Temporal Event Knowledge Base (TEKB)

Updated 4 January 2026

TEKBs are computational frameworks that represent, extract, integrate, and reason about events and their temporal relations using formal models like RDF and property graphs.
They employ diverse methodologies, including probabilistic text extraction, narrative-driven acquisition, and embedding-based inference, to construct robust event sequences.
TEKBs support applications such as temporal relation extraction, timeline generation, and predictive analytics while addressing challenges in scalability and data integration.

A Temporal Event Knowledge Base (TEKB) is a computational resource, data structure, or statistical framework designed to represent, extract, integrate, and reason about events and the explicit temporal (time-based) relations among them. TEKBs serve as foundational infrastructure for time-aware information extraction, temporal reasoning in NLP, event-centric knowledge graph completion, predictive analytics, and timeline generation. Research on TEKBs covers a heterogeneous methodological landscape, including probabilistic priors over event sequences, RDF/OWL-based temporal knowledge graphs, narrative-driven event chains, embedding-based temporal link prediction, and logical reasoning for temporal interval forecasting.

1. Formal Definitions and Core Data Models

TEKBs instantiate a variety of formal models to capture event-centric temporal knowledge. A canonical abstraction represents a TEKB as a multigraph, tuple, or labeled property graph, with the following common components:

Event Vertices: Each node corresponds to an event, typically defined by a verb frame (e.g., “die.01”), event instance with associated attributes, or quadruple (e.g., ⟨subject, predicate, object, time⟩).
Temporal Relations: Directed and typed edges model relations such as “before,” “after,” “includes,” or general interval relations (Allen’s interval algebra). These are often explicit (annotated time spans) or implicit (logical ordering).
Entity Vertices: Real-world entities, participants, or objects involved in events.
Time Representations: Time may be encoded as instants, intervals, or entire timestamp nodes.
Schema/Types: Event and entity types (ontology-backed or data-driven), roles, and relations.

A TEKB may be formally described as:

$\mathrm{TEKB} = (L,\,V,\,\mathcal{E},\,R,\,\tau,\,\lambda,\,\delta)$

where $L$ is the set of languages, $V$ the entity set, $\mathcal{E}$ the event set, $R$ the edge/relation set, $\tau$ a mapping from events to intervals, $\lambda$ and $\delta$ are labeling and description functions across $L$ (Gottschalk et al., 2018, Gottschalk et al., 2018). Alternative formalisms include labeled property graphs (LPGs) with strict node and edge type constraints as in tEKG (Khayatbashi et al., 2024), or temporal event graphs with interval-annotated edges (Xiong et al., 2023).

2. Construction Methodologies

TEKBs are constructed through one or more of the following methodological paradigms:

2.1 Statistical Extraction from Text Corpora

Probabilistic Priors: Large-scale extraction of event pairs and their temporal ordering, as realized in TemProb/TEKB, leverages corpora such as the NYT Annotated Corpus. Events are extracted as verb semantic frames, and pairwise temporal relations are classified using learned perceptron models with input features (POS, connectives, modal verbs, etc.) (Ning et al., 2018).
Global Aggregation: Sticky transitive closure enforces coherence within single documents, with all temporal graphs then globally summed to induce empirical probabilities and pairwise event ordering statistics.

2.2 Temporal Knowledge Graph Integration

Knowledge Graph Merging: Multilingual event-centric resources such as EventKG are produced by integrating event and temporal data from DBpedia, Wikidata, YAGO, Wikipedia event lists, and current events portals through light-weight alignment (OWL:sameAs, time windows, title heuristics) (Gottschalk et al., 2018, Gottschalk et al., 2019).
Temporal Fusion and Provenance: Conflicting time information is resolved through majority voting and trust hierarchies among sources; all triples retain provenance via named graphs.

