Compositional Temporal Semantics in Natural Language Processing

80 concepts54 papers1 proposal

When humans talk about time, we rarely rely on simple timestamps. Instead, we construct complex, layered temporal meanings out of simpler linguistic building blocks. Consider the difference between "John ran," "John ran a marathon," and "John was running a marathon every Saturday since March."

The first sentence describes an unbounded past activity. The second introduces a specific object ("a marathon"), which dynamically creates a logical endpoint—a property known as telicity. The third sentence adds grammatical aspect ("was running"), iteration ("every Saturday"), and a relational anchor ("since March").

Capturing how these pieces fit together is the core challenge of compositional temporal semantics. Traditional Natural Language Processing (NLP) models often treat temporal expressions as flat strings or assign them simple categorical labels like before or after. However, a long line of research demonstrates that to truly understand, reason about, and translate human language, computational systems must model how tense, aspect, verbs, and syntactic arguments interact compositionally to construct a unified timeline.

Dynamic Aspectual Determination and Event Boundaries

A fundamental hurdle in temporal processing is that a verb's inherent meaning (its lexical aspect) can be radically altered by its surrounding context. For instance, the verb "build" implies an ongoing activity, but "build a house" defines a complete event with a strict boundary.

To capture this computationally, Dorr et al. developed an algorithmic framework using Lexical Conceptual Structure (LCS) to dynamically derive aspectual properties at both the individual verb level and the full sentence level. This allows a system to update the temporal meaning of a sentence compositionally without conflicting with the base meaning of the verb.

The compositionality of event boundaries becomes even more apparent when machine learning models attempt to classify telicity. As Friedrich et al. demonstrate, the aspectual nature of a sentence depends heavily on the entire verb phrase rather than the verb alone; the presence or absence of specific objects shifts the event's boundaries. Furthermore, modifiers like prepositional phrases play a crucial role in determining whether an event is interpreted as finished or ongoing, a phenomenon that Schilder formalized using situation theory and interval calculus.

Contextual composition also governs our understanding of event duration. As Yang et al. show, the expected timeframe of an action cannot be reliably predicted from a verb in isolation. Variations in the subject, object, or setting can change the expected duration by orders of magnitude, requiring models to learn "temporal commonsense" from contextualized representations.

Multi-Dimensional Temporal Logics

If simple verbs cannot be understood in isolation, how should we represent the broader temporal structure of a sentence? Early natural language understanding systems often extracted tense and aspect but failed to integrate them into the final semantic logic, limiting their capacity for temporal reasoning (Clark et al.).

To bridge this gap, researchers adapted formal linguistic frameworks—most notably Hans Reichenbach’s 1947 tense theory—into computational models. Reichenbach's system maps verb tenses into a language-independent framework defined by the relationships between three distinct points in time:

  • Speech Time (S): The moment the sentence is uttered.
  • Event Time (E): The moment the action occurs.
  • Reference Time (R): The perspective or vantage point from which the event is viewed.
Past Perfect: "John had eaten"
   E ----------- R ----------- S
(Event occurs)  (Vantage point) (Speech time)

By using this tri-point representation, Horie et al. created a semantic interlingua that captures the underlying temporal logic of a sentence regardless of its specific surface syntax. Similarly, Derczynski et al. demonstrated that integrating Reichenbachian reference times into machine learning models consistently improves their ability to classify the chronological order of events in text, moving beyond restrictive categorical labels like before or after (Bethard et al.).

When generating natural language or interpreting complex narratives, these reference points are frequently combined with interval algebras. For example, Dorr et al. combined Allen's interval algebra with neo-Reichenbachian tense theory to ensure that natural language generation systems select grammatically correct tenses and connecting words (like "before" or "while") based on raw timestamped data. This multi-dimensional approach allows systems to resolve highly complex narrative timelines, such as the past future perfect tense, which traditional linear models fail to map accurately (Song et al.).

Parsing Complex Temporal Expressions

Beyond verbs and tenses, natural language contains multi-word temporal expressions (e.g., "the day before yesterday" or "every Saturday since March 6") that must be normalized into standardized formats like 2013-04-12.

Treating these phrases as flat strings ignores their compositional nature, where complex expressions are built from simpler, recurring temporal building blocks. To address this, Bethard introduced a compositional approach using a Synchronous Context Free Grammar (SCFG) to map natural language directly to a formal math of temporal operators. This insight was expanded in the SemEval 2018 SCATE schema task (Laparra et al.), which challenged systems to parse complex time expressions into structured semantic graphs. Because expressions like "the following day" and "the following year" share an underlying compositional logic, parsing them as structured graphs allows models to generalize across shared patterns rather than learning every variation independently.

"Every Saturday since March 6"
              │
      ┌───────┴───────┐
  Repeating       Bound Anchor
      │               │
[Saturday]       [March 6]

When building logic-based inference systems, handling these multi-word expressions requires tight integration with syntax. For example, the authors of a recent system based on Japanese Combinatory Categorial Grammar (CCG) demonstrated that multi-word temporal expressions must be unified into single semantic units during syntactic tree modification (Compositional Semantics and Inference System for Temporal Order based on Japanese CCG). This unification prevents variable mismatch in logical representations, ensuring that complex date and time phrases are treated as a single interval during the reasoning process.

Cross-Linguistic and Structural Challenges

The necessity of a compositional approach becomes starkest in Machine Translation (MT) and cross-linguistic analysis. As Appelo argues, tense, aspect, and time adverbials are so deeply intertwined that they cannot be translated in isolation; their translation must be handled jointly via shared semantic derivation trees to preserve nuanced temporal meanings across languages.

This challenge is magnified when translating between languages with fundamentally different grammatical structures. For instance, when translating from a language without overt tense markers (like Chinese) to a tensed language (like English), Zhang et al. showed that models must use broader contextual features to compositionally predict the correct grammatical aspect (such as progressive or perfective forms).

Finally, compositionality must account for complex syntactic structures like nested clauses. In Japanese, relative clauses do not always determine their timing based on the clause immediately above them. Instead, as Matsuoka et al. demonstrate, they can "skip" levels to find the nearest higher clause with an opposite tense, or default to the main clause. Modeling these interactions requires a unified formal grammar that can dynamically compute evaluation times across both dynamic and stative predicates.

By treating time not as a static attribute but as a dynamically assembled semantic structure, modern computational linguistics continues to close the gap between high-level linguistic theories and the practical requirements of automated reasoning systems.

Key Foundational Work

International Conference on Computational Linguistics COLING 1969: Preprint No. 60 (1969) — contributed 4 concept elements to this direction.
The ‘Time Category’ in Natural Languages and Its Semantic Interpretation (1969) — contributed 2 concept elements to this direction.
Uber Zeitkeferenz Und Tempus (1969) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_compositionality_of_temporal_expressions_de2249f5,
  title = {compositionality of temporal expressions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/compositionality-of-temporal-expressions/rdr_de2249f5db69449d8d8b790951309f5d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_compositionality_of_temporal_expressions_de2249f5_v1,
  title = {compositionality of temporal expressions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/compositionality-of-temporal-expressions/rdr_de2249f5db69449d8d8b790951309f5d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake