemporary and historical events of global impact such as the US elections and the Olympic Games is a major prerequisite for cross-lingual event analytics that investigate event causes, perception and consequences across country borders. In this paper, we present the Open Event Knowledge Graph (OEKG), a multilingual, event-centric, temporal knowledge graph composed of seven di erent data sets from multiple application domains, including question answering, entity recommendation and named entity recognition. These data sets are all integrated through an easy-to-use and robust pipeline and by linking to the event-centric knowledge graph EventKG. We describe their common schema and demonstrate the use of the OEKG at the example of three use cases: type-speci c image retrieval, hybrid question answering over knowledge graphs and news articles, as well as language-speci c event recommendation. The OEKG and its query endpoint are publicly available.

1 Introduction

Contemporary and historical events such as the US presidential elections, the Olympic Games and major earthquakes change the world. Their media coverage, their varying perception by di erent communities, their historical evolution and potentially global impact make cross-lingual event analytics a signi cant research topic in various elds of studies, including social science, computer science and digital humanities [ 4, 19 ].

When performing cross-lingual event analytics, the requirements towards event knowledge representation are manifold, given the heterogeneity, dynamicity and multilingualism of events [ 9 ]. Until now, there exists a large variety of event-related data sets [ 1, 4, 11, 20 ] that may help understand speci c characteristics of events, but they are barely connected by now. This calls for new models and processes that enable intuitive access to the event-related knowledge spread across the world.

In this paper, we present the OEKG , the Open Event Knowledge Graph, which makes a step towards a holistic representation of event knowledge by the integration of event-related data sets from multiple and diverse application domains such as Question Answering, entity recommendation and Named Entity Recognition. Also, these data sets originate from di erent data collections, including knowledge graphs and news articles. One of these knowledge graphs is EventKGlight, a new version of the event-centric and multilingual knowledge graph EventKG [ 11 ]. The OEKG is built on top of EventKGlight, allowing for easier integration of additional data sets using RDF named graphs. We propose an e cient and robust pipeline facilitating this integration of several data sets in an easy-to-use manner.

Fig. 1 shows four example resources of OEKG and thus demonstrates its versatility resulting from the integration of several data sets: { Events (Fig. 1a): Events are at the core of the OEKG . For example, the re of the Notre-Dame in Paris is covered with its locations, labels in multiple languages, related events such as \The Notre Dame Cathedral holds its rst mass since the April 15 re", and more event characteristics. { Places (Fig. 1b): Most events happen at speci c event locations which are also part of the OEKG . Such places do not only hold labels and coordinates, but also images and further characteristics. { News articles (Fig. 1c): Events are often reported in the media [ 5,13 ]. Therefore, the OEKG provides access to annotated news articles. For example, the news article entitled \Boris Johnson takes charge of Olympic Park's future" is related to the Olympic Games. { Questions and answers (Fig. 1d): Question Answering over knowledge graphs is an important natural language understanding task. The OEKG provides questions about events such as the Apollo 11 space ight, plus their answers (here, Neil Armstrong, Michael Collins and Buzz Aldrin).

Furthermore, the OEKG covers several other event-related aspects, including but not limited to (temporal) event relations, language-speci c relevance scores and specialised class hierarchies. Put together, this makes the OEKG a versatile resource targeting a variety of potential information needs. (a) Example event in the OEKG. (b) Example place in the OEKG. (c) Example news article in the OEKG.

(d) Example question in the OEKG. The OEKG contains more than 400 million triples from seven data sets and is publicly available: We provide the triple dumps for download, a SPARQL endpoint and access to all nodes on the OEKG website9. We also provide permanent access to the OEKG on Zenodo10.

The remainder of this paper is organised as follows: First, we present our integration pipeline (Section 2). Then, we describe the data sets integrated into the OEKG (Section 3) and the OEKG schema (Section 4). In Section 5, we provide two example use cases of the OEKG . Finally, we conclude in Section 6. 2

Creation of the OEKG

The creation of the OEKG requires an integration pipeline where a set of data sets is transformed into a single, integrated knowledge graph that provides links between all the involved resources. EventKGlight { a multilingual, event-centric knowledge graph later described in Section 3 { serves as the base data set of the OEKG that contains nodes representing real-world entities and events.

