Introduction
In the early 1980s, natural language interfaces were often brittle, treating each user query as an isolated event. However, human conversation is rarely a series of independent data requests; it is a goal-oriented process. This 1983 classic introduces TRACK, a system designed to follow the "moving target" of user intent during a conversation. By building a hierarchical model of a user's goals and plans, TRACK provides cooperative responses that go beyond literal answers, effectively anticipating what a user needs to achieve their objective.
The Challenge of Context in Information-Seeking
Sandra Carberry identifies a major gap in information-seeking environments: users often have high-level goals (e.g., graduating with a specific degree) but ask low-level questions (e.g., "What are the prerequisites for French 112?"). Without understanding the context—the plan the user is currently executing—a system cannot handle ill-formed input, detect confusion about requirements, or provide "cooperative responses" that address underlying needs. The challenge was to build a model that could hypothesize, track, and update these shifting task-level goals in real-time.
The TRACK System: Hierarchical Plan Modeling
TRACK utilizes a hierarchical structure of component goals and actions to maintain context.
Plan Representation
The system uses the STRIPS formalism to represent plans as sets of preconditions, partially ordered actions, and effects. It distinguishes between three levels of goals:
- Immediate Goal: The literal semantic interpretation of the utterance.
- Derived Goal: An inferred intent (e.g., asking about a prerequisite implies a goal of taking the course).
- Focused Plan: The specific plan the user is currently pursuing.
The Context Mechanism
The "brain" of TRACK is its context tree, which represents the user's global and local plans. As dialogue progresses, the system uses six heuristics (H1-H6) to decide how to attach new information to the tree. For example, it checks if a new query is a sub-action of the current focus or if the user has jumped to a different part of a higher-level plan.
Key Insights and Results
The paper demonstrates TRACK's logic through qualitative walkthroughs within a university knowledge base representing the requirements of the University of Delaware.
- Local vs. Global Context: By distinguishing between immediate tasks and higher-level goals, the system understands why a user asks a specific question.
- Pragmatic Overshoot: TRACK can provide more information than requested. If a user asks about a course that doesn't help satisfy their known language requirement, the system explains the mismatch rather than just answering the literal question.
- Handling Ambiguity: When a user asks "Can I obtain a Computer Science major?", TRACK identifies potential ambiguity by constructing multiple context trees. For instance, it might generate trees for both Bachelor of Arts (BA) and Bachelor of Science (BS) degrees.
Conclusion and Future Directions
While TRACK demonstrated significant progress in plan recognition, Carberry noted several limitations that remain relevant today. The 1983 system could not handle recursive plans, user preferences (e.g., avoiding 8:00 AM classes), or abrupt goal switching. Furthermore, it assumed the user's beliefs were correct, lacking a mechanism for robust belief revision. Despite these limitations, the general heuristics for plan attachment remain a foundational concept in goal-oriented dialogue systems.
Summary updated