1 / 35

Mixed-Initiative Dialogue Systems for Collaborative Problem-Solving

Mixed-Initiative Dialogue Systems for Collaborative Problem-Solving. George Ferguson & James Allen University of Rochester. Motivation. Eric Horvitz on “mixed-initiative” (2000):

cdan
Download Presentation

Mixed-Initiative Dialogue Systems for Collaborative Problem-Solving

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. Mixed-Initiative Dialogue Systems for Collaborative Problem-Solving George Ferguson & James Allen University of Rochester

  2. Motivation • Eric Horvitz on “mixed-initiative” (2000): • I shall use the phrase to refer broadly to methods that explicitly support an efficient, natural interleaving of contributions by users and automated services aimed at converging on solutions to problems. • Natural language dialogue systems for mixed-initiative problem solving • Efficient: easy to say complicated things • Natural: no training • Emphasizes role of the user

  3. Goals • Describe an architecture for mixed-initiative dialogue systems such that: • The system is able to do many things (in parallel) • e.g., search the web, plan activities, learn new tasks, observe its environment, ... • Dialogue-based interaction with a user is one of those things (albeit an important one) • Dialogue is in service of collaboration--we talk together in order to work together to solve problems • Collaboration is driven by principles of collaborative activity (joint intention) • Initiative (and mixed-initiative interaction) arises naturally from agents managing their joint intentions

  4. Outline • Architecture of Mixed-Initiative Dialogue Systems • Dialogue Systems • Agents • Collaborative Agents • Components of Collaboration • Ontology of collaborative problem solving acts • User- and system-initiative APIs • Extended Example • User initiative • Interpretation • System initiative • Collaborative behavior • Proposals • Agreement • Other Issues

  5. Outline • Architecture of Mixed-Initiative Dialogue Systems • Dialogue Systems • Agents • Collaborative Agents • Components of Collaboration • Ontology of collaborative problem solving acts • User- and system-initiative APIs • Extended Example • User initiative • Interpretation • System initiative • Collaborative behavior • Proposals • Agreement • Other Issues

  6. Dialogue Systems:The Standard Approach • A standard dialogue system is not an agent

  7. BDI Agents (briefly) Goal AssessmentGoal PrioritizationUtility Assessment... Hand-coded ProceduresReactive ControlLearned PolicyAdaptive PolicyMeans-Ends PlanningInteractive Control... Beliefs Desires Intentions What to do? Execute Observe

  8. Architecture of a Collaborative Dialogue Agent • Overall behavior controlled by Task Manager • Dialogue behavior under Dialogue Manager’s control • Dialogue Manager components operate independently and asynchronously • All components use shared BDI knowledge base (KB)

  9. Outline • Architecture of Mixed-Initiative Dialogue Systems • Dialogue Systems • Agents • Collaborative Agents • Components of Collaboration • Ontology of collaborative problem solving acts • User- and system-initiative APIs • Extended Example • User initiative • Interpretation • System initiative • Collaborative behavior • Proposals • Agreement • Other Issues

  10. Collaboration Manager Interpretation Generation • Ontology of collaborative problem solving acts • User initiative: intentions from CM to TM • System initiative: intentions from TM to CM CPS Act Ontology Collaboration Manager User Initiative API System Initiative API JointIntentions Task Manager

  11. Collaborative Problem Solving Act Ontology Interpretation Generation • Ontology of collaborative acts that an agent can perform • Propose, Accept, Reject, Report, Check, Ask • Modality-independent • Uses BDI language for content • Bel, Des, Commit • “Know-ref” or “Know what is” forms • e.g. “Let’s buy a book” • (Propose USR SYS (Commit (USR SYS) (Do PURCHASE123)))where PURCHASE123 has object = BOOK456, ... Collaboration Manager CPS Act Ontology User Initiative API System Initiative API JointIntentions Task Manager

  12. User Initiative API Interpretation Generation • User utterances interpreted using BDI state and principles of collaborative activity • Under the right conditions, user initiative results in suggestion that the Task Manager adopt a joint intention • Decision to adopt is up to the agent (Task Manager), not the dialogue subsystem • Asynchronous accept/reject from TM drives dialogue • If accepted, new intention drives agent behavior CPS Act Ontology Collaboration Manager User Initiative API System Initiative API JointIntentions Task Manager

  13. System Initiative API Interpretation Generation • Arises from Task Manager assigning a collaborative goal (joint intention) to Collaboration Manager • CM determines dialogue behavior based on BDI state and principles of collaborative activity • TM asynchronously informed of success or failure CPS Act Ontology Collaboration Manager User Initiative API System Initiative API JointIntentions Task Manager

  14. Outline • Architecture of Mixed-Initiative Dialogue Systems • Dialogue Systems • Agents • Collaborative Agents • Components of Collaboration • Ontology of collaborative problem solving acts • User- and system-initiative APIs • Extended Example • User initiative • Interpretation • System initiative • Collaborative behavior • Proposals • Agreement • Other Issues

  15. Example Dialogue USER: I need to purchase an LCD projector for my class. I need XGA resolution and it has to be portable. Find an appropriate machine and get approval to purchase it by this afternoon. SYSTEM: Ok. What is the budget? USER: 700 dollars. SYSTEM: Ok. ... SYSTEM: I’ve found three projectors that match your specifications. [displays table] USER: Let’s go with the Hitachi. ... SYSTEM: There’s a problem with the projector purchase. It’s 4 p.m. and I haven’t yet received approval from Fred. USER: You don’t need to get approval if the cost is less than 1000 dollars. SYSTEM: Ok.

