1 / 46

Document Transformations and Information States

Document Transformations and Information States. Staffan Larsson, Annie Zaenen SIGdial workshop, ACL 2000, Hong Kong. Goals of this talk. sketch a model of the relation between monologue and dialogue in instructional discourse, based on the TRINDI information state approach

Download Presentation

Document Transformations and Information States

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Document Transformations and Information States Staffan Larsson, Annie Zaenen SIGdial workshop, ACL 2000, Hong Kong

  2. Goals of this talk • sketch a model of the relation between monologue and dialogue in instructional discourse, based on the TRINDI information state approach • investigate possibilities of using domain task plans to generate discourse with various degrees of interactivity • adapt an existing TrindiKit dialogue system, designed for information-seeking dialogue, to handle instructional dialogue and monologue • NB. We are not dealing with educational (a.k.a. instructional) dialogue

  3. Overview of talk • motivation • monologue and dialogue • the TRINDI information state approach • IMDiS (Instructional Monologue and Dialogue System) • from domain plan to monologue and dialogue plans • monologue and dialogue behaviour in IMDiS • conclusions & prospects

  4. Practical motivation for document transformation • Practical advantages of being able to adapt the “same” document to different uses and media • Costly to develop presentations from scratch for each new media

  5. Theoretical issues • Written discourse and human-human dialogue are the extremes of the scale of interactivity • We explore some intermediate levels of interactivity using the TRINDI information state approach • How can we model the differences between these levels, independently of domain-specific knowledge? • Concepts: • interactivity in monologue and dialogue • interactivity vs. initiative • spoken vs. written discourse • maxim of quantity in monologue and dialogue

  6. Degrees of interactivity spoken monologue • Dialogue is interactive, since participants can influence each other’s moves • Monologue is typically non-interactive • Written monologue is minimally interactive; reader can influence order & speed of reading human-human dialogue written monologue inter- activity none minimal high

  7. Interactivity vs. initiative • Interactivity does not imply shared initiative; it is possible to have a dialogue with both task and dialogue initiative staying on one side • In instructional dialogue, task initiative is typically held by instructor; dialogue initiative may shift

  8. Spoken vs. written instructional discourse • separate dimension from interactivity (and initiative) • written monologue has advantage of providing minimal interactivity • spoken monologue (e.g. tape recording) has advantage of leaving hands and eyes free to do other things • prerecorded repetitions and reformulations can be used as a substitute for interactivity

  9. Quantity of information in monologue • manuals are cooperative, and are based on implicit assumptions about the reader • setting: instructor has know-how information which instructee does not • all information, and preferably only that information, which is needed by the instructee to complete the task should be transferred (Grice’s maxim of quantity) • however, if education is an issue, additional information (explanations) should be supplied • in manuals, instructor must keep provide more information than may be necessary for any given instructee (cf. Young ’97, how to generate text from dynamically generated plans which are very detailed)

  10. Quantity of information in dialogue • interactivity makes it possible for instructee to control the amount of information given by the instructor • instructee must be able to inform instructor about her state, thus influencing the instructors moves • instructor needs to keep track of the state of the instructee to determine what information to provide

  11. From written manual to spoken dialogue – 3 levels of interactivity • 1. written/spoken monologue • no overt interaction (but minimal interaction in written m.) • 2. recapture minimal interactivity in spoken presentation • system can ask yes/no questions and deals with simple user feedback (“done”, “OK”, “what did you say?”) • 3. increase interactivity • user can control level of detail by indicating whether she already knows certain subprocedures • requires more detailed domain task plan

  12. Example from a Xerox manual • Reinstalling the print head • Caution: make sure that the green carriage lock lever is STILL moved all the way forward before you reinstall the print head • 1. Line up the hole in the print head with the green post on the printer carriage • Lower the print head down gently into position • 2. Gently push the green carriage lock lever up until it snaps into place • This secures the print head • 3. Close the tops cover and re-attach the scanner • 4. Press and release the yellow LED button • The printer will prepare the cartridge for printing • Note: if the carriage does not move from the center position after you press the carriage change button, remove and reinstall the print head

  13. The TRINDI information state approach • Information states represent information available to dialogue participants, at any given stage of the dialogue • Dialogue moves trigger information state updates, formalised as information state update rules • TrindiKit: software package for implementing dialogue systems; based on the information state approach to dialogue management • Explicit information state data-structure • makes systems more transparent • closer to dialogue processing theory • easier comparison of theories • modularity for simple and efficient reconfiguration and reusability

  14. Building a system domain-specific system domain knowledge (resources) dialogue theory (IS, rules, moves etc) domain-independent system TRINDIKIT software engineering (basic types, control flow)

  15. control DME input inter- pret gene- rate output update select Information State resource resource

