Modeling Speech Acts and Joint Intentions in Markov Logic

1
Modeling Speech Acts and Joint Intentions in
Markov Logic

• Henry Kautz
• University of Washington

2
Goal

• Infer and track the goals, intentions, and
  beliefs of the participants in a meeting
• Sources of information
• Communicative actions
• Gestures
• Documents (agenda, minutes, emails)
• Commonsense background knowledge

3
Requirement

• Handle
• Uncertain observations
• Incomplete user models
• Non-categorical background knowledge
• Integrate with rest of CALO
• Build a working prototype
• In 12 months ?!
• Can leverage three key developments...

4
Key Pieces

• Joint intention theory
• Unified logical theory of belief, intention,
  communicative actions
• Result of 20 years of work by Phil Cohen, Ray
  Perrault, James Allen, and others
• U Texas CLib Knowledge Base
• Commonsense background knowledge
• Bruce Porter et al.

5
3. CALO Core Inference

• A general learning and probabilistic state
  estimation module
• Basis Markov Logic, a probabilistic
  generalization of clausal first-order logic
• Good support for importing rule-based background
  knowledge
• Good algorithms for learning MLE inference
• Tomas Uribe et al (SRI) main implementation
  support
• Other team members Pedro Domingos, Tom
  Dietterich, Stuart Russell, Alon Halevy, Henry
  Kautz

6
Work Plan

• Encode (simplified) version of Joint Intention
  Theory in Markov Logic
• Dynamic version ML already exists
• Adding modal operations underway
• Translate appropriate parts of CLib to ML
• Can be largely automated
• Define implement inputs from vision audio
• Use existing (eventually, new) ML engine for
  parameter learning (from annotated scenarios) and
  state tracking