Our integration pipeline is driven by the goal to make the inclusion of a new data set into the OEKG as simple as possible, which allows a robust and e cient process. Only then, it is possible to integrate a large variety of data sets in an e cient and faultless way. To do so, we follow a strategy de ned by Galkin et al. [ 8 ] where the data from di erent sources is stored under respective named graphs. Starting from EventKGlight, new data sets are added consecutively, each accompanied by a unique named graph. Fig. 2 exempli es this integration process when adding the rst new data set to EventKGlight, under the named graph new graph.

new_graph event_kg

Entity Linking

Graph Creation

Graph Upload new_graph event_kg OEKG

In detail, the integration process follows the following three steps: 1. Entity Linking: We require that each graph added to OEKG is connected to EventKGlight. That means any resource representing a real-world entity or event is represented by an OEKG resource URI. To facilitate this linking, we provide a web API that allows easy access to the OEKG resource URIs given Wikidata or DBpedia URIs. In our example in Fig. 2, some input table cells are successfully linked to EventKGlight. Algorithm 1 Example: Extension of the OEKG with a data set news that has an article about Barack Obama 1: procedure ExtendOEKG(e) 2: graphName "news" . 4: 5: 6: 7: 8: entityId

getId("en";"Barack Obama") G new Graph(graphName) articleId "article1" G.add(oekg-r:articleId, rdf:type, so:Article) G.add(oekg-r:articleId, so:mentions, oekg-r:entityId)

leName storeGraphIntoFile(G) uploadGraph( leName; graphName) . Entity Linking . Graph Creation . Graph Upload 2. Graph Creation: After retrieval of the OEKG resource URIs, a set of triples is created for each data set and serialised as an N-Triples11 le, using the RDFLib Python library12. In our example, a graph consisting of ve nodes is created, two of them being already part of the OEKG. 3. Graph Upload: We provide another API method that allows uploading an N-Triples le together with the identi er of a named graph. The respective triples are then added to the OEKG. In our example, the resulting graph consists of two subgraphs that can be queried in isolation or together. 2.1

Example Consider Algorithm 1 for an example of our integration pipeline. In this example, the new data set to be added to the OEKG under the named graph news contains one news article about Barack Obama. First, the OEKG URI of Barack Obama is retrieved via the provided API method using the English Wikipedia label (line 3). Second, a graph is created consisting of two triples and serialised into an RDF le (lines 4 - 8)13. Third, this le is uploaded via the provided API method (line 9). In this example, one new node is added to the OEKG (oekg-r:articleId) connected to an existing node (oekg-r:entityId). 2.2

Schema Extension If possible, the data sets were transformed into triples using the EventKGlight schema of the base graph. Otherwise, the use of standard vocabularies such as schema.org14 was encouraged. In every other case, schema extensions were uploaded into the OEKG through separate schema les using the same procedure. We will present the resulting OEKG schema in Section 4. 11 https://www.w3.org/TR/n-triples/ 12 https://rdflib.dev/ 13 Relevant pre xes used by the OEKG are later de ned in Table 2. 14 https://schema.org/ The OEKG integrates seven data sets which are described in this section. Table 1 provides an overview of these data sets, including the number of triples in the OEKG within their respective named graph. While some of these data sets are implicitly related to events, others add to the event knowledge from a di erent perspective, which will also prove useful as we will later show at the example of three use cases. provide a reference source to train and evaluate novel models for eventcentric cross-lingual user interaction. It directly builds upon EventKG and language-speci c information on user interactions with events, entities, and their relations derived from the Wikipedia clickstream.

