Introduction
Knowledge Bases (KBs) such as Freebase and Wikidata are invaluable for structured data retrieval, yet they suffer from significant sparsity regarding recurrent events. Events like the "Super Bowl," "IJCAI Conference," or "LA Marathon" occur periodically, but KBs often fail to capture their specific instances. Researchers from IIT Delhi, IBM, and Stanford found that while Freebase lists approximately 15,770 recurring events, it contains only about 16,347 specific instances—a staggering 85% gap compared to the expected 146,000 instances over a decade.
The Challenge of Near-Periodicity
The primary difficulty in completing these KBs lies in the "near-periodic" nature of real-world events. While many follow a strict rule (e.g., "the first Sunday in February"), they are subject to "noise"—shifts caused by holidays, venue availability, or leap years. Traditional Information Extraction (IE) tools often lack the precision to handle these nuances, failing to provide the day-level accuracy required for high-quality KB maintenance.
The TRINE Framework
To address this, the authors developed TRINE (TempoRal Inference for Near-periodic Events). TRINE is a joint inference engine designed to extract and predict dates for recurrent events by synthesizing information from multiple sources: schedule descriptions in Wikipedia, historical dates in KBs, and event mentions in unstructured news corpora.
As shown in the architecture above, TRINE utilizes an Instance Extractor to find specific dates in news text and a Schedule Extractor to parse natural language descriptions of event timings. These inputs are then processed by the Temporal Inference Engine to produce a refined schedule database and predict future instances.
Joint Inference via Probabilistic Graphical Models
At its core, TRINE treats temporal inference as a joint task using an undirected probabilistic graphical model. This approach allows the system to reconcile conflicting or vague information. For example, if a news article mentions an event "in February," TRINE can use historical data to determine the exact day-level rule.
The model balances three primary "potentials":
- Schedule Potential: Evaluates how well a candidate schedule (e.g., "1st Sunday of Feb") matches textual descriptions.
- Occurrence Potential: Measures the confidence of dates extracted from news text or existing KB facts.
- Pairwise Potential: Assesses the compatibility between a specific date and a proposed schedule.
A critical innovation is the inclusion of a noise parameter ($\rho$). This parameter allows the model to remain robust against "off-schedule" aberrations. If an event moves one year due to an outlier event, the model can "pay a penalty" to ignore that instance rather than compromising the accuracy of the overall periodic rule.
Key Results and Insights
TRINE's performance was evaluated against 4,350 recurrent events from Freebase. The results demonstrate the power of joint inference:
- Predicting the Future: TRINE achieved a Mean Reciprocal Rank (MRR) of 0.388 for future event prediction, significantly outperforming human workers on Amazon Mechanical Turk (0.308). Humans often struggle to aggregate years of calendar data to identify complex patterns like "the weekend after the third Monday."
- KB Completion: For missing instances where text extractions were available, TRINE reached an MRR of 0.508, typically placing the correct date in the top two predictions.
- Ablation Findings: The study revealed that while textual schedules are helpful, historical data remains the most critical source for high-precision predictions.
Conclusion
TRINE represents a significant step forward in automated KB maintenance. By modeling the inherent noise in real-world schedules and leveraging joint inference across structured and unstructured data, it achieves superhuman accuracy in temporal prediction. Future work aims to expand the system's grammar to handle multi-year cycles (like the World Cup) and integrate human-in-the-loop verification to further streamline the curation of global event knowledge.
Summary updated