  16. GoDiS - An experimental dialogue system built using the TrindiKit • Information-seeking dialogue • Information state based Ginzburg’s notion of Questions Under Discussion (QUD) • Dialogue plans to drive dialogue; plans are dynamically interpreted by information state update rules • Simpler than general reasoning and planning • More versatile than frame-filling and finite automata • Moves: ask, answer, greet, thank, requestRepeat, repeat

  17. GoDiS information state type AGENDA : stack( Action ) PLAN : stackset( Action ) PRIVATE : BEL : set( Prop ) TMP : (same type as SHARED) BEL : set( Prop ) SHARED : QUD : stack( Question ) LM: set( Move )

  18. from GoDiS to IMDiS information state SHARED.ACTIONS: stack of actions that the system has instructed the user to perform but that have not yet been confirmed as done

  19. IMDiS information state type AGENDA : stack( Action ) PLAN : stackset( Action ) PRIVATE : BEL : set( Prop ) TMP : (same type as SHARED) BEL : set( Prop ) ACTIONS : stack( Action ) SHARED : QUD : stack( Question ) LM: set( Move )

  20. IMDiS in monologue mode • 2 moves: Instruct, Inform • All actions use the update rule autoConfirm • rule autoConfirm val(SHARED.ACTIONS, A) pre: pop(SHARED.ACTIONS) add(SHARED.BEL, done(A)) eff:

  21. IMDiS in dialogue mode • 9 moves (6 GoDiS moves, the 2 above and Confirm) • Confirmations are integrated by assuming that the current top-most action of the shared.actions has been performed • rule integrateUsrConfirm (slightly simplified) val(SHARED.LM, confirm(usr) ) val(SHARED.ACTIONS, A) pre: pop(SHARED.ACTIONS) add(SHARED.BEL, done(A)) eff:

  22. Example from a Xerox manual (repeated) • Reinstalling the print head • Caution: make sure that the green carriage lock lever is STILL moved all the way forward before you reinstall the print head • 1. Line up the hole in the print head with the green post on the printer carriage • Lower the print head down gently into position • 2. Gently push the green carriage lock lever up until it snaps into place • This secures the print head • 3. Close the tops cover and re-attach the scanner • 4. Press and release the yellow LED button • The printer will prepare the cartridge for printing • Note: if the carriage does not move from the center position after you press the carriage change button, remove and reinstall the print head

  23. Simple task plan NAME: reinstall(print_head) PRECOND: moved_forward(carriage_lock) line_up(hole, post) lower(print_head) Action push(lever) close(top_cover) condition reattach(scanner) press_and_release(button) N Final state Y moved_from_center(carriage) remove(print_head) reinstall(print_head) EFFECT: reinstalled (print_head)

  24. From domain task plan to monologue and dialogue plans

  25. SampleIMDiS information state AGENDA = { findout(?moved(carriage)) } if_then( not moved( carriage ), [ remove(print_head), reinstall(print_head) ] ) PRIVATE = PLAN = BEL = { } TMP = (same structure as SHARED) moved_forward(carriage_lock) done(secure(print_head)) done(close(top_cover)) done(reattach(scanner)) COM = SHARED = ACTIONS= < press_and_release(yellow_button) > QUD = < > LM = { instruct(sys, press_and_release(yellow_button)}

  26. Monologue and dialogue behaviorin IMDiS • interactivity level 1 – (spoken) monologue • the text you have seen above but presented orally so the user can direct her attention to the task

  27. level 2 - dialogue behavior with minimal interactivity • yes/no questions • confirm (”ok”) • requestRepeat (”what?”) • grounding and avoidance of irrelevant information • example: • S: Press and release the yellow button • U: ok • S: Has the carriage moved from the center position? • U: yes • S: the print head is now installed • U: what? • S: the print head is now installed

  28. Complex task plan NAME: reinstall(print_head) PRECOND: moved_forward(carriage_lock) secure(print_head) Final state close(top_cover) reattach(scanner) Complex action press_and_release(button) N Y moved_from_center(print_head) remove(print_head) Action reinstall(print_head) Condition EFFECT: reinstalled (print_head)

  29. Subtask plan for complex action NAME: secure(print_head) PRECOND: - line_up (hole, post) lower(print_head) push(lever) EFFECT: secured (print_head)

  30. level 3: increased interactivity • skipping subtask instructions • S: Put the print head in place • U: Ok, done, what now? • S: Close the top cover • requesting subtask instructions; popping out of subdialogue • S: Put the print head in place • U: how? • S: Line up the hole in the print head with the green post on the printer carriage • U: done, I now remember the rest, the print head is secured

  31. Domain task plans – where do they come from? • What we did: read the text and construct a corresponding domain plan that seems plausible; this is not a principled approach • Alt.1: Annotating existing manuals with procedural relations (Scott et. al. 1995); use this to reconstruct domain plan • Alt.2: Author uses authoring tool to specify domain plan (“knowledge model”); this has been done in the context of multilingual manual authoring (Scott 1996, DRAFTER project, GIST project)

