1. MITRE Dialog Management Workshop – a review Dan Bohus Dialogs on Dialogs reading group CMU, November 2003

2. MITRE Dialog Management Workshop The Workshop  MITRE Dialog Workshop  @ MITRE, Bedford/Boston  October 27-28, 2003  Idea  Bring together researchers working on dialog management  Give them a homework Adapt you dialog manager to a medical diagnosis domain (details in a sec)  Discuss, compare, learn workshop : godis : ravenclaw : collagen : themes

3. MITRE Dialog Management Workshop The Homework  Implement a dialog system for the medical diagnosis domain  Task left open-ended (diagnosis, tutoring, etc)  No speech, just text in and out  Backend provided backend.docbackend.doc Java version and web-based interface version 3 diseases: malaria, coccidioidomycosis, another one List of symptoms: headache, nausea, muscle pain, etc. Decision tree involving symptoms and tests (fever, blood tests, travel patterns, etc)  Small enough to presumably not be lots of work, but large enough to allow illustration of functionalities, and provide some skeleton to the discussions… workshop : godis : ravenclaw : collagen : themes

4. MITRE Dialog Management Workshop Participants  MITRE (Carl Burke et al) MiDiKi  Gothenburg (Staffan Larsson) GoDiS (TRINDIKit)  USC ICT (David Traum) ICT Dialogue Manager  NTT/CMU (Matthias Denecke) Ariadne  CMU (Dan, Alex) RavenClaw  Ames (Beth-Ann Hockey) NASA Dialogue Manager  DFKI (Norbert Reithinger) DFKI Dialogue Manager  MERL (Candy Sidner, Charles Rich) COLLAGEN … and others invited but not present workshop : godis : ravenclaw : collagen : themes

5. MITRE Dialog Management Workshop GoDiS workshop : godis : ravenclaw : collagen : themes

6. MITRE Dialog Management Workshop GoDiS  TRINDIKit – information state update dialogue management toolkit  Information state Private: dialog plan, beliefs, agenda (short term goals) Shared: established facts, QUD, last utterance information  Dialog moves  Update rules  GoDiS: dialog management system implemented in TRINDIKit, handing:  information oriented dialogue  action oriented dialogue workshop : godis : ravenclaw : collagen : themes

7. MITRE Dialog Management Workshop TRINDIKit / GoDiS architecture input inter- pret TIS DEVICES LEXICON DOMAIN backend interface control update select gene- rate output lexicon domain knowledge DME Dialog plans Ontology Connectio n to Java Backend workshop : godis : ravenclaw : collagen : themes

8. MITRE Dialog Management Workshop GoDiS: Task Representation  Plans; propositional logic  Dialogue plans for dealing with diagnosis (issues opened at dialogue start)  ?x.disease(x): ”which disease is diagnosed?”  ?confirmed_by_interview: ”Is the diagnosis confirmed by additional information?”  ?confirmed_by_tests: ”Is the diagnosis confirmed by medical tests?”  Additional plans  ?x.info(x): ”What information is there about a given disease?”  ?x.treatment(x): ”What treatment is there for a given disease?” workshop : godis : ravenclaw : collagen : themes

9. MITRE Dialog Management Workshop GoDiS: Alternate Tasks  User-driven dialogue (implemented)  Not load issues when resetting; user has to raise all issues  User can ask system to Provide a diagnosis Confirm whether user has given disease  Decision trees as dialogue plans  Move backend knowledge into dialogue plans  Information conversion could be done automatically  Separate genre: expert system dialogue  Add special purpose update rules  Dynamic dialogue planning by expert workshop : godis : ravenclaw : collagen : themes

10. MITRE Dialog Management Workshop GoDiS: Highlights / Lowlights  Highlights:  Reuse, you get for free: Grounding Accomodation / plan recognition Multiple simultaneous issues & info sharing  High-level abstraction for dialog plans Rapid prototyping  Lowlights  Not used in this type of domain so far, so not entirely straight-forward (update rule changes)  Dynamic dialog plans (backend decides) workshop : godis : ravenclaw : collagen : themes

