This short paper describes ongoing work on the design of an event ontology that supports state-ofthe-art event extraction in the archives of the Dutch East India Company (VOC). The ontology models Dynamic Events (actions or processes) and Static Events (states). By modelling the transition of a given to a new state as a logical implication that can be inferred automatically from the occurrence of a Dynamic Event, the ontology supports implied information extraction. It also considers implied sub-event detection and models event arguments as coreferential between event classes where possible. By doing so, it enables the extraction of much more information than is only explicitly stated in the archival texts with minimal annotation e昀ort. We de昀椀ne this complete event extraction task that adopts both Natural Language Processing techniques as well as reasoning components as Event Reconstruction. The Event Reconstruction module will be embedded in a search interface that facilitates historical research in the VOC archives.
and detecting and classifying event arguments (who, what, where, etc.). The second step involves the classi昀椀cation of latent events denoted by verbs, nominal expressions or nouns (e.g., friendship evokes a relationship event): a type of implicit information extraction. Finally, the last step of ER deals with a deeper level of implied information extraction, where we extract implied events that cannot be anchored directly to a mention in the text, but are the logical consequence of a mentioned event. In our case this entails logical reasoning using an ontology to extract implied states. Obvious consequences or assumptions are o昀琀en not reported explicitly but need to be inferred to reconstruct a complete story (a mention of one geopolitical entity losing a battle from another implies they were at war). Other information that is o昀琀en implied in text that we also aim to make explicit is how multiple smaller events s(ubevents) sometimes implicitly describe a relevant supra-event (e.g., when the same shipsleave a harbour and arrive at a di昀erent harbour at the same time these subevents describe a voyage event).
Figure 1 illustrates these three levels of information extraction applied on a sample sentence of our corpus.1 It shows how the custom ontology enables the modeling of ship voyages and their locations over large time frames. The aim of the ontology presented in this paper is thus to allow inferencing of states orStatic Events (such as being at a location or being in con昀氀ict ) from extracted Dynamic Events stated explicitly or implicitly in the text (such asleaving a location or attacking). This type of common sense reasoning can be classi昀椀ed as causal reasoning: inferring logical implications of events.
We provide a taxonomy of event classes de昀椀ned by specialist historians, covering a wide range of researchable topics in the VOC archives. It serves as the basis for a more elaborate event ontology that enables the detection of both explicit and implicit information needed for event reconstruction (currently under development)2. To our knowledge, this paper describes the 昀椀rst event ontology designed for a source dating as far back as the Early Modern period. We describe a project that applies state-of-the-art automatic inferencing for the humanities, presenting the 昀椀rst ontology made for the historical 昀椀eld that supports reasoning with implicit information from unstructured text for comprehensive event reconstruction.
A vast amount of research has been conducted on knowledge representation for NLP, resulting in many ontologies and resources. However, none were made speci昀椀cally for our domain, and little leverage commonsense reasoning over event causality. In this section, we discuss the ontologies, lexical resources and theories most relevant to the current project. When assessing re-usability, we consider what information is needed to perform event reconstruction on (pre)colonial sources.
Most historical event detection research has been conducted on datasets from the 20th
century [
ER deals with storyline plausibility. For a system to be able to understand stories (chains
of events) it should have knowledge of causal reasoning 1[
Since the ontology supports an ongoing project, we aim to create a 昀氀exible and
comprehensive ontology to which event classes can be added in an iterative manner. Comprehensibility
is also important to guard annotation quality and modeling possibilities. General-purpose
ontologies like SUMO [13] and FrameNet [
The Rich Event Ontology aims to integrate lexical and ontological resources1[] and also
models causal relations between events [
Modeling causal relationships between events ontologically has been investigated
extensively by Segers [
A key aspect of event reconstruction is modelling the participants of an event, o昀琀en referred to as semantic roles or event arguments. Li et al. (2022) show how event prediction can be informed by implied information extraction9[], taking semantic roles into account. They utilize graph schema induction as a means of predicting new nodes that represent future events. They employ a copy mechanism to generate coreferential arguments (for example, theDetainee argument is the Attacker of a previous Attack event).
We adopt this idea and rely on PropBank 1[
The project in which our ER module is embedded focuses on a subset of the VOC archives, namely the Overgekomen Brieven en Papieren (‘Received Letters and Papers’, OBP) and within that the Generale Missiven (‘General Missives’). These are documents written by the governor general and council of Batavia summarizing the status of, for example, trade and con昀氀ict in Asia. They are written in Early Modern Dutch, which di昀ers from contemporary Dutch with respect to lexical variation, spelling, style, and grammar. Most noteworthy is the frequent occurrence of very long-range dependencies, with pronouns sometimes referring to an entity several sentences or even pages back. Imperfect automatic transcriptions of handwritten material result in even more variation. Given the nature of the historical data and the lack of training data, an ontology can have a strong added value to guide the interpretation of any models that need to process these texts and reconstruct events. See8.2 for some corpus characteristics.
