Abstract
We propose TRACIE, a novel temporal reasoning dataset that evaluates the degree to which systems understand implicit events—events that are not mentioned explicitly in natural language text but can be inferred from it. This introduces a new challenge in temporal reasoning research, where prior work has focused on explicitly mentioned events. Human readers can infer implicit events via commonsense reasoning, resulting in a more comprehensive understanding of the situation and, consequently, better reasoning about time. We find, however, that state-of-the-art models struggle when predicting temporal relationships between implicit and explicit events. To address this, we propose a neuro-symbolic temporal reasoning model, SYMTIME, which exploits distant supervision signals from large-scale text and uses temporal rules to combine start times and durations to infer end times. SYMTIME outperforms strong baseline systems on TRACIE by 5%, and by 11% in a zero prior knowledge training setting. Our approach also generalizes to other temporal reasoning tasks, as evidenced by a gain of 1%-9% on MATRES, an explicit event benchmark.
Introduction: The Challenge of Implicit Events
Most temporal reasoning research focuses on events explicitly mentioned in text (e.g., "I ate dinner, then I slept"). However, human communication relies heavily on what is not said. If a story says, "The person tried to brake but it was too late," a human reader automatically infers an implicit event: the person was likely distracted or looking away before the braking attempt.
Current state-of-the-art NLP models struggle with these "latent timelines." While they are proficient at spotting keywords, they often lack the commonsense logic required to place unmentioned events in time. Specifically, models find it very difficult to predict when an event ends, as end-times are rarely stated directly in natural language. This "end-time bottleneck" prevents models from building a cohesive situational model.
Methodology: Distant Supervision and Symbolic Logic
To address these challenges, we introduce a two-pronged approach involving specialized pre-training and a novel neuro-symbolic architecture.
PTNTIME (Pattern-based Pre-training)
Before task-specific training, the model undergoes pre-training on a massive amount of "distant supervision" data. We mined Wikipedia for sentences containing temporal keywords (like "before" or "after") and cross-sentence cues (like specific dates/times). This teaches the model the relative distances between event start times without requiring expensive human annotation.
SYMTIME (The Neuro-symbolic Model)
SYMTIME is the core innovation of this work. Instead of asking a transformer to guess an end-time, SYMTIME employs a symbolic logic rule: End Time = Start Time + Duration.
As shown in the architecture above, the model uses one neural module to predict the distance between two events' start times and another to estimate the typical duration of an event. These predictions are mathematically combined using a differentiable framework inspired by Allen’s interval algebra. This allows the model to reason about end-times even when it has never seen an explicit "end-time" label in its training data.
Experimental Setup
We evaluated our approach using two primary benchmarks:
- TRACIE (TempoRAl Closure InfErence): A new benchmark containing 5.4k human-curated instances based on ROCStories. It uses a textual entailment format: given a story (premise), is the temporal hypothesis (e.g., "Event A starts before Event B") true or false?
- MATRES: An existing benchmark for explicit event ordering used to test if the model's improvements generalize to standard temporal tasks.
Results and Analysis
SYMTIME demonstrated significant improvements over standard language models:
- TRACIE Performance: A standard T5-Large model achieved 67.9% accuracy, while SYMTIME reached 78.9%.
- Zero-Shot Success: In a "zero-knowledge" setting (no training on TRACIE), SYMTIME outperformed the baseline by 11%, proving it learned the underlying logic of time rather than just memorizing dataset patterns.
- Generalization: On the MATRES benchmark, the model showed a 1%–9% gain, suggesting that understanding implicit events enhances the comprehension of explicit ones.
Conclusion
By treating end-time as a derived property (Start + Duration) and leveraging distant supervision, SYMTIME bypasses the lack of explicit temporal data in natural language. While the model currently faces challenges with multi-hop reasoning and causal edge cases (where events are exactly simultaneous), it represents a significant step toward AI that understands the unwritten timelines of human narratives.
Summary updated