11. MITRE Dialog Management Workshop GoDiS RavenClaw workshop : godis : ravenclaw : collagen : themes

12. MITRE Dialog Management Workshop RavenClaw  Captures all domain-specific dialog (task) logic with a hierarchical description  The authoring effort is focused entirely here Dialog Task (Specification) Domain-independent Dialog Engine  Manages dialog by executing the dialog task specification  Provides domain-independent conversational strategies workshop : godis : ravenclaw : collagen : themes

13. MITRE Dialog Management Workshop RavenClaw Architecture Dialog Stack Madeleine E:LoadSymptomsGeneralFeel R:HowAreYou?I:GladI:Sorry Diagnose FeverTravel R:AskFeverE:MeasureTempI:InformFever I:Welcome Expectation Agenda general_feeling chart have_fever diagnostic workshop : godis : ravenclaw : collagen : themes

14. MITRE Dialog Management Workshop RavenClaw Architecture Dialog Stack Madeleine E:LoadSymptomsGeneralFeel R:HowAreYou?I:GladI:Sorry Diagnose FeverTravel R:AskFeverE:MeasureTempI:InformFever I:Welcome Expectation Agenda general_feeling chart have_fever diagnostic workshop : godis : ravenclaw : collagen : themes

15. MITRE Dialog Management Workshop RavenClaw Architecture Dialog Stack Madeleine Welcome Madeleine E:LoadSymptomsGeneralFeel R:HowAreYou?I:GladI:Sorry Diagnose FeverTravel R:AskFeverE:MeasureTempI:InformFever I:Welcome Expectation Agenda general_feeling chart have_fever diagnostic workshop : godis : ravenclaw : collagen : themes

16. MITRE Dialog Management Workshop RavenClaw Architecture Dialog Stack Madeleine Hi, this is Madeleine, the automated… Madeleine E:LoadSymptomsGeneralFeel R:HowAreYou?I:GladI:Sorry Diagnose FeverTravel R:AskFeverE:MeasureTempI:InformFever I:Welcome Expectation Agenda general_feeling chart have_fever diagnostic workshop : godis : ravenclaw : collagen : themes

17. MITRE Dialog Management Workshop RavenClaw Architecture Dialog Stack Madeleine Hi, this is Madeleine, the automated… Madeleine E:LoadSymptomsGeneralFeel R:HowAreYou?I:GladI:Sorry Diagnose FeverTravel R:AskFeverE:MeasureTempI:InformFever I:Welcome LoadSymptoms R:HeadacheR: Expectation Agenda general_feeling chart have_fever diagnostic headache workshop : godis : ravenclaw : collagen : themes

18. MITRE Dialog Management Workshop RavenClaw Architecture Dialog Stack Madeleine Hi, this is Madeleine, the automated… Madeleine E:LoadSymptomsGeneralFeel R:HowAreYou?I:GladI:Sorry Diagnose FeverTravel R:AskFeverE:MeasureTempI:InformFever I:Welcome R:HeadacheR: Expectation Agenda general_feeling chart have_fever diagnostic headache workshop : godis : ravenclaw : collagen : themes

19. MITRE Dialog Management Workshop RavenClaw Architecture Dialog Stack Madeleine Hi, this is Madeleine, the automated… Madeleine E:LoadSymptomsGeneralFeel R:HowAreYou?I:GladI:Sorry Diagnose FeverTravel R:AskFeverE:MeasureTempI:InformFever I:Welcome R:HeadacheR: GeneralFeel Expectation Agenda general_feeling chart have_fever diagnostic headache workshop : godis : ravenclaw : collagen : themes

20. MITRE Dialog Management Workshop RavenClaw Architecture Dialog Stack Madeleine Hi, this is Madeleine, the automated… Madeleine E:LoadSymptomsGeneralFeel R:HowAreYou?I:GladI:Sorry Diagnose FeverTravel R:AskFeverE:MeasureTempI:InformFever I:Welcome R:HeadacheR: GeneralFeel How are you feeling today? general_feeling chart have_fever diagnostic HowAreYou Expectation Agenda general_feeling: [good], [bad], [soso] GeneralFeel I:GladI:Sorry Not so good, I think I have a fever [soso](not so good) [fever](I think I have a fever) headache GeneralFeel workshop : godis : ravenclaw : collagen : themes