In the OEKG , EventKG+Click can be used for recommending events to users based on actual user interaction traces. Examples of particular relevant events from a language-speci c view include the 2016 Berlin truck attack from the German perspective and the 2009 Russian Premier League from the Russian perspective [ 1 ]. { VQuAnDa [ 12 ]: The Verbalization QUestion ANswering DAtaset is a dataset for Question Answering (QA) over knowledge graphs that includes the verbalization of each answer. Through this verbalisation, VQuAnDa intends to completely hide any semantic technologies and provides a uent experience between the users and the knowledge graph. VQuAnDa consists of 5; 000 questions accompanied by SPARQL queries and DBpedia entity links. QA over Knowledge Graphs is a common task in natural language processing [ 6 ]. Via the integration of question/answer pairs into the OEKG , both the question/answers pairs and the background knowledge are encapsulated into the same resource, enabling seamless training and application of QA systems. { MLM [ 3 ]: The Multiple Languages and Modalities data set is a resource for training and evaluating multitask systems in multiple modalities, for example, cross-modal (text/image) retrieval and location estimation. MLM comprises text in three languages, images and location data, extracted from the Wikidata entries of 236; 000 human settlements.

MLM is added to the OEKG for adding images as an additional modality to the knowledge graph. As locations are typical event characteristics, photos of locations are an immediate bene t to the representation of events. { InfoSpread [ 20 ]: The data set for Information Spreading over the News provides news articles covering three contrasting events (Global Warming, FIFA world cups and earthquakes). Initially, the goal of this data set was to understand information spreading patterns over news articles. InfoSpread contains 7; 773 news articles related to these events in ve languages. News articles are often used as a means to identify events [ 13 ] and oftentimes it is the media itself that makes events known to the public [ 5 ]. Therefore, the inclusion of news articles into the OEKG is an important step towards coverage of event-centric data from di erent viewpoints. { TIME [ 4 ]: The temporal discourse analysis applied to media articles data set is a collection of Brazilian, British and Spanish news articles covering the concept of Olympic legacy and the concept of Euroscepticism.

With the collection of news articles to speci ed events, the OEKG serves as an example for in-depth analysis of single events through knowledge graphs. { UNER [ 2 ]: The Universal Named Entity Recognition framework proposes a 4-level class hierarchy for training and testing Named Entity Recognition tools. For example, UNER contains the class Earthquake, which is a leaf node of the following branch of superclasses: Natural, NaturalPhenomenon, Event and Name.

In the OEKG, UNER adds to the already given class hierarchy from the DBpedia ontology. Given how challenging it is to recognise named events in texts [ 16 ], we envision that the inclusion of UNER classes into the OEKG can help training and evaluating NER systems in the speci c context of event-centric data.

Following the integration pipeline described in Section 2, the described data sets were added to the OEKG. For additional information or increased interlinkage with EventKGlight, some data sets were extended before: { Via the Wiki er15 and spaCy16, entities and events mentioned in news articles (TIME and InfoSpread ) were identi ed. This is to establish a connection between the news articles and EventKGlight: Given this connection, one may query for news articles about speci c events or entities. { Sentiment analysis, i.e., the computational study of people's opinions, sentiments, emotions, moods, and attitudes [ 14 ], contributes towards the understanding of natural-language texts and can, in particular, facilitate an analysis of news articles across languages [ 15 ]. In the OEKG, we enrich news articles by employing the sentiment detection system SentiStrength [ 21 ] on their headlines. That way, the OEKG enables queries for particularly positive or negative news articles, potentially initiating further event-centric analyses of the news articles in the context of speci c events. { To further increase the linkage between di erent sources, the UNER classes were aligned to the DBpedia ontology using the skos vocabulary17 when possible. 4

