Abstract
This paper introduces a formal framework for determining the chronological order of events in natural language by combining discourse relations with defeasible (default) reasoning. By modeling world knowledge and pragmatic maxims as rules that can be overridden by specific information, the authors explain why syntactically identical sentences—such as "Max fell. John pushed him" versus "Max stood up. John greeted him"—result in opposite temporal interpretations.
Introduction: The Temporal Paradox of Syntax
In natural language, the linear order of sentences does not always correspond to the chronological order of the events they describe. Consider the following pair:
- Max stood up. John greeted him. (Event A precedes Event B)
- Max fell. John pushed him. (Event B precedes Event A)
Syntactically, these examples are identical, yet they evoke different temporal realities. Traditional linguistic theories struggled with this "backwards movement of time." If a system assumes events always follow the text's order, it fails on the second example; if it allows any order, it lacks predictive power. This paper addresses this challenge by proposing a formal logic that derives temporal order as a byproduct of discourse relations.
The Formal Framework: Defeasible Logic and MASH
The authors utilize a logic for non-monotonic reasoning called MASH. Unlike standard logic, where new information never invalidates old conclusions, non-monotonic logic allows for "defeasible" rules—conclusions that are true by default but can be retracted if more specific evidence appears.
A key heuristic used is the Penguin Principle, where specific information overrides general information. In linguistics, Paul Grice’s "Maxim of Manner" (be orderly) is formalized as a defeasible rule: assume a Narration relation (events in order) unless there is a reason not to. When world knowledge provides a causal link (e.g., pushing causes falling), this specific "Causal Law" overrides the default narration, triggering an Explanation relation instead.
Discourse Relations as Temporal Anchors
The core insight of the paper is that temporal interpretation is a side effect of identifying the relationship between sentences. The authors define five primary relations: Narration, Elaboration, Explanation, Background, and Result. Each relation carries a specific temporal constraint (e.g., Narration implies $e_1 < e_2$, while Explanation often implies $e_2 < e_1$).
As illustrated in the figure above, discourse is structured hierarchically. An "Elaboration" provides more detail about a previous event, while "Narration" moves the timeline forward within that detail. The model successfully blends linguistic knowledge (syntax and tense) with world knowledge (causality) within this single logical framework.
Structural Constraints and Discourse Attachment
To manage complex narratives with sub-plots, the framework tracks "open" and "closed" nodes in a discourse tree. A new sentence can only attach to an "open" node. If a logical conflict arises—such as a "Nixon Diamond" where two defaults conflict—the system "pops" back up the tree to find a valid attachment site.
This structural approach explains why certain sequences feel "awkward" or incoherent: they attempt to attach information to a part of the discourse that has already been "closed" by a subsequent topic shift.
Conclusion
By integrating linguistic knowledge with defeasible world knowledge, the authors provide a rigorous foundation for computational discourse analysis. The paper demonstrates that temporal order is not a primary feature of text but an emergent property of the underlying discourse structure. While future work is needed to address explicit temporal connectives (like "before" or "after") and computational complexity, this framework offers a powerful tool for understanding how humans navigate complex stories.
Summary updated