Intention Cooperation Discourse and Dialogue CS 359

Intention & Cooperation Discourse and Dialogue CS 359 October 18, 2001

Questions • Do systems do conversational implicature? • What sorts of knowledge representations are used in dialogue systems? • Are there systems that incorporate planning, dialogue act recognition? • Has anyone tried applying these techniques to other conversational styles - e-mail, IM? • 60 -80% accurate subsystems, how bad is the whole thing?

Dialogue Management: Reading & Reporting • Dialogue Management overview – Spoken Dialogue Technology, Mc. Tear • State based systems – Design issues (Mc. Tear) – Automatic learning (Woszczyna & Waibel) • Frame-based systems: Sun. DIAL (Peckhem at al) • Plan based systems – Theorem proving (Smith et al) – TRAINS (BDI): Allen et al – Rational Agency “Artimis” - Sato & de Mori

Roadmap • Structure of Discourse (G&S 1986) – Attention to Intention • Planning and Cooperation – Cooperative meaning: Grice’s Maxims – Cooperative action: • STRIPS planning basics • Shared plans • Discourse and Domain plans

Discourse Structure • Attentional Structure: – Focus – Reference – Information Structure: Given/New • Intentional Structure: – Discourse purpose (DP) • Discourse segment purpose (DSP) • Contribute to overall goal of conversation • Linguistic structure organizes/executes

Cooperation in Communication • Conversational participants act together – Speakers must provide sufficient information about beliefs and intentions for hearers to interpret as part of plan – Hearers must recognize cues in language and structure of discourse to constrain inference of plan

Cooperative Meaning • Cooperative Principle: Grice 1975 – Make conversational contribution as required at stage of discourse by accepted discourse goal • Maxims: – Quantity – Quality – Relation – Manner

Maxim of Quantity • Be as informative as necessary – Be no more informative than necessary • E. g. ”I saw three ducks”-> – “I saw EXACTLY three ducks”

Maxim of Quality • Do not say that which you believe to be false – Don’t say things without evidence • E. g. “I saw ducks” implicates that you really did

Maxim of Relation • Be relevant – Utterances should relate to each other and overall discourse goals – Focus, coherence, reference all rely on relation • E. g. A: I am out of gas. • B: There’s a garage around the corner.

Maxim of Manner • “Be perspicuous” – Avoid ambiguity – Avoid obscurity – Be brief – Be orderly

Maxims and Meaning • In cooperative discourse, expect maxims will be followed. • However, – Violate or “opt out” – One or another may be violated in case of clash – Flout: Deliberately, blatantly break • “Exploit” maxim to create conversational implicature – Meaning outside of literal sense • E. g. irony, metaphor, hyperbole, etc. . .

Planning & Plan Recognition • Discourse planning based on classic AI – STRIPS (Nilson et al) • Plan: Sequence of actions from start to goal • Action model: “Operator” – “IF”: precondition for action – “ADD”, ”DELETE”: effect on state of action – “BODY”: subactions • Recognition: Links beliefs & desires to preconditions and goals

Plan Example

Planning Issues • Complexity – Forward-chaining: simple, but exponential – Backward-chaining: Can reduce search • Assumes single actor, single plan – Full control- ‘master-slave’ • Need notions of generation, enablement, simultaneous action, maintenance

Shared Plans • Tie belief and intention to plans (Pollack 86) – Beliefs about: relations among actions (enablement, generation) and executability – Intentions (of agent) about actions • • Multiple collaborative agents Not just simultaneous private plans Belief => Mutual belief Different agents, different actions

Collaborative Plans

Plan => Shared Plan • Simple. Plan(G, an, [a 1. . an-1], t 1, t 2) • Shared. Plan(G 1, G 2, A) • MB(G 1, G 2, EXEC(aj, Gaj) • BEL(G, EXEC(ai, G, t 2), t 1) • & • & • • • BEL(G, GEN(ai, ai+1, G, t 2), t 1) INT(G, ai, t 2, t 1) INT(G, BY(ai, ai+1), t 2, t 1) BEL: Believe INT: Intend EXEC: Execute GEN: Generate BY: By • MB(…) • MB(G 1, G 2, INT(Gaj, aj)) • MB(G 1, G 2, INT(Gaj, BY(a j, A)) • INT(Gaj, aj) • INT(Gaj, BY(aj, A))

Cooperative Plan Maxims • Conversational Default Rule 1 (CDR 1) • MB(G 1, G 2, Desire(G 1, P) & – Cooerative(G 1, G 2, P) & – Communicating (G 1, G 2, Desire(G 1, P) • MB(G 1, G 2, Desire(G 1, • Achieve(Shared. Plan(G 1, G 2, Achieve(P)))) •

Cooperative Plan Maxims • • Conversational Default Rule 2 (CDR 2) Shared. Plan* = Partial Shared Plan [Shared. Plan*(G 1, G 2, Achieve(P)) & MB(Desire(G 1, Do(G 2, Action)) & MB(G 1, G 2, Exec(G 2, Action) & MB(G 1, G 2, Contribute(Action, Achieve(P)))] Intend(G 2, Action) & MB(G 1, G 2, Intend(G 2, Action)|)

Action Schemas • Simultaneous action • Conjoined actions • Sequential actions • Single Actor plans

Shared Plan Summary • Intentional structure – Intentions • Relations: Dominance, Satisfaction-precedence – Discourse segments correspond to intentions • Plans in collaborative, task-oriented discourse – Not fixed, negotiated – Intended to be recognized – Propose plan; accept/deny; refine beliefs, intentions, plans, . ,

Limitations of Shared Plans • Only handles domain planning – No treatment of discourse plans • Turn-taking, clarification, openings… • Only addresses intentional structure – Doesn’t integrate attentional structure • Information flow, focus, reference

Proposal: Unified Framework • Integrate disparate components of discourse theory – Semantics: accessible referents – Attentional state – Intentional structure • Common structures form”threads”, ”scripts” – Speech acts - functional, informational – Dialogue acts

DRT-Style Combined Structure A: There is an engine at Avon B: It is hooked to a boxcar. ce 1 ce 2 s s’ s’’ x w e ce 1: asrt(A, B, engine(x) Avon(w) e: at(x, w) y u e’ ce 2: asrt(B, A, boxcar(y) e’: hook(y, u) u is x (s) (s’) (s’’))

Conversation Acts

Mental States & Dialogue Acts • Incorporate mental states in “s” of structure – Encode belief, attention, obligation. . – Belief = MB – Situation S’ inherits all of S • Dialogue Acts – Statement (assert), Open-option, Offer, Commit – Acts bring about effects • Me; ntal states, event types • E. g. Commit -> Obliged

Threads • Intentional organization: – Events grouped into “threads” • Threads “dominate” events • Events are ordered • Identify specific thread types – Argumentation acts= rhetorical relations: RST • Elaboration, etc – Predictable: activities - “scripts”, • Known conversational styles • Provide expectations, predict subsequent moves

Summary • Discourse as collaboration – Gricean conversational maxims • Cooperative principle – Cooperative task-oriented plans • “Shared. Plan’ • Use mutual belief, negotiation of plan, act timing • Integrated discourse model – Combine semantics, attentional, intentional state, conversational act strategy

Challenges • Conversational act recognition – “Okay’ • Domain plan recognition – Collection

Conversation Acts Extend speech acts for conversational control