Schema

{ EventKGlight: The EventKG schema is based on the Simple Event Model (sem)19 and its three main classes sem:Event, sem:Actor and sem: Place, that are connected via sem:hasPlace and (temporal) relations modeled by oekg-s:Relation (omitted from Fig. 3 for brevity). EventKG further distinguishes between di erent types of events (oekg-s:TextEvent, oekg-s:EventSeries and oekg-s:EventSeriesEdition). In 15 http://wikifier.org/ 16 https://spacy.io/ 17 https://www.w3.org/TR/swbp-skos-core-spec/ 18 For a full list of pre xes used in the OEKG, see oekg.l3s.uni-hannover.de/ sparql. 19 https://semanticweb.cs.vu.nl/2009/11/sem/ comparison to the EventKG schema, EventKGlight omits link count relations and adds the skos:prefLabel to entities for a more e cient access to their labels. { EventKG+Click : To model language-speci c, weighted relations for the representation of event-centric cross-lingual user interaction, we have introduced two new classes: oekg-s:LanguageSpecificRelation that assigns one or more instances of oekg-s:LanguageSpecificRelationScore to a source entity and a target entity. Such instances hold the score between the source and target entity in a speci c language. { VQuAnDa : A question, its suggested answer and their verbalisation are represented using schema.org's classes so:Question and so:Answer. Entities that appear in the question text are linked to EventKGlight instances via so:mainEntity, entities in the answer via so:mentions. { MLM : Images are assigned to places via so:image, descriptions via so: description. { InfoSpread and TIME : News articles are represented via so:Article and the respective properties denoting the headline (so:headline), for instance. News articles are connected to EventKGlight instances via so: mentions, which denote the appearance of an OEKG entity or event in the text. For the representation of news articles' sentiment, we follow the schema of the TweetsKB [ 7 ], using the onyx vocabulary and its classes onyx:EmotionSet, onyx:Emotion and onyx:EmotionCategory to assign a set of emotions of di erent strengths to a news article. { UNER: Entities are assigned UNER classes using rdf:type. Furthermore, the UNER class hierarchy and its connection to the DBpedia ontology are established using the owl and the skos vocabulary.

xs:language xs:date so:inLanguage so:date

Published xs:string so:publisher

TIME & InfoSpread onyx:Emotion

Category onyx:hasEmo tionCategory VQuAnDa

UNER xs:string so:text so:Question so:suggested

Answer rdfs:sub

ClassOf owl:equivalent

Class so:mainEntity so:Answer so:mentions so:text prefLabeslkos: xs:string xs:string skos:narrower owl:Class rdf:type sem:Core sem:Actor dbo:previousEvent / dbo:nextEvent / sem:hasSubEvent EventKG oekg-s: TextEvent xs:string xs:anyURI so:mentions rrddff::soubbjejecctt / so:image

xs:anyURI sem:Event sem:hasPlace

sem:Place oekg-s: EventSeries

Edition oekg-s: EventSeries so:head line so:url onyx:hasEmo tionIntensity xs:double so:Article onyx:has EmotionSet onyx:has Emotion onyx: Emotion onyx: Emotion

Set oekg-s:Language SpecificRelation so: contained

InPlace so:

description xs:string MLM oueakgge-Ss:claonreg LaRnegluoaateigkogenS-Sspc:oecreific

Loaenkggu-asg:secore osVeckaoglrue-es: xs:language

xs:double EventKG+

Click In this section, we demonstrate the OEKG and its ability to enable integrated access over multiple datasets via three example use cases. 5.1 Image Retrieval: EventKGlight, MLM & UNER Event classi cation in images is an important task for various applications in the elds of computer vision, including geolocation estimation and place classi cation [ 17 ]. Such tasks typically rely on the existence of a well-de ned class hierarchy and the availability of images. The OEKG facilitates queries both for the UNER type hierarchy speci cally designed for Named Entity Recognition, and for images of locations, using the MLM data. In combination, event locations in EventKGlight, MLM 's image links, and the UNER type hierarchy enable retrieval of images relevant for speci c event types.

We demonstrate the OEKG's potential for image retrieval by an example query for images from earthquake regions shown in Listing 1.1: It queries for entities typed as earthquakes using the uner:Earthquake class, their locations (EventKGlight) and the images assigned to such locations (MLM ). Table 3 presents selected results of this query, including a photo of the port of Messina and more.

SELECT DISTINCT ?Location ?Image WHERE { ?earthquake rdf:type uner:Earthquake ;

sem:hasPlace ?Location .

?Location so:image ?Image . }

Listing 1.1. SPARQL query: Images of locations where earthquakes happened.

Question Answering over News Articles: EventKGlight,

VQuAnDa , InfoSpread & TIME Question Answering (QA) is the task of supplying precise answers to questions, posed by users in natural language, and is typically divided into QA over free text and QA over knowledge graphs [ 6 ]. Through the integration of EventKGlight, VQuAnDa, TIME and InfoSpread into the OEKG, the OEKG facilitates a combination of these two tasks, i.e., hybrid approaches: We can query for news articles which speci cally mention the entities part of the question/answer pair. 20 These photos are taken from Wikimedia Commons. They are licensed under the following licenses. Ferrara: Creative Commons Attribution 2.5 Italy license. Messina: Creative Commons Attribution-Share Alike 3.0 Unported, 2.5 Generic, 2.0 Generic and 1.0 Generic. Guaranda: Creative Commons Attribution 2.0 Generic license.

