A Survey on Temporal Reasoning for Temporal Information Extraction from Text (Extended Abstract)

Artuur Leeuwenberg, Marie-Francine Moens

Abstract Time is deeply woven into how people perceive and communicate about the world. Almost unconsciously, we provide our language utterances with temporal cues, like verb tenses, and we can hardly produce sentences without such cues. Extracting temporal cues from text and constructing a global temporal view about the order of described events is a major challenge of automatic natural language understanding. Temporal reasoning, the process of combining different temporal cues into a coherent temporal view, plays a central role in temporal information extraction. This article presents a comprehensive survey of the research from the past decades on temporal reasoning for automatic temporal information extraction from text, providing a case study on the integration of symbolic reasoning with machine learning-based information extraction systems.

1. Introduction

Temporal Information Extraction (TIE) aims to identify events, time expressions, and their relationships within natural language. However, human language rarely provides a neat, timestamped log. Instead, it relies on relative cues ("after the meeting"), vague durations ("for nearly an hour"), and overlapping intervals ("during the surgery").

The core challenge in TIE is that local extractions are often globally inconsistent. A machine might predict that Event A happened before B, and B before C, but then mistakenly predict that C happened before A, creating a logical paradox. Temporal Reasoning (TR) is the process of combining these separate, often underspecified cues to derive a coherent, "legal" timeline. This survey explores how the field has evolved from rigid symbolic logic to modern "neuro-symbolic" approaches that integrate temporal constraints directly into machine learning models.

2. Representational Frameworks and Inference Strategies

The paper categorizes the primary frameworks and algorithmic strategies used to solve the temporal puzzle.

2.1 Representational Frameworks

  • Allen’s Interval Algebra (AIA): The classic 1983 framework defining 13 basic relations between time intervals (e.g., overlaps, meets, contains). While expressive, reasoning over the full set is computationally expensive (NP-complete).
  • Point Algebra (PA): A more efficient alternative that treats events as sets of start and end points. By reasoning about whether point $x < y$ or $x = y$, researchers can build faster systems that still capture most linguistic nuances.
Allen's Interval Algebra relations mapped to point-based constraints.

2.2 Inference Strategies

  • Greedy & Post-Hoc: Adding relations one-by-one or fixing contradictions after a model has made its initial guesses.
  • Integer Linear Programming (ILP): Using hard mathematical constraints to "cut off" impossible timelines during the prediction phase.
  • Markov Logic Networks (MLN): A probabilistic approach that treats temporal rules as "soft" constraints, allowing the model to learn which rules are most important.
  • Direct Timeline Models: A newer shift where models predict absolute coordinates for event endpoints directly, rather than classifying every possible pair of events, which improves scalability from quadratic ($O(n^2)$) to linear ($O(n)$).

3. Datasets and Annotation Standards

The evolution of TIE has been driven by the development of standardized annotation schemes and corpora:

  • TimeML: The industry standard for annotating events and "TLinks" (temporal links).
  • TimeBank: A widely used corpus based on TimeML.
  • TempEval: A series of shared tasks/competitions that have served as the primary benchmarks for TIE systems.
  • Recent Variations: Newer datasets move beyond simple interval relations to include "vague" relations, point-wise relations, and absolute calendar anchoring.

4. Key Insights and Results

  • The Efficiency Trade-off: There is a constant tension between expressivity (using the full Allen Algebra) and tractability. Most successful modern systems use "tractable sub-fragments" of logic, like Point Algebra, to maintain reasonable computation times.
  • Reasoning as a Training Signal: TR isn't just for post-processing. Incorporating temporal logic during the training phase (e.g., using structured perceptrons or ILP-augmented loss functions) helps models learn the underlying structure of time more effectively than just looking at local features.
  • The "Dense Graph" Problem: Annotators often miss many implicit temporal relations. Using TR to calculate the "transitive closure" (inferring that if A < B and B < C, then A < C) can help densify training data, though it must be done carefully to avoid introducing noise.
  • Global Reasoning is Essential: Systems that incorporate global reasoning consistently outperform those that rely solely on local, pairwise classifiers. Sieve-based architectures—applying high-precision rules first, then moving to machine learning—are particularly effective.

5. Limitations and Future Work

The authors identify several gaps in current research:

  • Fragmented Approaches: Most current models focus on specific types of cues (e.g., just event-event relations) rather than extracting events, times, and durations jointly.
  • The Need for Flexibility: There is a lack of frameworks that can handle the full spectrum of temporal information—absolute dates, relative orders, and vague quantifiers—in a single, efficient model.
  • Neuro-symbolic Integration: The next frontier is better "neuro-symbolic" integration: combining the representational power of deep learning with the logical guarantees of symbolic reasoning.

Summary updated

Abstract

Time is deeply woven into how people perceive, and communicate about the world. Almost unconsciously, we provide our language utterances with temporal cues, like verb tenses, and we can hardly produce sentences without such cues. Extracting temporal cues from text, and constructing a global temporal view about the order of described events is a major challenge of automatic natural language understanding. Temporal reasoning, the process of combining different temporal cues into a coherent temporal view, plays a central role in temporal information extraction. This article presents a comprehensive survey of the research from the past decades on temporal reasoning for automatic temporal information extraction from text, providing a case study on the integration of symbolic reasoning with machine learning-based information extraction systems.
BibTeX
@inproceedings{leeuwenberg2020survey,
  author = {Artuur Leeuwenberg and Marie-Francine Moens},
  title = {{A Survey on Temporal Reasoning for Temporal Information Extraction from Text (Extended Abstract)}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=DFsbu6FdRa2},
}
Figure