PlanVerb: Domain-Independent Verbalization and Summary of Task Plans

Gerard Canal, Senka Krivic, Paul Luff, Andrew ColesAAAI 2022 · DOI

Introduction

Automated task planners serve as the decision-making engines for autonomous systems, yet their outputs—sequences of technical, parameterized actions—are often incomprehensible to non-expert users. PlanVerb addresses this gap by providing a domain-independent framework that translates machine-readable plans into clear, natural language. Unlike previous template-based approaches, PlanVerb focuses on the underlying causality of actions, explaining not just what a robot is doing, but why.

The Verbalization Space

To accommodate diverse user needs, PlanVerb defines a multi-dimensional "Verbalization Space." This allows the system to tune its output across four primary axes:

  1. Abstraction (A1-A4): Ranges from raw technical data (A1) to high-level conceptual summaries of "who and what" (A4).
  2. Locality: Determines whether the narration covers the entire plan or a specific subset of actions.
  3. Specificity: Offers choices between a "General Picture," a "Summary," or a "Detailed Narrative."
  4. Explanation (E1-E5): Controls the depth of causal reasoning, from simple action lists (E1) to complex "deferred justifications" (E5) that link early actions to final goals.
The impact of Abstraction (A1-A4) and Explanation (E1-E5) levels on the total word count of generated plans.

As shown in the figures, increasing the levels of explanation and specificity naturally leads to longer, more detailed outputs. The system manages this complexity through action compression, merging repetitive tasks (like multiple navigation steps) into concise phrases using prepositions like "via."

Word count variations across different levels of Specificity and Explanation depth.

Methodology: Tagging and Causality

PlanVerb achieves domain independence through Semantic Domain Tagging. Instead of rigid templates, domain experts add syntactic tags (verbs, subjects, prepositions) to planning files (PDDL or RDDL). The system then uses the mlconjug3 library to dynamically conjugate verbs based on the plan's temporal context, ensuring grammatical correctness across past, present, and future tenses.

Crucially, the system identifies "immediate justifications" and "goal-achieving explanations" by analyzing the plan's internal logic. This causal scripting ensures that users understand how a seemingly minor action, like picking up a tool, is a necessary precursor to a distant objective.

Evaluation and Key Results

The system was tested against standard AI planning benchmarks (IPC, IPPC) and custom robotics domains. A user study involving 42 participants revealed several key insights:

  • Comprehension Gains: Participants provided with summarized plans containing causal justifications achieved comprehension accuracy between 80% and 97.62%. In contrast, those viewing simple step-by-step lists scored significantly lower (50-60%).
  • User Trust: Causal explanations significantly boosted user confidence, with satisfaction scores rising from 4.11 to 5.67 out of 7 on a Likert scale.
  • Linguistic Quality: Despite the automated nature of the generation, users rated the output as highly readable and grammatically sound, with an average score of 4.18/5.

Conclusion and Future Directions

PlanVerb demonstrates that providing the "why" behind actions is more critical for human understanding than simply listing the "what." While the system currently requires manual tagging and can be repetitive with nouns, future iterations will incorporate pronominalization and support for continuous processes in PDDL+. By bridging the communication gap between machines and humans, PlanVerb paves the way for more transparent and trustworthy autonomous systems.

Summary updated

BibTeX
@inproceedings{canal2022planverb,
  author = {Gerard Canal and Senka Krivic and Paul Luff and Andrew Coles},
  title = {{PlanVerb: Domain-Independent Verbalization and Summary of Task Plans}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i9.21204},
}
Figure