This way, two sources for answering the question can be provided: the OEKG itself, as well as the news article potentially holding the answer to the initially posed question.

For example, the query in Listing 1.2 asks for a question in VQuAnDa (?question) that is about an event (?questionEntity rdf:type sem:Event). The query then searches for news articles (?article) mentioning both that event and one of the suggested answer entities. It returns the question \Whose wife is a presenter at WWE? (en)" and its verbalised answer \The people whose partners are presenters at WWE are John Cena, Dwayne Johnson." together with the Spanish news articles entitled \>Que luchador tiene el mayor porcentaje de victorias en la historia de WWE?" (Which wrestler has the highest percentage of victories in in the history of WWE? ). The question entity \WCE (en)" is mentioned in the news article, as well as both answers: John Cena and Dwayne Johnson.

SELECT DISTINCT ?questionText ?answerText ?headline

?questionEntity ?answerEntity WHERE { ?question so:suggestedAnswer ?answer; so:mainEntity ?questionEntity ; so:text ?questionText . ?questionEntity rdf:type sem:Event . ?answer so:mentions ?answerEntity ;

so:text ?answerText . } ?article rdf:type so:Article ; so:mentions ?questionEntity, ?answerEntity ; so:headline ?headline .

Listing 1.2. SPARQL query: News articles that mention entities of a question/answer pair. 5.3 Event Recommendation: EventKGlight & EventKG+Click As de ned by Ni et. al, entity recommendation is the problem of suggesting a contextually-relevant list of entities in a particular context [ 18 ]. This task is particularly relevant in Web search. With the OEKG, we can speci cally create language-speci c recommendations for events and further enrich them with relevant event characteristics.

The query in Listing 1.3 asks for events relevant to the First World War, from the Russian point of view. We lter for the most relevant related events (FILTER(?value >= 0.8)) and retrieve EventKGlight's event characteristics SELECT ?Label ?StartDate WHERE { ?event owl:sameAs dbr:World_War_I. ?r oekg-s:source ?event ; oekg-s:target ?target ; oekg-s:hasLanguageSpecificRelationScore [ oekg-s:scoreValue ?value ; oekg-s:scoreLanguage ’ru’ˆˆxsd:language ] . ?target skos:prefLabel ?Label ;

sem:hasBeginTimeStamp ?StartDate .

FILTER(?value >= 0.8) . } ORDER BY ?StartDate Listing 1.3. SPARQL query: Events related to the First World War from a Russian point of view. to order the resulting list of events chronologically. Table 4 lists the results of this query, that clearly show a Russian focus. This result could be used for creating a language-speci c event timeline similar to the link-based EventKG+TL system [ 10 ], but now inferred from actual user interaction traces in EventKG+Click . In this paper, we have introduced the OEKG { the Open Event Knowledge Graph21. The OEKG comprises event-related knowledge from seven data sets of various application domains. We have presented an easy-to-use, e cient and robust pipeline that facilitated a seamless integration of seven data sets into the OEKG. At the examples of image retrieval, question answering over text and event recommendation, we have exempli ed three use cases of the OEKG. Acknowledgements The project leading to this publication has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No. 812997 (Cleopatra). 21 http://oekg.l3s.uni-hannover.de/

References

1. Abdollahi , S. , Gottschalk , S. , Demidova , E.: EventKG+Click: A Dataset of Language-speci c Event-centric User Interaction Traces . In: CLEOPATRA Workshop @ ISWC ( 2020 )

2. Alves , D. , Kuculo , T. , Amaral , G. , Thakkar , G. , Tadic , M. : Uner: Universal namedentity recognitionframework . In: CLEOPATRA Workshop @ ISWC ( 2020 )

3. Armitage , J. , Kacupaj , E. , Tahmasebzadeh , G. , Maleshkova , M. , Ewerth , R. , Lehmann , J.: MLM: A Benchmark Dataset for Multitask Learning with Multiple Languages and Modalities . In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management . pp. 2967 { 2974 ( 2020 )