2.3 Narrative-Driven Acquisition

Narrative Principle: The “double temporality” property of narratives aligns textual order with chronological order, enabling extraction of high-confidence “before/after” event relations across sentences using actantial syntax patterns and bootstrapped weakly-supervised classifiers (Yao et al., 2018).
Scoring and Selection: Event pairs and chains are ranked by “causal potential,” combining PMI and directionality measures, to select the most probable temporal sequences.

2.4 Embedding-Based and Logical Inference

Latent Embedding Spaces: TEKBs constructed for temporal knowledge base completion embed entities, relations, and time in high-dimensional spaces, supporting joint link and interval prediction (Jain et al., 2020, Park et al., 2022).
Logical Rule Extraction: TEILP converts TKGs to explicit event graphs with timestamp nodes, enabling differentiable random-walk inference and rule-based conditional densities for robust, interpretable time prediction (Xiong et al., 2023).

2.5 Business Process and Event Log Analysis

Object-Centric Event Log Transformation: Recent approaches formalize tEKG as an extension of OCEL 2.0, explicitly modeling snapshot nodes for per-object attribute changes with qualified temporal edges, yielding graph-based TEKBs suitable for temporal join queries and windowed analytics (Khayatbashi et al., 2024).

3. Schema, Representation, and Storage

Specifications for TEKBs are tightly coupled to their target use cases:

Event Representation: Frames (disambiguated verb senses), event instances with argument roles, or (subject, relation, object, time) tuples.
Temporal Relations and Labels: TimeBank-Dense six-way relations (before, after, includes, included, equal, vague), Allen interval relations, or custom role/predicate sets.
Intervals and Time Points: Properties sem:hasBeginTimeStamp/sem:hasEndTimeStamp in RDF (EventKG); explicit timestamp nodes in tEKG/TEKG.
Data Structures: RDF triple stores (SPARQL endpoint queries), JSONL or property graphs, tensor/embedding memory layouts, or matrices for logical walks.

Example: The EventKG schema extends SEM with classes (sem:Event, sem:Actor, sem:Place, eventKG-s:Relation) and properties sem:hasBeginTimeStamp, sem:hasEndTimeStamp, sem:hasPlace, eventKG-s:links and eventKG-s:mentions (Gottschalk et al., 2018, Gottschalk et al., 2019).

4. Applications and System Integration

TEKBs provide generalizable priors and direct supporting features for multiple downstream and integrative tasks:

Temporal Relation Extraction: Priors derived from a TEKB improve F1 in temporal relation extraction, whether included as real-valued features or as global ILP regularizers (Ning et al., 2018).
Timeline and Biography Generation: TEKBs underpin automatic construction of entity-centric or event-centric timelines, with systems such as EventKG+TL leveraging mention/link statistics and relevance scores (Gottschalk et al., 2018, Gottschalk et al., 2019).
Narrative Understanding and Script Induction: Extracted event chains support narrative cloze and story completion tasks, outperforming neural baselines when leveraging pairwise causal potential (Yao et al., 2018).
Temporal Knowledge Base Completion: State-of-the-art link and time prediction in temporal KBC tasks are enabled by embedding-based TEKBs (TIMEPLEX, EvoKG) and logical inference (TEILP), with rigorous time-aware evaluation protocols (Jain et al., 2020, Park et al., 2022, Xiong et al., 2023).
Business Process Analytics: tEKG representations allow for powerful, snapshot-based temporal joins, attribute window analysis, and reasoning over object state histories (Khayatbashi et al., 2024).
Adaptation to New Domains: Modular pipelines (e.g., EventPlus) allow straightforward replacement of the encoder/label taxonomy, facilitating adaptation to biomedical or other domain-specific event schemas (Ma et al., 2021).

5. Statistical Properties, Evaluation, and Benchmarks

TEKBs have been quantitatively characterized at scale and validated across multiple evaluation axes.

