The Interpretation of Temporal Relations in Narrative

Fei Song, Robin CohenAAAI 1988

This foundational 1988 paper explores the computational reconstruction of timelines from natural language stories. By synthesizing classical linguistics with interval logic, the authors propose a deterministic algorithm capable of navigating the non-linear and often vague structures of human storytelling.

The Challenge of Narrative Time

When humans read a story, they instinctively construct a coherent timeline. For instance, in the sequence "John went to the store. He had bought a gift earlier," the reader understands that the purchase preceded the trip, despite the reverse order of narration.

For artificial intelligence, this task presents significant hurdles:

  1. Tense Complexity: Traditional models, such as Reichenbach’s, often fail to capture the full spectrum of English tenses (16 in total).
  2. Non-Linearity: Narratives frequently employ flashbacks and "temporal focus" shifts that disrupt simple linear progression.
  3. Vagueness: Stories often describe events as happening "during" others without providing precise timestamps.

A Three-Tiered Methodology

The authors address these challenges through a tiered linguistic framework:

Level 1: Aspect

Verbs are classified into "situations"—States, Processes, or Transition Events—to determine their inherent temporal properties.

Level 2: Tense

The model expands upon Reichenbach’s classic triad (Speech Time, Event Time, and Reference Time) by introducing additional Reference Times ($RT'$, $RT''$). This expansion allows for the formal mapping of all 16 English tenses, including complex constructions like the "past future perfect."

Level 3: Discourse and Temporal Focus

The core of the system is a deterministic algorithm that manages Temporal Focus (TF) using a stack-based approach. As the narrative progresses, the algorithm evaluates "situational descriptions" (participants and locations). Significant shifts in context trigger a search of the stack to determine if the narrative has returned to a previous focus or initiated a new one.

To represent these relations, the authors utilize a tree-like structure based on Allen’s interval logic. In this model, "Includes" relations establish parent-child nodes, while "Precedes" relations define sibling nodes.

Figure 3: Mapping temporal intervals (i1-i5) to a relational graph structure.

The algorithm's ability to track the current focus point relative to event times and speech time is crucial for maintaining a consistent narrative state.

Figure 6: Representation of Temporal Focus (TF) and Event Times (ET) relative to Speech Time (ST).

Key Insights

  • Situational Anchors: The use of non-temporal data (who and where) serves as a powerful heuristic for identifying temporal shifts. A change in location is often a primary indicator of a new temporal context.
  • Structured Vagueness: By prioritizing high-level relations like "Includes" and "Precedes" over exact coordinates, the system maintains a robust model that can be incrementally refined as more information becomes available.
  • Deterministic Management: Treating temporal focus as a stack allows the system to handle complex nested narratives and flashbacks with computational efficiency and predictability.

Results and Impact

The paper successfully provided a formal logical mapping for all 16 English tenses, surpassing the limitations of contemporary models. Furthermore, the deterministic nature of the algorithm made it highly suitable for software implementation, moving away from the parallel heuristics common in earlier research.

While the model is optimized for narrative, the authors noted that further work is needed to handle goal-oriented discourse (like instructions) and to integrate specific durations (e.g., "for three hours"). Additionally, the system assumes that anaphoric references (e.g., resolving "he" to "John") are handled externally, though these tasks are often interdependent.

Summary updated

BibTeX
@inproceedings{song1988interpretation,
  author = {Fei Song and Robin Cohen},
  title = {{The Interpretation of Temporal Relations in Narrative}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure