Reasoning with Temporal Intervals and Qualitative Relations

191 concepts139 papers1 proposal

Traditional models of time often treat it as a sequence of discrete points or snapshots. While mathematically convenient, this approach struggles with the inherent vagueness of human language and the complexity of overlapping physical actions. If a person is "cooking dinner," there is no single point in time that captures the activity; rather, it is a span of time with internal structure.

Interval-based temporal representation shifts the fundamental primitive of time from the zero-duration point to the interval. This shift allows systems to reason about duration, persistence, and concurrency without requiring precise, absolute timestamps.

The Interval as a Fundamental Primitive

In foundational work, James F. Allen argued that representing temporal knowledge as a network of relationships between intervals is more natural for AI tasks like natural language understanding. As noted in An Interval-Based Representation of Temporal Knowledge, intervals effectively handle the imprecision and uncertainty common in real-world descriptions.

Allen et al. further proposed a "common-sense" theory where time points are not the building blocks of intervals, but rather the meeting places between them. By defining a single primitive relation—the adjacency of two intervals—a rich library of temporal relationships can be derived. This avoids the boundary paradoxes inherent in point-based systems, where it is often unclear whether a point belongs to the state before or after a change.

Qualitative Algebra for Relative Timing

The power of this approach lies in Allen’s Interval Algebra, a framework of thirteen mutually exclusive relations (such as before, meets, overlaps, during, and starts) that describe how two time spans interact. This algebra allows for qualitative reasoning: if we know Event A occurs during Event B, and Event B occurs before Event C, we can logically infer the relationship between A and C without knowing their exact start or end times.

This logic is particularly useful for resolving complex queries. For example, the TEQUILA framework proposed by Abujabal et al. uses interval algebra to enable temporal question answering over knowledge bases, allowing systems to filter results based on complex constraints like "Which teams did Neymar play for before joining PSG?"

The algebra has been extended to handle more complex scenarios:

  • Non-convex intervals: Morris et al. and Peter B. Ladkin extended the logic to "n-intervals"—sequences of time blocks that represent recurring events like "every Tuesday."
  • Metric distances: Montanari et al. integrated algebraic truths about time distances into the logic, allowing for reasoning about specific durations alongside qualitative relations.

Modeling Composite Activities and Workflows

Interval logic is essential for recognizing complex activities that consist of multiple overlapping sub-events. In computer vision, Wang et al. developed the Interval Temporal Bayesian Network (ITBN) to recognize facial expressions by modeling how muscle movements persist and interact over time.

Similarly, Ryoo et al. demonstrated that using interval logic instead of frame-by-frame analysis allows systems to recognize composite human activities (like "shaking hands") regardless of their duration. By defining activities through the relative timing of constituent actions, the system remains robust even when the speed of the movement varies. This approach prevents the exponential state growth that occurs when trying to model concurrent, overlapping tasks using traditional finite state machines, as noted by Ryoo et al..

Bridging Logic and Natural Language

One of the most significant applications of interval representation is in the generation and understanding of tense and aspect in language. Bonnie J. Dorr and Kathleen McKeown both proposed systems that map temporal intervals to linguistic permutations. By comparing event intervals against speech and reference times, these systems can automatically select the correct grammatical aspect (e.g., distinguishing between "he ran" and "he was running").

Despite the maturity of these logical frameworks, modern Large Language Models (LLMs) still struggle with these concepts. Li et al. introduced the COTEMPQA benchmark, finding that models often fail at "co-temporal" reasoning—understanding events that occur simultaneously. They suggest that framing temporal reasoning as a formal mathematical problem of interval intersections, rather than a linguistic association task, is necessary for models to achieve human-level performance.

Key Foundational Work

What's Necessary to Hide?: Modeling Action Verbs (1981) — contributed 3 concept elements to this direction.
An Interval-Based Representation of Temporal Knowledge (1981) — contributed 2 concept elements to this direction.
An Interval-Based Representation of Temporal Knowledge (1981) — contributed 2 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 142 papers
  2. Version 1 · 2026-10-04 · 169 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_interval_based_temporal_representation_bb601b04,
  title = {interval-based temporal representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interval-based-temporal-representation/rdr_bb601b045c45405dae6f73ae2ace95d0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_interval_based_temporal_representation_bb601b04_v2,
  title = {interval-based temporal representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interval-based-temporal-representation/rdr_bb601b045c45405dae6f73ae2ace95d0/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake