AgentBased Dialogue Management Discourse Dialogue CMSC 35900 1

Agent-Based Dialogue Management Discourse & Dialogue CMSC 35900 -1 November 10, 2006

Possible Projects • Alternatives: – Implementation with short write-up – Analysis with longer write-up (max 15 p) • Topics: • Use toolkit to implement components of SLS • Implement/Extend a pronoun resolution algorithm • Identify phenomena not covered by theory & extend – E. g. Dialogue type functions in RST – Zero-anaphora in pronoun resolution • Analysis: – Compare reference behavior in chats/IM to other – Turn-taking in Call. Home vs Switchboard

Roadmap • Agent-based Dialogue Management – Dialogue by Theorem Proving – Dialogue by Planning – Dialogue by Conversational Agency – Event-Driven Dialogue – Dialogue by Rational Interaction • Advantages and Disadvantages

Agent-Based Dialogue • Contrasts with State- and Frame- Based – Structurally constrained – Simplified, activity based semantics • Agent-based dialogue management – Employs AI planning or reasoning processes – More full linguistic interpretation – More complex model of interaction, belief – Wide variety of models, reasoning

Dialogue by Theorem Proving • Exemplar: Circuit Fix-it-Shop (Smith) – Interaction between student and expert (S) – Debugging a fault in a circuit • Dynamically develops solution – Based on state and user knowledge • Theorem prover determines task completion – Dialogue used to acquire missing axioms

Theorem Proving in Dialogue • Example: system needs to know if there is a wire b/t connectors 84 and 99 – goal(c, learn, . . (prop(84, 99), exist, X), true))) – “missing” axiom about wire present yes/no – Use language to ask user, and ask to add – Answers, then asks for help • Adds new subgoal to teach user • Integrates problem-solving, theorem proving, and dialogue

Dialogue by Planning • Integrate utterances with plan actions – Interpret intention behind utt, how achieve goal – Utterance=speech act, same structure as plan op – Example: Convinceby. Inform – Constraints: Agent(Spkr), Agent(Hear), Prop(prop), Bel(Spkr, prop) – Preconditions: At(Spkr, Loc(hear)) – Effects: Bel(Hear, prop) • Buying ticket, requires agent using CBI to tell price

Dialogue by Planning • Issues: – Intention inference error-prone, cascading – Planning AI & NP-complete in general case • Strategies: – Differentiate domain plans and discourse plans • Train routing vs clarification subdialogues – Later approaches revise

Conversational Agency • Multi-agent planning system – Exemplar: TRAINS (Allen et al) • Collaborative, incremental: – System: Low-level route planning, track info – User: High-level goals, can relax constraints if needed – Mixed-initiative planning: • Plan generation under constraints; criticism & revision – System as full-fledged agent in conversation • Utts are speech acts, interp user, maintain models

Discourse Obligations • Prior work: “co-operativeness” – Interpret utt to determine speaker goals – System adopts goals as own • Trains: – Distinguish intentions and obligations of conv – Obligations represent what agent should do • E. g. if promise(A), generate obligation to achieve A – If S 1 requests A, S 2 should accept or reject – If utt not understood or wrong, should repair/correct, etc

Discourse Actor Algorithm While conversation not finished If system has obligations then address obligations else if system has turn then if system has intended conv. Acts then call NL generator else if some material is ungrounded then do grounding else if some proposal not accepted then consider proposals else if high-level goals unsatisfied then address goals else release turn to try to end conv else if no one has turn then take turn else if long pause then take turn • Act on any obligations before doing own goals • If user doesn’t take lead, system will take initiative

Event-driven Dialogue • Extension of frame-based approaches – Doesn’t model or recognize user intention – Models system state, intentions – Uses more complex representation of context to perform semantic interp of user utt • Modular approach: dialogue, task, belief – Based on current belief/task state, select dialogue options: e. g. confirmations, requests

Contextual Interpretation • Interpretation of utt generates contextual functions • • • E. g. new_for_system(X): confirm(X) or spec(X, Y) repeated_by_user(X): cancel “confirm(X)” Inferred_by_system(X): goal confirm(X) Modified_by_user(X): goal repair confirm(X) Negated_by_user(X): repair(X) • Combine goals with initiating actions • Include meta-strategy of increasing system control when need repair

Rational Interaction