21. MITRE Dialog Management Workshop Illustrated Features  Dynamic generation of dialog task structure  Symptoms loaded from backend, appropriate structures to “talk about them” created on-the-fly  New symptoms – no DM changes  Dynamic dialog control policy  The order in which symptoms are addressed is controlled by the backend  Conversational skills workshop : godis : ravenclaw : collagen : themes

22. MITRE Dialog Management Workshop Illustrated Features  Dynamic generation of dialog task structure  Symptoms loaded from backend, appropriate structures to “talk about them” created on-the-fly  New symptoms – no DM changes  Dynamic dialog control policy  The order in which symptoms are addressed is controlled by the backend  Conversational skills workshop : godis : ravenclaw : collagen : themes

23. MITRE Dialog Management Workshop Dynamic Dialog Control … Dialog Stack Madeleine Hi, this is Madeleine, the automated… How are you today? Not so good, I think I have a headache Sorry to hear you’re not feeling so good, Tell me more about your symptoms… Do you have abdominal pain? Madeleine E:LoadSymptomsGeneralFeel R:HowAreYou?I:GladI:Sorry Diagnose FeverTravel R:AskFeverE:MeasureTempI:InformFever I:Welcome R:HeadacheR: Diagnose Expectation Agenda general_feeling chart have_fever diagnostic headache Backend Decision Tree workshop : godis : ravenclaw : collagen : themes

24. MITRE Dialog Management Workshop Illustrated Features  Dynamic generation of dialog task structure  Symptoms loaded from backend, appropriate structures to “talk about them” created on-the-fly  New symptoms – no DM changes  Dynamic dialog control policy  The order in which symptoms are addressed is controlled by the backend  Conversational skills workshop : godis : ravenclaw : collagen : themes

25. MITRE Dialog Management Workshop Conversational Skills  Corresponding agencies added automatically to the dialog task tree  Help  What Can I Say?  Repeat  Suspend / Resume  Start Over  Timeout handling (not illustrated)  Still need all the language generation prompts and grammar, but some of those are develop-once, too workshop : godis : ravenclaw : collagen : themes

26. MITRE Dialog Management Workshop RavenClaw Conclusion  Highlights  Set task posed no challenges to the framework Easy to implement  Dynamic dialog structure and control  Automatic use of domain-independent conversational skills  Lowlights?  Toolkit perspective: how easy would it be for someone else to build it?  Asynchronous behaviors? (timing)  Couple of bugs / fixes (or is that a highlight?) workshop : godis : ravenclaw : collagen : themes

27. MITRE Dialog Management Workshop GoDiS Collagen workshop : godis : ravenclaw : collagen : themes

28. MITRE Dialog Management Workshop COLLAGEN Collaborative Interface Agent communicate interact observe plan tree focus stack * Collagen workshop : godis : ravenclaw : collagen : themes

29. MITRE Dialog Management Workshop COLLAGEN Systems  air travel planning  email reading and responding (w. IBM/Lotus)  GUI design tool operation  car navigation system operation  airport landing path planning (w. MITRE)  gas turbine operator training (w. USC/ISI)  personal video recorder operation  programmable thermostat operation (with Delft U.)  multi-modal web-based form-filling workshop : godis : ravenclaw : collagen : themes

30. MITRE Dialog Management Workshop Collagen: Theory and Implementation Intentional purposes, contributes Linguistic segments, lexical items Attentional focus spaces, focus stack SharedPlan Discourse Theory (Grosz, Sidner, Kraus, Lochbaum 1974-1998) Java Implementation focus stack purpose tree workshop : godis : ravenclaw : collagen : themes

31. MITRE Dialog Management Workshop Collagen: Discourse Segments and Purposes (Grosz, 1974) E: Replace the pump and belt please. A: Ok, I found a belt in the back. A: Is that where it should be? A: [removes belt] A: It’s done. E: Now remove the pump. … E: First you have to remove the flywheel. … E: Now take the pump off the base plate. A: Already did. replace belt replace pump replace pump and belt (fixing an air compressor, E = expert, A = apprentice) workshop : godis : ravenclaw : collagen : themes

