TL;DR
This paper proposes a unified Bayesian network model for plan recognition that explains why we make different inferences in stories versus real life. By adjusting a single parameter representing the prior probability of association between entities, the model can capture both the "story-like" tendency to jump to conclusions and the "life-like" caution.
What Problem Does This Paper Address?
When we encounter information, our reasoning process differs depending on whether it's presented as a narrative (a story) or as a factual account of the real world. For instance, if a story mentions a rope and a character wanting to kill themselves, we're more likely to infer they'll hang themselves. In real life, however, we'd consider a wider range of possibilities and be less certain about the rope being the chosen method, assuming other options (like pills or jumping) are available. This difference in inference, particularly concerning how objects influence our understanding of plans, is the core problem the authors tackle. They aim to develop a single, consistent theory of plan recognition that can account for these distinct behaviors.
Methods
The authors propose a Bayesian network framework to model plan recognition. They start by formalizing plans as actions with substeps, represented by "frames" and "slots" in a first-order logic. This formalism is then translated into a probabilistic model using Bayesian networks. The key innovation lies in identifying a specific network parameter – the prior probability of two arbitrary entities being the same (denoted as E) – as the "knob" that controls the difference between story-like and life-like inferences.
They further refine this by introducing a "mention" theory. This theory posits that in stories, authors are more likely to mention objects that are relevant to the narrative. By incorporating a rule that increases the probability of an object being mentioned if it's used in a plan, they show how this "mention" factor can influence the "knob" parameter (E) in a way that naturally leads to story-like interpretations.
Datasets
Not discussed in the paper. The paper uses hypothetical examples ("Jack wanted to kill himself") to illustrate its theoretical points rather than relying on specific datasets.
Key Insights and Ideas
- Unified Theory: The paper argues for a single, underlying mechanism for plan recognition, rather than separate models for stories and real life.
- The "Knob" Parameter (E): The prior probability of association between entities (E) is identified as the crucial factor that can be adjusted to produce different inference styles. A high E leads to story-like inferences, while a low E leads to life-like inferences.
- Relevance in Stories: The authors propose that the difference arises from the assumption of "relevance" in stories. Authors select and present information that is pertinent to the narrative, which in turn influences the probabilistic relationships within the plan recognition model.
- "Mention" Theory: The idea that authors mention objects that are relevant to plans provides a mechanism for how the "relevance" assumption can be encoded into the probabilistic model, effectively "turning the knob" towards story-like reasoning.
Key Results
The paper presents a theoretical framework and demonstrates how adjusting a single parameter (E) in a Bayesian network can yield both story-like and life-like plan recognition behaviors.
- Story-like inference: Achieved when E is high, leading to stronger probabilistic links between objects and plans, and thus more confident inferences.
- Life-like inference: Achieved when E is low, reflecting a more cautious approach where the presence of an object doesn't automatically imply its use in a plan, especially if other options exist.
- The "mention" theory provides a plausible explanation for why E might be higher in stories, by linking the author's act of mentioning an object to its relevance in a plan.
Limitations and Future Work
- Theoretical Framework: The paper presents a theoretical model and uses hypothetical examples. The authors acknowledge that they have no clear idea how to collect the necessary statistics for the probabilities needed in their Bayesian networks.
- Simplification: The model is a simplified first-order theory of plans and objects. The "mention" rule is also a basic one, and the authors suggest that more sophisticated rules would likely be needed.
- Implementation: While an appendix mentions ongoing work on implementing a story understanding program, the paper focuses on the theoretical underpinnings. The practical tractability of computing probabilities in complex networks remains a challenge.
- Future Work: The authors suggest expanding the theory to cover inferences about states and events, not just objects. They also plan to continue developing their implementation and exploring more sophisticated rules for "mention" and relevance.
Summary updated