1. Dialog Models 11-716 September 18, 2003 Thomas Harris

2. What is a (dialog) model? A model is an abstraction of a thing, dimensionally reduced, while still informative of the thing with respect to a particular perspective. A dialog model is a process calculus of a dialog, dimensionally reduced, while still informative of the dialog with respect to usability.

3. Why model? Not a good question. We always abstract, hence we always model. Ask instead, “Why this model?” Grosz and Sidner ’86 – the deep end of linguistics. TRINDI ’00ish – a modern survey.

4. Attentions, Intentions, and the Structure of Discourse Barbara J. Grosz and Candance L. Sidner Computational Linguistics, vol. 12, num 3, July-September 1986 11-716 Ariadna Font Llitjos September 25, 2001

5. New Theory of discourse structure As opposed to meaning (needs to partially rest on the discourse structure) Stresses discourse purpose and processing 3 separate but interrelated components (needed to explain interruptions, referring expressions, etc.): –Linguistic structure (sequence of utterances) –Intentional structure –Attentional state

6. This distinction simplifies both the explanations given + computation mechanism used Speaker/hearer ICP/OCP

7. Linguistic structure Utterances in a discourse are naturally aggregated into discourse segments (like words into constituent phrases) Segments are not necessarily continuous (interruptions) LS is not strictly decompositional 2-way interaction between discourse segment structure and utterances constituting the discourse: –linguistic expressions can convey info about discourse structure (cue phrases, ling. boundary markers) –Discourse structure constraints the interpretation of these ling. expressions

8. Intentional Structure Discourse (participants) have an overall purpose Even though there might be more than one, G&S distinguish one as foundational to the discourse (vs. private purposes) which needs to be recognized Each discourse segment has a discourse segment purpose (DSP), which contributes to the overall DP

9. Intentional structure cntd. 2 structural relationships between DSP: –Dominance DSP1 contributes to DSP2 = DSP2 dominates (DOM) DSP1 –Satisfaction-precedence (Parsing: linear precedence) DSP1 satisfaction-precedes DSP2 when 1 must be satisfied before 2 The dominance relation invokes a partial ordering on DSPs, i.e. a dominance hierarchy Determinations (complete specification of what is intended by whom) vs. recognition

10. Attentional State As opposed to cognitive state, which is a richer structure that includes knowledge, beliefs, desires and intentions Abstraction of the participants’ focus of attention as their discourse unfolds (a property of the discourse itself) Dynamic: records the objects, properties and relations that are salient at each point in the discourse

11. Attentional State cntd. Modeled by a set of focus spaces which constitute the focusing structure A focus space = segment + DSP Although each focus space contains a DSP, the focus structure does not include the intentional structure as a whole The stacking of focus spaces reflects the salience of entities in each space during the corresponding segments of the discourse

12. Attentional State cntd. Focusing structure depends on the intentional structure: the relationships between DSPs determine pushes and pops from the stack Focusing structure coordinates the linguistic and intentional structures during processing (p. 181) Like the other 2 structure, focusing structure evolves as discourse proceeds

13. Discourse examples Essay (p. 183) Task-oriented dialog (p. 186) –Intentional structure is neither identical nor isomorphic to the general plan

14. Processing issues Intention recognition –What info can the OCP use to recognize an intention –At what point does this info become available Overall processing module has to be able to operate on partial information It must allow for incrementally constraining the range of possibilities on the basis of new info that becomes available as the segment progresses

15. Info constraining DSP: –Specific linguistic markers –Utterance-level intentions (Grice’s maxims) –General knowledge about actions and objects in the domain of discourse Applications of the theory: –Interruptions (weak vs. strong) (p. 192) –Cue words (p. 196)

16. Properties and problems of discourse-level intentions DP/DSP are natural extensions of Grice’s utterance-level meanings… but G&S don’t address meaning Remains to be seen whether x and f are equivalent to DS and their features (p. 199) G&S state that the modes of correlation that operate at the utterance-level (c) also function at the discourse level

17. Basic Generalization Discourse sufficiency: the intentional structure need not be complete Belief case Action case

18. Conclusions Theory presented by G&S is a generalization of theories of task-oriented dialogs, but it’s domain independent Interesting and thorough but infeasible

19. More conclusions Asks more questions than it answers. How do we implement these aspects of dialog? Basically correct.

20. TRINDI circa 1998-2000 European Community sponsored Göteborg, Edinburgh, Saarbrücken, SRI, Cambridge, Xerox Research Centre EuropeGöteborgEdinburghSaarbrückenSRI, CambridgeXerox Research Centre Europe Effort to experiment and evaluate different theoretical dialog models in a real system

21. Basic Toolkit Architecture Informational Components Formal Representations Dialog Moves Update Rules Control Strategy

22. Informational Components Data Participants Beliefs Common ground Intentions

23. Formal Representations Formal representation of informational components Typed feature structures Lists Sets Propositions First order logic

24. Dialog Moves Trigger the update of the information state Grammatical triggers External events

25. Update Rules Govern information state updates Sometimes incorporates domain knowledge Sometimes govern behavior of dialog moves

26. Control Strategy Decide which update rule applies Simple priority list Game theory Utility theory Statistical methods

27. Dialog Theories Finite State Dialog Models Plan-based Models

28. Finite State Dialog Models Information is a state in the FSM Dialog moves are inputs matching transitions Update Rules are FSM lookups and transitions Control Strategy is static, the FSM itself

29. Plan-based Models Information state is the modeled beliefs, desires, and intentions of the participants Dialog moves are speech acts, e.g. request and inform Update rules are cognitive rules of evidence Control Strategies are classic AI plan-based strategies

30. Systems Implemented GoDiS (Questions under Discussion, Ginzburg ’96) PTT and EDIS (DRT) MIDAS (DRT) SRI Autoroute (Game Theory)