Tense Interpretation in the Context of Narrative

Fei Song, Robin CohenAAAI 1991

Introduction

Determining the chronological order of events in a narrative is a complex task for computational systems. While humans instinctively grasp temporal sequences, a computer must decode the linguistic cues embedded in verb tenses. This paper, originally published in the AAAI-91 proceedings, proposes a formal computational framework for interpreting tense within narrative contexts. The central thesis is that tense is anaphoric—its meaning is derived from the preceding discourse context rather than being absolute.

The SRE Framework and Tense Hierarchy

The authors build upon Reichenbach’s (1947) classic system, which characterizes tense using three temporal points: S (Speech time), R (Reference time), and E (Event time). To resolve ambiguities in previous models, the authors refined the SRE definitions for nine English tenses into precise triples (e.g., defining the Future Perfect as $[S < R > E]$).

A key innovation is the organization of these tenses into a Tense Hierarchy. This tree structure represents the logical relationships between different temporal configurations. In this model, moving down a branch signifies the creation of a new temporal focus (such as a flashback or elaboration), while remaining at a node represents maintaining the current focus.

The Tense Hierarchy tree structure used to model temporal focus movement.

Temporal Focus Structure (TFS)

To manage the shifting focus of a narrative, the system utilizes a Temporal Focus Structure (TFS) implemented as a stack. As the story progresses:

  1. Push: When a new time period is introduced (e.g., a "Past Perfect" introducing a prior event), it is pushed onto the stack.
  2. Pop: When the narrative returns to a previously established time, the stack is popped to restore the earlier focus.

This stack-based approach allows the algorithm to handle complex narrative movements, such as jumping back to the present after a detailed flashback, by treating both Reference (R) and Event (E) times as variables that refer back to the context stored in the TFS.

Evaluation and Results

The algorithm was evaluated using 20 short narratives from Real Stories (Katz et al., 1975). To isolate the performance of tense-interpretation rules, the authors pre-processed the texts to remove non-actual situations like indirect speech.

The system was tested in two modes: a fully automated round using only tense and aspect heuristics, and a user-assisted round where minimal hints were provided for difficult focus shifts (simulating the role of temporal adverbs like "before" or "after").

Performance metrics of the algorithm across 20 narrative stories.

As shown in the results, the automated system correctly interpreted an average of 14.15 out of 16.15 utterances per story (approximately 87%). With user assistance, accuracy rose to 16.0 out of 16.15, suggesting that while tense provides the majority of the logic needed for temporal ordering, full narrative comprehension requires integration with discourse connectives and world knowledge.

Conclusion

This research demonstrates that a hierarchical, stack-based model of tense can effectively reconstruct the temporal flow of simple narratives. By treating tense as a set of instructions for moving through a focus structure, the authors provide a robust foundation for more advanced temporal reasoning in Natural Language Processing. Future work aims to integrate these linguistic heuristics with quantitative time data and explicit temporal connectives.

Summary updated

BibTeX
@inproceedings{song1991tense,
  author = {Fei Song and Robin Cohen},
  title = {{Tense Interpretation in the Context of Narrative}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure