Abstract
Modern artificial intelligence has made massive strides in natural language processing, yet current dialogue systems remain largely uncooperative and rigid. This paper argues that the prevailing "intent + slots" (I+S) paradigm is fundamentally incapable of true collaboration. To build assistants that can navigate complex real-world scenarios, the research community must return to and modernize the plan-based models of the 1980s and 90s, pairing symbolic reasoning with modern neural semantic parsing.
The Problem with Intent + Slots
Most current Task-Oriented Dialogue Systems (TODS) operate on a "frame" or "slot-filling" basis. In this model, the system identifies a user's intent and then "badgers" the user until a predefined list of variables (e.g., date, time, location) is filled. Philip R. Cohen identifies three critical gaps that prevent these systems from being truly helpful:
- Restricted Meaning: I+S systems struggle with complex constraints. For example, a user requesting "the earliest time available" or "not before 7 PM" is providing logical variables and constraints rather than atomic values. Current systems often fail to map these into simple slots.
- Lack of Mental State Modeling: Current systems do not track why a user wants something, nor can they represent the knowledge states of others. They lack the ability to handle "quantifying-in"—reasoning about what another person knows (e.g., "John knows Mary’s phone number") without the system itself knowing that specific data point.
- Rigid Initiative: These systems typically follow a prescriptive script. If a user asks a counter-question or attempts to shift the goal mid-conversation, the system's state machine often breaks because it lacks a model of the user's underlying plan.
A Hybrid Plan-Based Methodology
Cohen advocates for a Plan-Based Model of Dialogue, where communication is treated as a sequence of "speech acts" intended to change the mental state of the listener. Rather than abandoning modern progress, he proposes a hybrid architecture:
- Neural Front-End: Neural network-based semantic parsers map messy natural language into formal Logical Forms (LFs). This can be achieved using an "overnight" approach—generating a grammar of LFs and using crowd-sourcing to find natural language paraphrases for training data.
- Symbolic Back-End: A Dialogue Manager (DM) uses multimodal logic to reason about beliefs, desires, and intentions. By utilizing specific logical constructs like
pgoal(persistent goals) andknowref(knowing the value of a variable), the system can "debug" a user's plan.
Key Insights: Collaboration as Plan Debugging
The paper posits that a true assistant does more than answer questions; it evaluates the feasibility of the user's objectives. If a user asks for a movie time at a theater that is sold out, a plan-based system recognizes the higher-level goal (seeing the movie) and proactively suggests an alternative theater or time.
This approach allows for Mixed-Initiative dialogue. When a user responds to a system question with another question (e.g., "What's the earliest time available?"), the plan-based system treats this as a necessary sub-dialogue to establish a constraint. This decoupling of semantic parsing (language) from dialogue policy (planning) allows for the creation of domain-independent dialogue managers that do not need to be rebuilt for every new application.
Limitations and Future Directions
While promising, returning to symbolic reasoning introduces challenges in scaling and uncertainty. Automated reasoning and plan recognition are computationally expensive and historically difficult to scale to open-ended domains. Cohen suggests two primary paths forward:
- Probabilistic Plan Recognition: Investigating a middle ground between the robustness of neural networks and the precision of symbolic logic to handle uncertainty.
- Automated Knowledge Acquisition: Using text mining and crowd-sourcing to help planners automatically acquire the "physics" of a domain (the rules of what actions lead to what results), reducing the need for manual engineering.
Conclusion
By looking backward at the foundational logic of dialogue, researchers can move forward toward agents that act as genuine collaborators. Moving beyond the slot-filling paradigm is essential for creating assistants capable of participating in the complex, multi-party, and goal-oriented conversations of daily life.
Summary updated