Abstract
Leon Derczynski, Robert Gaizauskas. Proceedings of the 10th International Conference on Computational Semantics (IWCS 2013) – Long Papers.
2013.
This paper provides the first large-scale empirical validation of Hans Reichenbach’s 1947 theory of tense. By mapping Reichenbach’s abstract "three-point" framework to modern human-annotated datasets, the authors demonstrate that the theory accurately predicts how events are ordered in natural language, achieving up to 91.9% consistency in specific contexts.
Introduction
In the field of Natural Language Processing (NLP), temporal reasoning—the ability to automatically determine the chronological order of events in a narrative—is often considered a "final boss" challenge. While humans intuitively grasp the flow of time in a story, machines frequently struggle with the nuances of tense and aspect.
In 1947, philosopher Hans Reichenbach proposed a seminal framework for describing tense using three distinct points: Speech time (S), Event time (E), and Reference time (R). While this theory has been a cornerstone of theoretical linguistics for decades, it remained largely abstract. Derczynski and Gaizauskas (2013) address this gap by asking a fundamental question: Does Reichenbach’s framework actually hold up when compared to how humans annotate real-world news text? If validated, this framework could significantly improve how AI systems understand narrative time.
Methodology
The authors bridge the gap between mid-century linguistic theory and modern computational linguistics through several key innovations.
Mapping Schemas
The first step involved translating the modern TimeML annotation standard, which is the current benchmark for NLP temporal tasks, into Reichenbach’s S/E/R structures. For instance, the "Past Perfect" tense (e.g., had slept) is mapped to a structure where the Event (E) occurs before the Reference time (R), which in turn occurs before the Speech time (S), denoted as $E < R < S$.
Semi-Interval Algebra and Context Modeling
Natural language is inherently "fuzzy." To account for this, the authors moved beyond strict point-based logic to Freksa’s semi-interval temporal algebra. This allows for "disjunctions"—situations where the start of an event is known relative to another, but its end point remains underspecified.
Furthermore, the study explores the "permanence of the reference point," a Reichenbachian principle suggesting that a sequence of verbs in a discourse usually shares the same perspective. The authors tested this by modeling temporal context, assuming that verbs within the same or adjacent sentences share a reference point.
Datasets and Validation
The study utilizes TimeBank v1.2, the gold-standard corpus for temporal semantics. Containing 6,418 temporal links and 7,935 event annotations, TimeBank provides a reliable benchmark due to its high inter-annotator agreement. The authors compared the temporal relations predicted by Reichenbach's framework against these human-provided "ground truth" labels.
Key Insights and Results
The empirical results strongly support the validity of the framework:
- The Power of the Reference Point: Most simple temporal models only consider Event and Speech time. Reichenbach’s inclusion of the "Reference" point (the perspective from which an event is viewed) is the critical element that allows the model to handle complex narrative shifts.
- High Consistency: Across all verb pairs, the framework was consistent with human data 81.5% of the time.
- Optimized Performance: When the scope was narrowed to verbs likely sharing a reference point (the "Same SR" model), consistency jumped to 91.9%.
- Constraint Power: Even when the framework did not predict a single specific relation, it successfully narrowed the possibilities to a small, accurate subset of potential temporal relations.
Limitations and Future Directions
The primary limitation identified is the method of Context Modeling. The authors used sentence proximity as a proxy for temporal context, which is a relatively blunt instrument. In complex discourse, reference points can shift due to quoted speech, conditional clauses, or explicit temporal signals like "meanwhile."
Future work involves developing more sophisticated automated methods for detecting the boundaries of temporal contexts. Validating this framework opens doors for improving downstream AI tasks such as machine translation, clinical report summarization, and more coherent natural language generation.
Summary updated