The design of the ontology is based on six requirements following from the interdisciplinary collaboration between NLP scientists and historians and the needs for an ER system.
R1 Domain-speci昀椀c event classes
The ontology should contain event classes that historians have indicated to be relevant in this speci昀椀c corpus, and the classes should be de昀椀ned by historians.
R2 Scalar and binary change
The ontology must accommodate for scalar change as well as binary change. This enables, for example, the modeling of price changes or the conditions of ships over time through ER.
R3 Implied event detection and reasoning
The ontology must accommodate automatic reasoning over events in a way that enables it to extract implied states and events.
R4 Flexibility
The ontology must have a (taxonomic) structure that allows for the addition of events in the future through an iterative process.
R5 Comprehensibility
The ontology must remain as comprehensible and transparent as possible in order to guard both annotation quality and modeling possibilities, taking into account the complexity and noisiness of the source data. The ontology should serve to reduce annotation e昀ort where possible.
A昀琀er considering several linguistic frameworks and ontologies, we opted to take building blocks from di昀erent resources to create an ontology that is as robust as possible for our dataset and research purposes. The CEO served as a basis for the de昀椀nition of event classes because it focuses less on how events are semantically manifested in text (like FrameNet), and more on the practical notion that dynamic events o昀琀en imply a static event. However, many event classes had to be adapted, deleted or added, since the CEO was built for processing modern new4s. We copied the notion that tokens can be annotated with semantic roles cross-sententially from FrameNet (as opposed to PropBank where the whole syntactic structure of one sentence should be represented in the annotation). We took the notion that event arguments do not have to be re-de昀椀ned for each event class speci昀椀cally from PropBank. We de昀椀ned our own limited set of argument types in order to guard annotation quality and modeling purposes.
The present ontology contains events that can be categorised under three themes deemed relevant by historians: ship movement, trade, and (geo)political/social relations. The event classes are de昀椀ned so that they mainly represent observable events, steering clear of concepts that have an inherently subjective character. For example, whether an action is legal or illegal depends on the context, which should be studied 昀椀rstly by a historian. Which events can be classi昀椀ed as observable and which as subjective is an open and ongoing interdisciplinary discussion. Since there also exist causal relations that are debatable or subjective2[4], we only aim to model logical implications of events (which means only modelling necessary and not possible consequences of dynamic events).
For an overview of the current taxonomy of dynamic events, see Figure 2. Most dynamic events automatically imply the transition from a given to a new static event. The static events are modeled as event classes with their own arguments. This allows us to annotate and extract them at the same complexity level as dynamic events. For a sample of the graph indicating how event classes relate, see Figure 3. Our Wiki describes all event classes5.
Previous research [
FrameNet and the CEO de昀椀ne participants speci昀椀c to each event class. In order to be able to model actors of implied changes of state, semantic roles have to be recyclable from one event class to another (if necessary with minimal conversion rules). That is to say, we want to infer that the Agent in an Attacking event is a Patient in a BeingInCon昀氀ict state. FrameNet shi昀琀s the semantics from the event class to the role, which limits the generalization across roles. The CEO o昀ers a solution to this with the incorporation of assertion rules. The goal for our ontology is to have even more intuitive and general roles than in the CEO because the textual data we are working with is extremely noisy and information about one event is stretched out over large portions of text (see Section3). By adopting more intuitive and general roles, 4The 昀椀nal ontology will link classes to the CEO, SUMO, FrameNet and other relevant ontologies with SKOS relations 5https://github.com/globalise-huygens/nlp-event-detection/wiki 6A ship being repaired impliesHavignInternalState+ and a deteriorated city impliesHavingInternalStatewe simplify the annotation process as much as possible, which will lead to more consistent annotation. Also, we suspect that the de昀椀nition of more general semantic roles might be more informative for a machine learning system that needs to generalize on limited training data.
Table 1 gives an overview of our arguments. Participants can be animate and non-animate. The AgentPatient role is speci昀椀cally de昀椀ned for events in which a causer is simultaneously undergoing the event SocialInteraction, BeginningARelationship, EndingAContractualAgreement. For now, Time can be anything from a date to an hour to a day in the week. Source and Target are de昀椀ned for Translocation events that may indicate both the starting and ending point of a movement. Roles are de昀椀ned per event in our Wiki. 7. In our ontology, event participants are represented as properties: i.e. event classes have properties likehasAgent, hasPatient.
