Upload
lambert-york
View
216
Download
1
Embed Size (px)
Citation preview
Modeling Speech Acts and Joint Intentions in Modal Markov Logic
Henry Kautz
University of Washington
Goal Unified way to specify and reason about
Communicative actions Domain specific actions Joint and individual obligations Beliefs of agents about other agents
Criteria Handle uncertain and incomplete knowledge Support well-founded and efficient inference Support learning
Markov Logic Language for statistical-relational learning
Developed by Pedro Domingos [2004+] Clausal (CNF) syntax
Clauses may be hard or soft Weights of soft clauses are learned from examples
Semantics: compilation to a Markov model
Example
Advantages of Markov Logic
Expressive power of (finite domain) first-order logic Ontologies: project_review(x) => meeting(x) Relations: manages(Bill,CALO) Rules:
manages(x,y) & DARPA_project(y) => has_headache(x) Dynamic worlds: at(A,L1,i) & go(L1,L2,i,j) => at(A,L2,j)
Supports both weight and structure learning Very efficient local-search algorithms for computing
most likely assignment (MPE) Language of CALO Probabilistic Consistency
Engine (Uribe & Dietterich)
What’s Missing?
Consider representing the felicity conditions for the speech act Ask_If(S,H,P): Preconditions:
Speaker does not know whether P holds Speaker wants to know whether P holds Speaker believes Hearer knows whether P holds
Effects Hearer believes Speaker wants to know whether P holds
Modal Logic
Logics for representing attitudes such as Knows, Believes, Wants, Ought, …
Traditionally formalized by rules & axiom schemas, e.g.: If p can be deduced, then Bp (necessitation) B(p => q) => (Bp => Bq) (distribution) Bp => BBp (introspection) …
Issues in Adding Modalities to Markov Logic
ML is not a deductive system: consequences follow from probabilistic semantics There cannot be an explicit rule of necessitation;
instead, must follow from probabilistic semantics ML only defined for finite structures
Distribution (and other axiom schemas) must not require infinite instantiations
Modal Markov Logic Ba
Ba P means agent a believes P Need not be certain belief Intuitively: the agent’s belief is actionable
Syntax KB = conjunction of weighted clauses Clause = disjunction of literals Literal = Atom or ~Atom Atom = Proposition or Ba(Clause) Extend to quantification over sets of constant
terms
Inference
Given a KB and a query, construct a Markov graph with Nodes for each (ground) atom and its negation Weighted hyperedges for each top-level clause Unweighted (strict) hyperedges connecting each
modal atom to the atoms for its disjuncts, and to the negations of its disjuncts Enforce consistency Enforce distribution
Example
B(p v q)
B p B q B ~qB ~p
B~p & B(p v q) => Bq
~B~p v ~Bp ~B~q v ~Bq
B~q & B(p v q) => Bp
Uses of soft rules: speech acts
Practically all preconditions and effects of communication acts are non-categorical E.g.: you may ask a question whose answer you
already know the answer Exceptions (and exceptions to exceptions…) need
not be explicitly written into each rule Higher-weighted rules can over-rule lower weighted
rules Can learn weights (& rules!) corresponding to different
styles of discourse
Uses of soft rules: joint obligations
Let M(a,b,g,i) = at time i, agent a is obliged to agent b to perform g
Simple soft persistence axiom: M(a,b,g,i) => does(a,g) v M(a,b,g,i+1) A purely logical persistence rule for obligations
would be extremely complex Such complexities (what if b dies? what is g
becomes impossible? etc) can be added as needed as additional soft rules
Uses of soft rules: plan recognition & cooperative behavior
Let W(a,p) = agent a wants p Cooperative agents
Try to recognize the goals of other agents
W(a,p) & enables(p,q) => W(a,q) Adopt those goals as their own (under proper
circumstances)
B(a,W(b,g)) & cooperative(a,b) => W(a,g)
Status 2nd generation (non-modal) UW Markov Logic
engine has been released Working on proofs of soundness & completeness of
modal extension Next steps
Implement Markov graph instantiation routines for modalities
Hand-code speech act, obligation persistence, and (simple) plan-recognition rules
Create or find annotated discourse transcripts and use to train weights Extend SRI/ICSI annotated corpus to include annotations
about agents’ mental state, as well as dialogs acts