4. Cheema , G.S. , Major , D. , Mello , C. , Sittar , A. : TIME: Temporal Discourse Analysis applied to Media Articles . http://cleopatra-project.eu/index.php/ 2020 /06/01/ time -temporal-discourse-analysis-applied-to-media-articles/ ( 2020 )

5. Dayan , D. , Katz , E.: Media Events. harvard university press ( 1994 )

6. Dimitrakis , E. , Sgontzos , K. , Tzitzikas , Y. : A Survey on Question Answering Systems over Linked Data and Documents . Journal of Intelligent Information Systems pp. 1 { 27 ( 2019 )

7. Fafalios , P. , Iosi dis , V., Ntoutsi , E. , Dietze , S.: TweetsKB: A Public and Largescale RDF Corpus of Annotated Tweets . In: European Semantic Web Conference . pp. 177 { 190 . Springer ( 2018 )

8. Galkin , M. , Auer , S. , Kim , H. , Scerri , S. : Integration Strategies for Enterprise Knowledge Graphs . In: 2016 IEEE Tenth International Conference on Semantic Computing (ICSC) . pp. 242 { 245 ( 2016 )

9. Gottschalk , S. , Bernacchi , V. , Rogers , R. , Demidova , E.: Towards Better Understanding Researcher Strategies in Cross-lingual Event Analytics . In: International Conference on Theory and Practice of Digital Libraries . pp. 139 { 151 . Springer ( 2018 )

10. Gottschalk , S. , Demidova , E.: EventKG+TL: Creating Cross-lingual Timelines from an Event-centric Knowledge Graph . In: European Semantic Web Conference . pp. 164 { 169 . Springer ( 2018 )

11. Gottschalk , S. , Demidova , E.: EventKG{the Hub of Event Knowledge on the Web{ and Biographical Timeline Generation . Semantic Web 10 ( 6 ), 1039 { 1070 ( 2019 )

12. Kacupaj , E. , Zafar , H. , Lehmann , J. , Maleshkova , M.: VQuAnDa: Verbalization QUestion ANswering DAtaset . In: European Semantic Web Conference. pp. 531 { 547 . Springer ( 2020 )

13. Leban , G. , Fortuna , B. , Brank , J. , Grobelnik , M. : Event Registry: Learning about World Events from News . In: Proceedings of the 23rd International Conference on World Wide Web . pp. 107 { 110 ( 2014 )

14. Liu , B.: Sentiment Analysis: Mining Opinions , Sentiments, and Emotions. Cambridge University Press ( 2020 )

15. Lo , S.L. , Cambria , E. , Chiong , R. , Cornforth , D. : Multilingual Sentiment Analysis: from Formal to Informal and Scarce Resource Languages . Arti cial Intelligence Review 48 ( 4 ), 499 { 527 ( 2017 )

16. Marujo , L. , Ling , W. , Gershman , A. , Carbonell, J.G., Neto , J.P. , Matos , D. : Recognition of Named-Event Passages in News Articles . In: Proceedings of COLING 2012: Demonstration Papers . pp. 329 { 336 ( 2012 )

17. Muller-Budack , E. , Springstein , M. , Hakimov , S. , Mrutzek , K. , Ewerth , R.: Ontology-driven Event Type Classi cation in Images . In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . pp. 2928 { 2938 ( 2020 )

18. Ni , C.C. , Sum Liu , K. , Torzec , N.: Layered Graph Embedding for Entity Recommendation using Wikipedia in the Yahoo! Knowledge Graph . In: Companion Proceedings of the Web Conference 2020 . pp. 811 { 818 ( 2020 )

19. Rogers , R.: Digital Methods. MIT press ( 2013 )

20. Sittar , A. , Mladenic , D. , Erjavec , T. : A Dataset for Information Spreading over the News . In: Conference on Data Mining and Data Warehouses (SiKDD) ( 2020 )

21. Thelwall , M. , Buckley , K. , Paltoglou , G. , Cai , D. , Kappas , A. : Sentiment Strength Detection in Short Informal Text . Journal of the American society for information science and technology 61(12) , 2544 { 2558 ( 2010 )