The roles can sometimes be recycled directly from a dynamic event class to their supra-class as well as the static class they change, and sometimes they need conversion rules. See Table 2 for an example of how roles are translated between classes. It shows how the Bene昀椀ciary of a Selling event (the person to whom is sold) is an Agent in the static eventHavingInPossession, and it can be deducted that this Agent has the Patient of theSelling event (the thing sold) in possession from the moment the event happened onward. 7https://github.com/globalise-huygens/nlp-event-detection/wiki
An annotation pilot has been conducted and the ontology and event guidelines have been updated accordingly. We plan to enhance the ontology in this iterative manner over a few annotation rounds. When the ontology and guidelines are stabilized, we will run consecutive annotations through the pipeline presented in Figure 4. The resulting Knowledge Graph (KG) is to be integrated in the complete KG supporting the online infrastructure for historical research.
When representing information extracted from text it is important to be transparent about provenance [4]. For each event triple represented in the KG, we will show whether this triple was retrieved through human annotation or through the NLP pipeline. The GRaSP model5[] allows for provenance tracing so that every triple is part of a claim by a source.
This paper has o昀ered an overview of considerations that arise when designing an event ontology for a speci昀椀c domain and a speci昀椀c purpose within an interdisciplinary e昀ort, relying, where possible, on existing resources. We make clear that event classes should be de昀椀ned AP AP A P A P A P A B P A B P A B
P
Ontological Reasoning
Event enrichment module: inferring (Static) Events Provenance tracking P1 P2 P1 P2 P1 P2 P1 P2 A1 A2 P A1 A2 P A2 A1 P
according to the characteristics of the dataset and needs of the 昀椀nal product: in our case de昀椀ning events speci昀椀c enough to the corpus but general enough to leave space for historical and contextual interpretation. We distinguish between event extraction and event reconstruction, de昀椀ning the latter as a combination of several NLP-tasks and reasoning. We propose an ontological method of extracting implicit information and incorporating it in an event reconstruction pipeline.
By focusing on this case-study of event reconstruction in Early Modern Dutch documents we expect to come to new insights about how to push the state-of-the-art in event extraction (using ontologies) forward. Our data and ontology can serve as a robust use-case for investigating to what extend LMs need external common sense knowledge integration to successfully perform certain NLP tasks. Finally, we aim to facilitate a new way of doing historical research with the so昀琀ware supported by the presented ontology.
This research is part of the GLOBALISE project, funded by the Dutch Research Council (NWO) under project number 175.2019.003. We thank Kay Pepping, dr. Lodewijk Petram, dr. Pia Sommerauer and dr. Micky Cornelissen for their help revising the paper, as well dr. Manjusha Kuruppath, Maartje Hids, Henrike Vellinga and Brecht Nijman for their involvement in the annotation pilot. [21] R. Sprugnoli and S. Tonelli. “One, no one and one hundred thousand events: De昀椀ning and processing events in an inter-disciplinary perspective”. In:Natural language engineering 23.4 (2017), pp. 485–506. doi: https://doi.org/10.1017/S1351324916000292.
• Original source ‘Tsedert onsen Jongsten (hiernevens in copie gaande) met de schepen Frederick Hendrik, Hollandia, ’t Wapen van Del昀, ’s landts Hollandia, ende de Galias den 12en November passato in compagnie van Bataviagescheyden ende den 15en ditto door de Strate Sunda geraeckt, sijn hier, Godtlo昀, den 17en ende den 21en van Teyouhan ende Jappan successive wel aengecomen ’t jacht Erasmus ende ’t schip de Vreede, t’samen geladen 810 picol rouwe Chineesche syde ende 332 picol coper8.’ • Literal translation
Since our youngest (herewith in copy) with the ships Frederick Hendrik Hollandia, ’t Wapen van Del昀, ’s landts Hollandia, and de Galias on the 12th of November passato in company from Bataviaseperated and the 15th of the same have made its way through the Sunda Strait, did here, God bless, on the 17th and the 21st from Teyouhan and Jappan successively arrive ’t yacht Erasmus and the ship de Vreede, together loaded 810 picul raw Chinese silk and 332 picul copper.
8‘General Missive of 6-1-1628’, National Archive, The Hague, The Netherlands, 1.04.02 (Archive of the VOC), inventory no. 1092, folio 1, r.
from Council of India (Batavia) to Gentlemen Seventeen
(Dutch Republic).
small summaries in margins
Avg doc length 923 documents; 191.725 handwritten pages c. 250,000 documents; c. 7 million handwritten pages c. 28 pages (rough estimate)