Inducing Temporal Relations from Time Anchor Annotation

Fei Cheng, Yusuke Miyao

Abstract

Traditional temporal relation annotation relies on pairwise comparisons, which is complex, sparse, and inconsistent. This paper proposes Time Anchor Annotation, where events are assigned absolute dates. This linear approach allows for the automatic induction of a dense temporal relation graph (TORDER), significantly improving performance in downstream tasks and reducing label vagueness.

Introduction

Understanding the temporal order of events is fundamental for NLP tasks like timeline generation. The standard approach uses Temporal Links (TLINKs) to manually label relations between event pairs. However, this method suffers from quadratic complexity, annotator bias toward "salient" pairs, and high inconsistency.

Methods: From Pairs to Anchors

The authors shift the focus from relative comparisons to absolute positioning on a timeline.

Time Anchor Annotation

Annotators assign an absolute time value to each event. The system distinguishes between single-day events and multi-day spans. Crucially, it supports imprecise anchors (e.g., "before 1998-02-06") when exact dates are unavailable.

Visualization of precise and imprecise time anchors for events on a timeline.

Relation Induction (TORDER)

Once events are anchored, Temporal Order (TORDER) relations (e.g., BEFORE, AFTER, SAME_DAY) are automatically generated for all pairs using logical rules. This transforms a quadratic annotation task into a linear one, providing inter-sentence relations for free.

Key Results and Insights

The proposed method was evaluated on the TimeBank-Dense corpus.

Improved Label Distribution

Compared to traditional TLINKs, the induced TORDER relations are much more balanced. Most notably, the "vague" label—often a default for annotators in traditional datasets—is significantly reduced because anchoring reveals clear temporal relationships that were previously overlooked.

Comparison of relation distributions between induced TORDER (a) and traditional TLINK (b) datasets.

Performance Gains

Models trained on TORDER data showed a 14.1 point increase in exact match accuracy for time anchor prediction. The method proved particularly effective at narrowing down time windows for imprecise events.

Conclusion

While the current one-day granularity limits intra-day ordering, the transition to time anchors provides a more efficient and consistent framework for temporal reasoning in news-like domains. Future work may explore finer granularities or hybrid models to handle narrative prose where absolute dates are scarce.

Summary updated

BibTeX
@inproceedings{cheng2018inducing,
  author = {Fei Cheng and Yusuke Miyao},
  title = {{Inducing Temporal Relations from Time Anchor Annotation}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
  url = {https://openreview.net/forum?id=6gxLyVfegkX},
}
Figure