A Synthetic Approach to Temporal Information Processing

Mark D. GroverAAAI 1982

Abstract

This paper introduces a "synthetic" framework designed to enable computational systems to interpret the complexities of human temporal language. By integrating formal linguistics, specifically Montague Grammar, with computer science models such as branching time and four-valued logic, the system processes nuanced temporal statements like "John was leaving on Thursday yesterday." The approach bridges the gap between narrow functional programs and broad linguistic modeling, offering a computationally feasible method for handling deictics, aspect, and uncertainty in natural language.

Introduction: The Challenge of Temporal Language

In the early 1980s, AI research faced a dichotomy between task-specific functional programs and broad models of human communication. Temporal information presented a significant hurdle due to its inherent messiness. Standard models struggled with "deictics" (context-dependent words like now), "aspect" (the distinction between completed and ongoing actions), and "uncertainty" regarding expected events. This research proposes a middle ground: a general-purpose temporal understanding system that is both linguistically grounded and computationally practical.

System Architecture and Methodology

The proposed model is "synthetic," drawing from disparate fields to create a unified processing pipeline.

Architecture of the General Temporal Understanding System.

The system operates through several key components:

  • Linguistic Analysis: Utilizing David Dowty’s Montague Grammar, the system parses English sentences into logical forms.
  • Branching Time Theory: Rather than a linear timeline, the system employs a graph of "possible worlds" to represent "inertia futures"—outcomes expected to occur unless intervened upon.
  • Interval Semantics: Time is treated as intervals rather than discrete points, facilitating the handling of events with vague durations.
  • Four-Valued Belief Logic: To process queries, the system utilizes a logic extending beyond binary True/False to include:
  • T (True): Event occurred.
  • F (False): Event did not occur.
  • B (Both): Contradictory information exists across different possible worlds.
  • Z (Zero): No information is available.

The implementation consists of a 3,100-line PASCAL program that translates syntax trees into a "Procedural Form" for the Temporal Information Processor.

Key Insights: Modeling Human Time-Keeping

A central innovation is the use of a dynamic "now" pointer. As this pointer advances, the system's answers to questions can change, reflecting the non-monotonic nature of human time-keeping.

A World/Time graph illustrating branching histories and possible worlds.

The model handles the "progressive" tense (e.g., "was leaving") by generating alternate futures where the event would have concluded. To manage complexity, the system uses a directed world/time graph without cycles, merging identical histories to prevent an exponential explosion of possible worlds. Additionally, the author suggests "reference counts" for events to simulate human-like memory decay, where unreferenced hypothetical events eventually fade from the system.

Evaluation and Results

The system was evaluated using complex illustrative sentences, such as "John was leaving on Thursday yesterday," to test its ability to handle past/future relations and the futurate progressive tense.

In sample runs, the system successfully navigated belief values within its internal timeline. For instance, when asked "Does John sleep?", the system returned different values (e.g., B for conflicting vs. T for true) depending on the position of the "now" pointer relative to scripted events. The model successfully integrated six distinct temporal phenomena into a single, cohesive processing pipeline.

Limitations and Future Directions

While the synthetic approach proved effective, several challenges remain. The model is a "fundamental scheme" rather than a complete solution for all English temporal nuances. Complex adverbials like "since yesterday" require further formal definitions. Furthermore, scaling the system to a full-scale natural language interface would necessitate a significantly larger repository of "scripted" world knowledge to handle rapidly changing contexts and metaphorical language.

Summary updated

BibTeX
@inproceedings{grover1982synthetic,
  author = {Mark D. Grover},
  title = {{A Synthetic Approach to Temporal Information Processing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1982},
}
Figure