This demonstration paper introduces a tool to analyse historical digital libraries with the bene t of publicly available knowledge bases, such as Wikidata and Wikipedia. In this paper, we focus on realworld-events of the past, such as festivals or assassinations. We introduce our method which merges knowledge from Wikidata and Wikipedia article summaries to gather entities involved in events, dates, types and labels. We hereby present the Web tool we designed and the implemented method to characterise events. It integrates easily into any event oriented pipeline: in our demonstration, we use it to nd event related documents in the NewsEye corpus 1.
We propose a software library 2 able to extract event characteristics from at
least two knowledge bases (KB), Wikidata and Wikipedia. Used together, they
contain su cient data to characterise real world historical events. In this paper,
we refer to events as happenings in the real world which have spatio-temporal
anchors and additional entities involved in it. We focus on participating entities
which are dates, places and participants [
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). 1 The NewsEye corpus is available for anyone at https://platform.newseye.eu 2 The package is a Python 3 library called wikivents on Pypi.org and available on the Software Heritage repository at https://archive.softwareheritage.org/swh: 1:dir:ef325a054ba6f7eb1121807da7b1c92b9ecde8f8 With such a tool, we propose to analyse a large corpora of historical documents. The method we propose in this paper will help ltering documents on whether they report a speci c historical event. To the best of our knowledge, such a tool does not exist. We propose a method to describe events from knowledge bases, and the interface to use it and nd event related documents.
In this paper, we will take the example of the Assassination of Rasputin event, a murder that occurred in December 1916 in the Russian Empire. 2
Our methodology uses an ontology (i.e. Wikidata) as a primary source of
information to obtain fundamental event characteristics: type and dates or times.
It also processes semi-structured databases (i.e. Wikipedia and others) to
identify involved entities [
This software supplies a solution when it is either too long or complicated
to aggregate event characteristics manually. It enables to add more data sources
(e.g. EventKG [
Ontologies: the Extraction of Elementary Event Information
Ontologies such has EventKG [
Our software takes Wikidata entities identi ers as input. Q2882749 is the input value for our example, the assassination of Rasputin. The entity is checked to con rm it is an instance of a WET. If so, the dates (in the Gregorian calendar), the event locations and labels are saved. These generic properties discriminate two similar events: it is unlikely that two distinct events have the same labels, types, and occurred at the same place at the same time. Other properties (the target for a political assassination for instance) do depend on the WET and are often missing. To overcome this limitation, we propose to analyse Wikipedia lead sections as well, looking for relevant named entities.
In our example, we get the entity labels as strings for every language, the date and the linked entities, identi ed by their URIs. 2.2
The event representation is supplemented with entities extracted from
semistructured data-sources. In this work, we analyse the lead sections of Wikipedia
articles. We focus on those written in ve well represented languages that are
English, French, German, Spanish and Italian. Languages are chosen arbitrarily
from the top-10 list of biggest Wikipedia versions, excluding ones written by bots.
For each lead section, involved entities [
To show the relevance of entities in relation to the event, the number of occurrences found in lead sections is counted. Entities found in multiple lead sections are important in the event description, synthesise historical knowledge and give an unbiased information about their implication in the event. In our example, there only exists Wikipedia articles written in Spanish and French. From them, we extract, for instance, these triples: (PER, Q312997 [Felix Yusupov, perpetrator], 3), (PER, Q43989 [Grigori Rasputin, target], 2), (GPE, Q34266 [Russian Empire], 1). The weights, respectively 3, 2 and 1 show that knowing who murdered the victim is more pertinent than where it happened. Beside of those entities, the system found 7 people involved, 5 organisations and 4 geo-political entities. There are respectively only 3, 0 and 3 where analysing Wikidata only. 2.3
The event representation then consists of an association of absolute properties such as dates and links to knowledge bases: it is language independent. In most cases, ontologies provide multiple names in di erent languages (i.e. with di erent writings) for each entity. In our example, in French the entity Q312997 is written Felix Youssoupo or Felix Youssoupov.
A nal step transforms an abstract event representation, based on KB identi ers, into a language-dependent description. It extracts all the alternative spellings for every entity involved in the event. The software makes it possible to get the event description in Italian even if, in this example, only French and Spanish Wikipedias were analysed. 3
It may happen that for some entities, there does not exist any language-speci c spelling in any ontology (e.g. the entity Q3129997, Felix Yussopov has no spelling given in some languages). This case is rare but may be encountered when processing events where local entities are involved, less known to speakers of other languages. We only took into account a list of arbitrarily chosen languages, without capitalising on the languages spoken where the event happened, excluding de facto Asian and African languages.
The process has drawbacks: SPARQL queries are used to retrieve event data and the procedure can be slow due to the multiple API calls. Our caching implementation accelerates the process by a factor of almost 20 (from one minute to ve seconds with the running example). We previously mentioned the possibility to extend the tool capabilities. This way, it is easy to locally process any triple store, such as a Wikidata dump, to overcome this limitation.
Evaluation of such a tool is hard, due to the lack of annotated data. We
propose, to evaluate this tool and the e ciency of the event representation to use
datasets such as those from the Topic Detection and Tracking (TDT) program [
In this paper, we release a set of tools, including a Web front-end to the library, able to gather into one speci c language, all the event related knowledge that exists. We provide an easy way to analyse events described on the Internet. The library can be easily extended and integrating more data sources (i.e. ontologies) is made easy. Extensive technical information may be found on the Pypi project page, as well as real world examples and outputs.
It is already in use as the entry point of a pipeline in a multilingual historical event based search engine which indexes millions of historical documents. For this demonstration, we used a private access to the NewsEye corpus to build queries with information collected from Wikidata and Wikipedia. Similarly to a search engine query, we weighted the entities according to their importance related to the event. The query is a conjunction of all weighted terms found by the wikivents library. 5
This work has been supported by the European Union's Horizon 2020 research and innovation program under grants 770299 (NewsEye) and 825153 (Embeddia).