Abstract: This paper presents a formal framework for the computational processing of temporal information in English. By defining compositional rules for tenses, adverbs, and temporal connectives based on point-based logic, the authors demonstrate how a system can logically reason about event ordering and automatically detect temporal inconsistencies in natural language.
Introduction
Human language communicates time with remarkable nuance, utilizing complex constructions like the past perfect progressive ("Jack had been running") to establish intricate relationships between distinct points in time. For computational linguistics, this presents two primary challenges: extraction—transforming strings into formal representations—and validation—recognizing logically impossible sentences, such as "I will run yesterday."
Previous symbolic AI models often struggled with compositionality, where the meaning of a complex sentence is derived from its constituent parts. This research addresses these gaps by proposing a flexible, point-based logic capable of handling events that cross tense boundaries.
Formal Framework
The proposed framework is built upon McDermott’s (1982) temporal logic, which treats time as a series of points rather than intervals. To represent a sentence, the model identifies three core event types:
- Utterance Event: The moment of speech.
- Main Event: The primary action described.
- Perfect/Progressive Events: Intermediate intervals used as reference points for complex tenses.
Compositional and Anchoring Rules
The authors define five basic rules (Past, Present, Future, Perfect, and Progressive) using relational operators ($<, >, =$) to constrain event boundaries. For instance, the "Past Rule" dictates that an event must conclude before the Utterance Event begins. To handle adverbs like "yesterday," the system employs an "Anchoring Rule," where the first verb acts as a temporal anchor. If an adverb cannot logically overlap with this anchor, the sentence is rejected as erroneous.
Key Insights and Results
A significant finding is that the "Present" rule must vary based on whether a verb is stative (e.g., "to have") or non-stative (e.g., "to leave"), as stative verbs allow for greater flexibility with future-leaning adverbs. Additionally, the authors argue that temporal connectives like "before" and "after" are not simple logical inverses in natural language, but carry subtle perceptual differences.
The primary result of this work is a formal system that successfully categorizes the compatibility of various tenses and adverbs. As shown in the compatibility matrix below, the model correctly identifies errors in combinations such as "Past + Tomorrow."
Furthermore, the model demonstrates that the connective "when" functions dynamically based on the tense of the clauses it joins, either aligning events exactly or nesting one within the duration of another.
Limitations and Future Work
While robust for intra-sentence logic, the model currently lacks the ability to handle the temporal flow of multi-sentence narratives. It also excludes habitual actions (e.g., "I run every Tuesday") and does not account for the inherent temporal aspects of specific lexical items like "plan" or "finish." Future iterations will need to address these context-dependent nuances and move beyond purely compositional rules for certain complex adverbs.
Summary updated