32. MITRE Dialog Management Workshop Discourse state representation E: Replace the pump and belt please. A: Ok, I found a belt in the back. A: Is that where it should be? A: [removes belt] A: It’s done Focus Stack replace belt replace pump and belt Purpose Tree replace pump and belt replace pumpreplace belt current focus space (Grosz & Sidner, 1986) replace belt replace pump and belt workshop : godis : ravenclaw : collagen : themes

33. MITRE Dialog Management Workshop Discourse interpretation algorithm (Lochbaum, 1998) starts a new segment/focus space (push) ends the current segment/focus space (pop) continues (contributes to) the current segment/... (add) The current (communication or manipulation) act either: focus stack directly achieves the purpose is a step in the plan for the purpose * identifies the recipe used to achieve the purpose identifies who should perform the purpose or a step in the plan identifies a parameter of the purpose or a step in the plan An act contributes to the purpose of a segment if it: purpose tree * does not include recursive plan recognition (see later topic) workshop : godis : ravenclaw : collagen : themes

34. MITRE Dialog Management Workshop COLLAGEN … my take  Separation of task from dialog/discourse engine  Recipes / Domain plans / Task tree  Full-blown HTN Hierarchical Preconditions (constraints) Effects Completion / failure Live nodes  Stack to keep track of focus and discourse structure  Tree explicitly contains agent and user nodes  Formalized / descriptive recipe specs (actually Java underneath), with procedure overwrites… workshop : godis : ravenclaw : collagen : themes

35. MITRE Dialog Management Workshop GoDiS Themes … workshop : godis : ravenclaw : collagen : themes

36. MITRE Dialog Management Workshop Themes: Task Representation  Task representation  Separation of task representation from dialog engine  High-level representations of task  Descriptive rather than procedural Procedural will be unavoidable for complex tasks Expressive power  GoDiS, RavenClaw, Collagen: plan based representations of task workshop : godis : ravenclaw : collagen : themes

37. MITRE Dialog Management Workshop Themes: Task/Domain/Gendre  The notion of dialog gendre  Tutoring  Diagnosis  Information Access  Where to fold it in a dialog manager?  GoDiS: update/select rules  Ariadne: plugins  RavenClaw: collapsed with task  How clear is that separation: task vs. gendre? workshop : godis : ravenclaw : collagen : themes

38. MITRE Dialog Management Workshop Themes: Development time  Systems took on the order of 3-5 days to develop  Significant effort in the backend connection Some sites shortcut it  Significant effort in grammar/language generation development Some sites shortcut it  Everyone that had an implementation: “fixed a couple of bugs, but no major changes required” workshop : godis : ravenclaw : collagen : themes

39. MITRE Dialog Management Workshop Themes: Development tools  Regression testing (GoDiS)  Systems are complex. Change something in a dialog management framework, can you prove that it did not screw up things that used to work?  System-wise, very intractable  Component-wise, maybe: i.e. DM with DM inputs/outputs  System diagnosis / log visualization tools (Collagen) workshop : godis : ravenclaw : collagen : themes

40. MITRE Dialog Management Workshop Themes: Timing  (Micro)timing  unaddressed  Turn-taking models  in general, very rudimentary  Asynchronous behaviors  Could be accomplished, but no-one seemed to have it  Multi-party conversation  unaddressed workshop : godis : ravenclaw : collagen : themes

41. MITRE Dialog Management Workshop Themes: the important problems  Different people have different views of what those are:  Plan / Intention recognition  Reference resolution  Backup in complex systems  Tense problems  Negations  Grounding; error prevention / recovery workshop : godis : ravenclaw : collagen : themes

42. MITRE Dialog Management Workshop Themes: Reasoning  Dialog Managers vs Backends  Where to draw the line?  Who does the reasoning?  Can we avoid duplicating it?  How rich is the interaction between them? Dialog systems - use language to act in a domain, so they are generally strongly tied Basic set of conversational skills can be identified  Drawing that line is still an “art”, no general agreement or solutions exist workshop : godis : ravenclaw : collagen : themes

43. MITRE Dialog Management Workshop Themes: Science of Dialog?  How much science do we have?  Theory vs. experiment  Interesting Collagen / RavenClaw similarities  Representation or not?  GUI analogy  Do we have the checkboxes and radio-buttons? workshop : godis : ravenclaw : collagen : themes