Coverage: TemProb captures 51,000 verb frames and 80 million event-pair count entries (Ning et al., 2018); EventKG contains 690,247 events and over 2 million temporal relations (Gottschalk et al., 2018).
Precision: Event extraction precision can exceed 90% in both news-based statistical resources and multi-KG integration (EventKG/Wikidata); narrative TEKBs report >84% annotation precision in bootstrapped narrative identification (Yao et al., 2018).
Temporal Prediction Accuracy: Embedding- and logic-based TEKB approaches (TIMEPLEX, EvoKG, TEILP) yield up to 77–116% relative gains in temporal prediction accuracy and significant lifts in link-prediction MRR compared to static or simple baselines (Jain et al., 2020, Park et al., 2022, Xiong et al., 2023).
Timeline and Biographical Relevance: TEKB-supported timeline generation is preferred by users over generic baselines by margins of 50% or more (Gottschalk et al., 2019).
Evaluation Protocols: Time-aware evaluation metrics, such as mean intersection-over-union (aeIOU), and time-filtered candidate ranking are established as standard for robust model comparison (Jain et al., 2020).

6. Limitations and Future Research Directions

Despite substantial advancements, TEKBs manifest several open challenges:

Event Sense Disambiguation and Synonymy: Narrative-based and statistical TEKBs remain sensitive to event polysemy; mapping to unified sense inventories (FrameNet, PropBank) is an ongoing goal (Yao et al., 2018).
Semantic Type Unification: Merging of event types across ontologies and harmonization of event series versus singleton events remains an open integration problem (Gottschalk et al., 2018).
Spatial and Argument Coverage: Location information is often missing from semi-structured and KG sources (e.g., only 12% of EventKG events carry location); argument role filling and temporal scoping are frequently incomplete.
Globally Consistent Inference: Many pipelines rely on local or document-level predictions; global inference through transitive closure or joint structured models is flagged as potential improvement (Ma et al., 2021).
Scalability and Adaptive Reasoning: Very large, growing TEKBs impose computational and memory demands (embedding updates, RNN state storage); inductive methods, graph subsampling, and meta-learning are being explored (Park et al., 2022).
Expressive Temporal Distributions: Current time-density models are often limited to simple log-normal or Gaussian mixtures; more flexible multimodal or flow-based distributions are needed for arbitrarily distributed event times (Park et al., 2022, Xiong et al., 2023).
Interpretability and Rule Extraction: Recent TEILP demonstrates how explicit logical rule extraction provides interpretable explanations for predictions, an area that merits further research and adoption (Xiong et al., 2023).

7. Resource Availability and Standard Querying

Public releases and interoperability support are critical to broad TEKB adoption:

Datasets and Dumps: EventKG and TemProb/TEKB offer downloadable RDF dumps or count tables (Gottschalk et al., 2018, Ning et al., 2018).
Query Interfaces: SPARQL endpoints enable complex time-based queries (sub-event retrieval, timeline construction by date or popularity) (Gottschalk et al., 2018, Gottschalk et al., 2018).
Schema Examples and API Access: JSONL/Neo4j schemas (EventPlus), timeline APIs (EventKG+TL), and sample Cypher queries (tEKG) facilitate integration into NLP and analytics pipelines (Ma et al., 2021, Khayatbashi et al., 2024).
Code and Benchmarks: Accompanying source code and evaluation splits provided in multiple works ensure reproducibility and comparative benchmarking.

In summary, Temporal Event Knowledge Bases define a flexible, multi-paradigm computational substrate for capturing, representing, and reasoning over event-based temporal information at scale. TEKBs integrate corpus-derived statistical priors, logic-driven reasoning, KG fusion, and process log traces to form the technical foundation of time-aware semantic analytics, predictive timeline construction, and machine reading for historical, narrative, and dynamic contexts (Ning et al., 2018, Gottschalk et al., 2018, Yao et al., 2018, Jain et al., 2020, Park et al., 2022, Xiong et al., 2023, Khayatbashi et al., 2024, Ma et al., 2021, Gottschalk et al., 2018, Gottschalk et al., 2019).