Introduction
In the mid-1980s, temporal reasoning in AI, pioneered by researchers like James Allen, focused primarily on single intervals of time. While effective for simple event sequences, these systems struggled with "collections" of intervals—recurring patterns such as "every Thursday" or "the first Monday of the month." Representing complex expressions like "U.S. Election Day" (the first Tuesday after the first Monday in November) often required ad-hoc logical formulas that obscured the original semantic meaning. This paper proposes a formal, compositional representation for sets of intervals, enabling AI systems to handle periodic events and relative scheduling with symbolic efficiency.
The Core Framework: Slicing and Dicing
The proposed representation is built upon three fundamental concepts: Calendars, Dicing, and Slicing.
- Calendars: These serve as the primitives of the system. A calendar is defined as an infinite sequence of contiguous intervals (e.g., Days, Weeks, Months). Crucially, these are defined "intensionally" by rules rather than exhaustive lists, allowing the system to manage leap years and infinite time horizons without pre-calculating every date.
- Dicing Operators: These operators "divide" a larger interval into its constituent parts based on a calendar. For instance, dicing a Year by the Months calendar yields a collection of twelve intervals. The authors distinguish between strict dicing (using the
:operator), which clips intervals to fit boundaries, and relaxed dicing (using the.operator), which includes full intervals even if they only partially overlap the boundary. - Slicing Operators: These operators "select" specific intervals from a collection. Common slices include "1st," "last," or "any" (a random selection).
Compositionality and Natural Language Mapping
A key strength of this representation is its compositionality. Complex temporal phrases are constructed from simple parts using a right-to-left precedence that mirrors English prepositions. For example, the expression 1/Mondays :during: Months represents "the first Monday of every month."
The distinction between strict and relaxed operators allows for precise mapping of natural language nuances. As shown in the figure below, the choice of operator determines how boundaries are handled during the dicing and slicing process, ensuring that the symbolic logic matches human intuition.
Advanced Concepts: Depth and Symbolic Efficiency
The authors introduce the concept of "order" or "depth" in collections. An order-1 collection is a simple list of intervals (e.g., all Thursdays). An order-2 collection is a nested list of lists (e.g., a list of months, each containing a list of days). This nesting capability allows for the representation of highly sophisticated queries.
Furthermore, the system is symbolically efficient. Because expressions are stored as symbolic rules, the system does not need to instantiate a massive data structure for every second in a year. It only "unrolls" the specific parts of the calendar required to resolve a query or conflict, such as finding a free appointment slot using the any operator to search through a collection.
Conclusion and Future Directions
This formal language provides a principled way to bridge the gap between human temporal speech and machine logic. While the current model assumes "convex" (contiguous) intervals and precise time, it lays the groundwork for more complex scheduling systems. Future work may address "non-convex" intervals (unions of disjoint blocks) and "fuzzy" temporal expressions like "around lunchtime," which currently require pre-defined calendar mappings to be resolved.
Summary updated