Human language is inherently non-linear and full of temporal ambiguity. When a text mentions "the meeting," "the explosion," or "next Tuesday," how does a natural language processing system determine exactly when these events occur relative to a real-world timeline?
The core insight of this research direction is that grammatical tense and aspect do not merely modify individual verbs; they serve as structural anchors and directional compasses that propagate temporal meaning across entire sentences, syntax trees, and discourse boundaries. By treating tense as a relational framework rather than a localized morphological feature, researchers have developed robust methods for anchoring relative expressions, resolving untensed elements, and segmenting narratives into coherent event timelines.
Tense as a Directional Compass for Relative Expressions
Consider expressions like "on Monday" or "next month." Without context, these phrases are completely ungrounded. However, the grammatical tense of the surrounding matrix verb provides a critical directional heuristic.
As demonstrated by Mani et al. in Robust Temporal Processing of News, a past-tense verb acts as a backward offset vector from a reference date, whereas a future-tense marker indicates a forward shift. For instance, "the committee met on Monday" signals a date prior to the document creation time, while "the committee will meet on Monday" points forward.
To formalize this computationally, Lee et al. integrated the tense and part-of-speech information of the governing verb into a context-dependent semantic parser in Context-dependent Semantic Parsing for Time Expressions. By evaluating the structural relationship between the verb and the time phrase, the model determines the mathematical directionality of the temporal offset.
Even in short-form, noisy media like Twitter, Qin et al. showed in Feature-Rich Segment-Based News Event Detection on Twitter that the presence of past-tense verbs serves as a high-signal indicator to distinguish completed, actual news events from ongoing chatter or future-oriented noise.
Syntactic Propagation to Untensed Elements
Many linguistic elements carry no inherent tense. Nouns ("the explosion"), adjectives ("a warm cake"), and nominalizations ("the failure") represent events or states but lack morphological markers. How do they get placed on a timeline?
The solution lies in syntactic propagation down the parse tree. Dalrymple established a foundational framework in The Interpretation of Tense and Aspect in English, demonstrating how a temporal variable from the main event is passed through the syntactic tree to provide a reference time for non-tensed modifiers. Similarly, Dahl et al. showed in Nominalizations in PUNDIT that the timing of a nominalized action can be systematically derived from the tense of its governing matrix verb.
[Matrix Verb: "announced" (PAST)] <-- Temporal Anchor
|
+-------------+-------------+
| |
[Noun: "the failure"] [Noun: "the strike"]
(Inherits PAST) (Inherits PAST)
When an event is realized as an infinitive or a noun phrase, Huang et al. utilized the surrounding matrix verb or specialized future-oriented lexicons as "temporal projectors" to determine whether an event is past, ongoing, or planned (Distinguishing Past, On-going, and Future Events: The EventStatus Corpus).
This structural inheritance even applies to "tenseless" languages like Chinese, which lack overt morphological inflections. Xue demonstrated in Automatic Inference of the Temporal Location of Situations in Chinese Text that the temporal location of nested clauses can be inferred by treating the state of the governing verb in the higher-level clause as a feature, propagating temporal context from the macro-structure down to embedded situations.
Avoiding the Simultaneity Trap via Local Scope
Early semantic frameworks often treated tense as a global operator that shifted the evaluation of an entire sentence. However, this creates a "simultaneity trap," forcing every entity and action in the sentence to exist at the exact same point in time. For example, in the sentence "The president was a student in 1980," a global operator might incorrectly imply that the individual was president while being a student.
To resolve this, Hinrichs proposed a compositional semantics in A Compositional Semantics of Temporal Expressions in English where tense is applied locally to the time index of the verb rather than globally. This narrow-scope approach allows nouns and verbs to be evaluated at distinct temporal indices. Under this model, temporal adverbials define a specific window of interest, and the local tense locates the event time within that window, cleanly resolving scope conflicts and maintaining discourse coherence.
Discourse Segmentation and Narrative Planes
Beyond individual sentences, shifts in grammatical tense serve as primary signals for mapping the macro-structure of a document. A transition from the past tense to the present perfect or future tense often signals a shift in the narrative plane—such as moving from a primary news event to historical background or future implications.
Jean‐Louis et al. leveraged this insight by using linear Conditional Random Fields (CRFs) to partition news articles into functional zones like "Main Event," "Secondary Event," and "Background Context" (Utilisation d'indices temporels pour la segmentation événementielle de textes, Text Segmentation and Graph-based Method for Template Filling in Information Extraction). By isolating the primary narrative plane via tense transitions, downstream information extraction systems can filter out noisy, irrelevant historical context (Besançon et al. in Une méthode d'extraction d'information fondée sur les graphes pour le remplissage de formulaires).
This macro-level understanding is further supported by identifying the "uppermost verb" in a query or sentence parse tree. As shown by Yu et al. in TUTA1 at the NTCIR-11 Temporalia Task, isolating the main predicate allows models to capture the primary temporal weight of an utterance, ignoring subordinate clauses that might otherwise skew the temporal classification.
Scaling to Global Context and Neural Architectures
When local sentences lack explicit markers entirely—either due to telegraphic speech or language constraints—the system must look to global discourse context. For instance, Linebarger et al. showed that for sentence fragments lacking explicit tense (common in technical logs), temporal values can be assigned based on the semantic category of the predicate (e.g., mapping event-based transitions to a default past tense) to maintain logical consistency (Sentence Fragments Regular Structures).
In machine translation and grammar correction, models use the tense of preceding sentences and document-level genre priors to resolve local ambiguities (Tajiri et al. in Tense and Aspect Error Correction for ESL Learners Using Global Context; Zhang et al. in Classifier-Based Tense Model for SMT).
Modern contextualized language models naturally excel at this task. As demonstrated by Ross et al. in Exploring Contextualized Neural Language Models for Temporal Dependency Parsing, fine-tuning architectures like BERT allows self-attention mechanisms to implicitly capture these long-range tense and discourse-level cues, mapping complex timelines with unprecedented accuracy.
Key Foundational Work
Some Linguistic Aspects for Automatic Text Understanding (1984) — contributed 1 concept elements to this direction.
A Model of Natural Language Processing of Time-Related Expressions (1980) — contributed 1 concept elements to this direction.