  16. Outline • Architecture of Mixed-Initiative Dialogue Systems • Dialogue Systems • Agents • Collaborative Agents • Components of Collaboration • Ontology of collaborative problem solving acts • User- and system-initiative APIs • Extended Example • User initiative • Interpretation • System initiative • Collaborative behavior • Proposals • Agreement • Other Issues

  17. User Initiative (1) USER: I need to purchase an LCD projector for my class. • Direct report of a want or need: (Report USR SYS (Des USR (Done PURCHASE123))) • Response: “Thanks for letting me know.”

  18. User Initiative (2) USER: I need to purchase an LCD projector for my class. • Statement of a goal being pursued independently: (Report USR SYS (Commit USR (Done PURCHASE123))) • Response: “Good luck with that.”

  19. User Initiative (3) USER: I need to purchase an LCD projector for my class. • Proposal that a joint commitment (goal) be adopted: (Propose USR SYS (Commit (USR SYS) (Done PURCHASE123))) • If adopted, drives collaborative behavior

  20. Supporting Interpretation • Not the focus of this paper but... • Crucial for natural language • Also crucial for understanding actions of other agents • Need to interpret others’ actions/utterances consistently with one’s beliefs and principles of collaborative activity • We use the same procedures that drive the agent’s collaborative behavior in reverse to help interpret the user’s utterances • Effectively, interpret it as a given CPS act if we might have performed that act given the current BDI state

  21. User Initiative API • User proposals handled as suggestions that the system adopt a joint commitment: (suggest (Commit (USR SYS) (Done PURCHASE123)) • If accepted: “Ok.” • And new commitment drives subsequent behavior • If rejected: “No. ...”

  22. Reasoning Requirements for User Initiative • Have to be able to decide whether to adopt new commitments • Have to to be able to do this for oneself anyway • Some strategies for committing to goals: • Hardcode goals that are acceptable • If I know a way of achieving the goal (by myself or collaboratively), then adopt it • If achieving the goal is not incompatible with my beliefs, desires, and intentions, then adopt it • Can take initiative to gather information necessary for decision • Response to dialogue sub-system is asynchronous • See next section...

  23. Outline • Architecture of Mixed-Initiative Dialogue Systems • Dialogue Systems • Agents • Collaborative Agents • Components of Collaboration • Ontology of collaborative problem solving acts • User- and system-initiative APIs • Extended Example • User initiative • Interpretation • System initiative • Collaborative behavior • Proposals • Agreement • Other Issues

  24. System Initiative • Knowing when collaboration is necessary • Collaborative dialogue behavior • Proposals • Agreement

  25. Knowing When Collaboration Is Necessary • Task Manager needs to know what aspects of a task require joint commitment • Hard-coded into task models • Compiled into procedures by combining task models with general principles of collaboration • Incremental meta-decision of execution system • ... • For our example, assume it knows that we need to agree on the budget for purchasing the projector • Perhaps other aspects it can decide on its own • Perhaps have to agree on what aspects need to be agreed!

  26. System Initiative API • Task Manager dispatches collaborative goal to Collaboration Manager: (Commit-What-Is (USR SYS)(the budget of PURCHASE123)) • Collaborative goals drive dialogue behavior

  27. Collaborative Dialogue Behavior • Collaboration Manager is an agent that achieves its collaborative goals by performing CPS acts (leading to dialogue behavior) • Reactive procedures use BDI state to select CPS acts • These rules are compiled versions of the axioms defining the CPS acts ala Cohen & Levesque • For our example, it decides to: (RFP SYS USR (the budget of PURCHASE123))) • Generation produces: “What is the budget?”

  28. Proposals • Suppose the user answers: “700 dollars” • Context leads to interpretation as proposal: (Propose USR SYS (Commit-What-Is (USR SYS) (the budget of PURCHASE123) (* 700 dollars))) • Handled uniformly as suggestion that Task Manager adopt the commitment

  29. Agreement • Generation produces: “OK” • Collaboration Manager execution will now realize that the goal of reaching agreement on a value for the budget has been achieved • Reports successful completion of the goal to the Task Manager • Shared BDI KB contains the new knowledge about the budget

  30. Outline • Architecture of Mixed-Initiative Dialogue Systems • Dialogue Systems • Agents • Collaborative Agents • Components of Collaboration • Ontology of collaborative problem solving acts • User- and system-initiative APIs • Extended Example • User initiative • Interpretation • System initiative • Collaborative behavior • Proposals • Agreement • Other Issues

  31. Some Other Issues (1) • Is this only about natural language? • No. • But need an interface where: • Content is explicitly represented • Actions are represented as CPS acts • Seems like A Good Thing in any event

  32. Some Other Issues (2) • Isn’t there more to collaboration than dialogue? • Yes. • But dialogue is useful enough (and hard enough) to be worth focusing on • Our approach embeds dialogue in a general theory of collaborative activity

  33. Conclusions • An architecture for mixed-initiative problem solving assistants • Particular emphasis on dialogue systems • True mixed-initiative system • Goals and commitments come from either party • The system’s collaborative behavior is driven by a formal model of collaborative activity • Which also supports interpretation of user’s behavior • Completely domain- and application-independent • But very knowledge-intensive

  34. Conclusions • A practical way to build collaborative dialogue systems in many domains • Logistics planning • Personal health care • Command and control • Agent team coordination • Office assistant • NL transcription and knowledge mining • Crisis management • ...

  35. For More Information... George Ferguson ferguson@cs.rochester.edu James Allen james@cs.rochester.edu http://www.cs.rochester.edu/~ferguson/ http://www.cs.rochester.edu/research/cisd/

More Related