Abstract
Previous work on ordering events in text has typically focused on local pairwise decisions, ignoring globally inconsistent labels. However, temporal ordering is the type of domain in which global constraints should be relatively easy to represent and reason over. This paper presents a framework that informs local decisions with two types of implicit global constraints: transitivity (A before B and B before C implies A before C) and time expression normalization (e.g. last month is before yesterday). We show how these constraints can be used to create a more densely-connected network of events, and how global consistency can be enforced by incorporating these constraints into an integer linear programming framework. We present results on two event ordering tasks, showing a 3.6% absolute increase in the accuracy of before/after classification over a pairwise model.
Introduction
Understanding the chronological flow of a story is fundamental to document comprehension. However, most machine learning models treat temporal ordering as a series of isolated, local decisions: Is Event A before Event B? Is Event B before Event C?
The primary challenge is that local classifiers often produce "impossible" timelines—for example, concluding that A is before B, B is before C, but C is before A. While the Timebank Corpus provides a foundation for this research, it is notoriously sparse; human annotators only label a fraction of possible relations, leaving the "event graph" largely disconnected. Without a way to enforce global consistency and fill in these missing links, models struggle to build a coherent chronological narrative.
Methodology
The authors propose a three-pronged global framework to move beyond simple pairwise classification:
- Pairwise SVM Classifiers: Standard Support Vector Machines (SVMs) are used to predict relations based on local features like tense, aspect, part-of-speech tags, and syntactic dominance.
- Global Optimization (ILP): Instead of simply accepting the SVM's top choice for each pair, the system uses Integer Linear Programming to find the most likely set of labels for the entire document that satisfies transitivity constraints (e.g., if $A < B$ and $B < C$, then $A < C$).
- Temporal Reasoning & Normalization: To solve the "sparsity" problem, the authors introduced a rule-based system to normalize time expressions (e.g., resolving "last Friday" and "Oct 13" to specific dates). By linking events to these normalized timestamps, they created "bridges" that connected previously isolated parts of the event graph.
Key Insights and Results
Connectivity as the Bottleneck
A critical discovery was that global optimization (ILP) provides virtually zero gain on raw Timebank data. The graph is so fragmented that transitivity constraints rarely have three connected events to act upon. By resolving dates and times (Time-Time relations), the authors inferred a massive number of new Event-Event relations, increasing training data by 81% and providing the density needed for global constraints to be effective.
Performance Gains
On the core before/after classification task, the global model (incorporating time normalization and supplemental data) improved accuracy from 66.8% to 70.4%—a statistically significant 3.6% absolute increase. The automated time normalization procedure added 2,459 new relations to the corpus, compared to only 157 tagged by human annotators. Furthermore, the ILP framework successfully "repaired" local contradictions, consistently outperforming the pairwise baseline across varying levels of data noise.
Conclusion
This work demonstrates that temporal ordering is significantly improved by moving from local to global reasoning. By using time normalization as a "universal glue," the authors were able to bridge disconnected event clusters, allowing transitivity constraints to enforce document-wide consistency. Future work should focus on better identifying "unknown" relations—pairs that are truly ambiguous versus those simply missed by annotators—and developing more sophisticated implicit constraints to further tighten the temporal network.
Summary updated