  32. Conclusions • We provide a simple schema modeling the relation between task plan and discourse plans for intermediate interactivity level • The schema also gives a model of how different levels of interactivity (from monologue to dialogue) are related • The degree of interactivity is limited by the level of detail with which the manual encodes the underlying task • Document transformation requires separation of task plans and assumptions about interactions • Task plans can be obtained either by marking up existing manuals, or by constructing them using authoring tools • Extended existing information-seeking dialogue system to handle instructional dialogue (reusability!); based on information state approach

  33. Prospect • Use existing domain task plans produced by annotating text or using authoring tools • TrindiKit dialogue system for information seeking dialogue, adapted to instructional dialogue; uses dialogue plans (NB: current system very simple) • Our model sketches a mapping from the domain plans to dialogue plans • So, building on existing components, this indicates a possible way to get instructional dialogue systems without a lot of extra work

  34. Possible future work • Explore and implement further levels of interactivity • Scale up experiment, e.g. a complete manual • Explore use of domain plans generated using existing tools • Extend IMDiS (and GoDiS) to handle referent resolution subdialogues, asynchronous turntaking etc. • Do real NL generation, possibly using existing components

  35. extra slides... • Relation to previous work • Carlson, Kuppevelt • M. Young • Smith & Hipp • Interactive tutoring (Lester, Rickel & Johnson) • N.B. We have not read all papers by everyone...

  36. Relation to previous work: Kuppevelt, Carlson • – • monologue as a special case of dialogue where all moves are answers to implicit questions resulting from the previous move(s) • example • The printer will prepare the cartridge for printing • Q: What if the carriage does not move ... • A: If the carriage doesn’t move, ... • but this is not a processing theory, and it is unclear how it could be implemented; the number of possible questions that could arise at any given moment is too big

  37. Relation to previous work:Young • Young (1997) deals with how to give the right amount of information in a (monologue) text, given a dynamically (automatically) generated task plan which may be very complex • Young deals only with monologue • We deal with both monologue and dialogue • Our plans are not automatically generated, and not as complex

  38. Relation to previous work: Smith & Hipp • Handles similar type of dialogue (problem solving) using “Missing Axiom Theory”, using general inference over FOL representation • Handles some phenomena which we don’t (e.g. referent disambiguation) • S&H do not generate monologue, or explore several levels of interactivity; however, they allow mixed initiative to a larger extent than we do • S&H assume complex representation of the domain task plan (?) • Our approach is less complex and therefore easier to implement and more computationally tractable • We provide an explicit model of the relation between monologue and simple dialogue

  39. Dialogue models and monologue • Unfortunately, dialogue models don’t look at monologues • Fortunately, dialogue models based on information states can be extended to account for monologues • Ways to adapt dialogue models to monologue • General theory of action and communication • Monologue as a degenerate case of dialogue

  40. Relation to previous work:interactive tutorials using text

  41. Similarities between monologue and dialogue • Hierarchically structured • Can be represented as a sequence of questions and answers: Carlson, Van Kuppevelt

  42. Importance of Information States • The main difference between monologue and dialogue lies in what each participant knows about what the others know • Instructional manuals as a genre make assumptions about information states (as do other text genres) • Between monologue and full-blown ‘natural’ dialogue one can devise various modes of interaction: everything is a question of style and genre

  43. theoretical aside: plan representation • traditionally, plans are represented as decomposition trees • we add conditionals to the plan representation format, adding expressivity • plans are interpreted dynamically by information state update rules • our domain task plans are not constructed dynamically by a planner; rather, we assume that they are constructed by a human (the author of the manual), possibly using some markup language • discourse plans are generated from task plans using a simple schema which encapsulates the different levels of interactivity

  44. Sample GoDiS information state(travel agency domain) AGENDA = { findout(?return) } findout(?x.month(x)) findout(?x.class(x)) respond(?x.price(x)) PRIVATE = PLAN = BEL = { } TMP = (same structure as SHARED) dest(paris) transport(plane) task(get_price_info) COM = SHARED = QUD = < x.origin(x) > LM = { ask(sys, x.origin(x))}

  45. What we have not done • NL generation – instead, we used canned text derived from manuals • referent disambiguation issues • more complex user utterances, e.g. alternative questions (though latest version of GoDiS handles this partially)

  46. Possible extensions • Larger scale experiments • More descriptive work, including on more fine- grained linguistic aspects • Implementation of more complex dialogue behaviours, e.g. referent resolution subdialogues, explanatory subdialogues • Explore higher levels of interactivity • Build tools to make the creation of multi-purpose documents possible • Mark-up schemes (XML)

More Related