World Knowledge Resolves Some Aspectual Ambiguity

Katarzyna Prus, Mark Steedman, Adam LopezACL 2025

Introduction

This paper investigates the "imperfective paradox," a long-standing linguistic puzzle regarding whether the past progressive form of a verb (e.g., "I was doing X") logically entails its past simple counterpart (e.g., "I did X"). While linguistic theory provides clear predictions based on event types, human intuition often diverges from these models. This study explores whether providing explicit world-knowledge context—specifically, evidence of an interruption—can align human judgments with formal aspectual theory.

Background: The Imperfective Paradox

In linguistics, aspect describes the internal temporal structure of an event. Events are typically categorized into three primary types based on their duration and telicity (whether they have a natural endpoint):

  • Activities: Atelic and durative (e.g., "to dance"). They have no inherent endpoint; if you were dancing, you have danced.
  • Accomplishments: Telic and durative (e.g., "to build a house"). They have a natural culmination; if you were building a house, you might not have finished it.
  • Achievements: Telic and punctual (e.g., "to land"). These are near-instantaneous transitions to a state.
Classification of event types based on telicity and duration.

The paradox arises because, for Accomplishments, the progressive does not necessarily entail the simple past. Previous research found that non-experts often ignore this, assuming that common actions reach completion. The authors hypothesize that humans use world knowledge to "fill in the blanks," and that providing explicit context about an interruption will break this default assumption.

Methodology

The researchers conducted a human study using 46 predicates categorized into Activities, Accomplishments, Achievements, and "Contested" cases. They introduced a template to test entailment under interruption: "If the sentence 'I was Xing, but [Interruption]' is true, does it necessarily mean 'I Xed' is also true?"

Participants used a slider interface (ranging from -50 for "Definitely No" to 50 for "Definitely Yes") to indicate their confidence. To compare human intuition with computational models, the researchers also prompted GPT-4 with the same scenarios across multiple temperature settings.

Results and Analysis

The study revealed that explicit context significantly shifts human interpretation, though the effect varies by predicate type.

Comparison of human judgment distributions between decontextualized (top bar) and interrupted (bottom bar) settings across predicate categories.
  • Activities: These remained stable. Interruption does not negate the fact that the activity occurred (e.g., if you were listening to music but were interrupted, you still "listened to music").
  • Achievements: This category showed the most significant shift toward "No" answers. Explicitly stating that a punctual event was interrupted (e.g., "I was catching the ball, but my hand cramped") successfully aligned human intuition with the theory that the event did not reach completion.
  • Accomplishments: Results here were mixed. While some shifted toward "No," others like "read a book" remained stubbornly ambiguous. The authors suggest these are "underspecified," where some participants interpret the phrase as an Activity (the act of reading) rather than a completed Accomplishment.

Discussion: LLMs and Human Intuition

A key finding was the divergence between human and LLM behavior. While GPT-4 followed general aspectual trends, it was far more "extreme" and sensitive to context than humans. For instance, GPT-4 often shifted from 100% "Yes" to 100% "No" upon the addition of context, failing to capture the nuanced distribution of human uncertainty. Furthermore, GPT-4 failed to predict human answers on an individual predicate level, highlighting a lack of grounded real-world experience in current language models.

Conclusion

The study demonstrates that while world knowledge and explicit context can resolve some aspectual ambiguity—particularly for Achievements—many predicates remain open to individual interpretation. This suggests that linguistic categories like "Accomplishment" may be more fluid in common usage than formal theory suggests, and that aligning AI with human common-sense reasoning remains a significant challenge.

Summary updated

BibTeX
@inproceedings{prus2025world,
  author = {Katarzyna Prus and Mark Steedman and Adam Lopez},
  title = {{World Knowledge Resolves Some Aspectual Ambiguity}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure