资源预览内容
第1页 / 共14页
第2页 / 共14页
第3页 / 共14页
第4页 / 共14页
第5页 / 共14页
第6页 / 共14页
第7页 / 共14页
第8页 / 共14页
第9页 / 共14页
第10页 / 共14页
第11页 / 共14页
第12页 / 共14页
第13页 / 共14页
第14页 / 共14页
亲,该文档总共14页全部预览完了,如果喜欢就下载吧!
点击查看更多>>
资源描述
Click to edit Master title style,Click to edit Master text styles,Second level,Third level,Fourth level,Fifth level,*,*,*,Agent-Based Dialogue Management,Discourse&Dialogue,CMSC 35900-1,November 10,2006,Possible Projects,Alternatives:,Implementation with short write-up,Analysis with longer write-up(max 15p),Topics:,Use toolkit to implement components of SLS,Implement/Extend a pronoun resolution algorithm,Identify phenomena not covered by theory&extend,E.g.Dialogue type functions in RST,Zero-anaphora in pronoun resolution,Analysis:,Compare reference behavior in chats/IM to other,Turn-taking in CallHome vs Switchboard,Roadmap,Agent-based Dialogue Management,Dialogue by Theorem Proving,Dialogue by Planning,Dialogue by Conversational Agency,Event-Driven Dialogue,Dialogue by Rational Interaction,Advantages and Disadvantages,Agent-Based Dialogue,Contrasts with State-and Frame-Based,Structurally constrained,Simplified,activity based semantics,Agent-based dialogue management,Employs AI planning or reasoning processes,More full linguistic interpretation,More complex model of interaction,belief,Wide variety of models,reasoning,Dialogue by Theorem Proving,Exemplar:Circuit Fix-it-Shop(Smith),Interaction between student and expert(S),Debugging a fault in a circuit,Dynamically develops solution,Based on state and user knowledge,Theorem prover determines task completion,Dialogue used to acquire missing axioms,Theorem Proving in Dialogue,Example:system needs to know if there is a wire b/t connectors 84 and 99,goal(c,learn,.(prop(84,99),exist,X),true),“missing axiom about wire present yes/no,Use language to ask user,and ask to add,Answers,then asks for help,Adds new subgoal to teach user,Integrates problem-solving,theorem proving,and dialogue,Dialogue by Planning,Integrate utterances with plan actions,Interpret intention behind utt,how achieve goal,Utterance=speech act,same structure as plan op,Example:ConvincebyInform,Constraints:Agent(Spkr),Agent(Hear),Prop(prop),Bel(Spkr,prop),Preconditions:At(Spkr,Loc(hear),Effects:Bel(Hear,prop),Buying ticket,requires agent using CBI to tell price,Dialogue by Planning,Issues:,Intention inference error-prone,cascading,Planning AI&NP-complete in general case,Strategies:,Differentiate domain plans and discourse plans,Train routing vs clarification subdialogues,Later approaches revise,Conversational Agency,Multi-agent planning system,Exemplar:TRAINS(Allen et al),Collaborative,incremental:,System:Low-level route planning,track info,User:High-level goals,can relax constraints if needed,Mixed-initiative planning:,Plan generation under constraints;criticism&revision,System as full-fledged agent in conversation,Utts are speech acts,interp user,maintain models,Discourse Obligations,Prior work:“co-operativeness,Interpret utt to determine speaker goals,System adopts goals as own,Trains:,Distinguish intentions and obligations of conv,Obligations represent what agent should do,E.g.if promise(A),generate obligation to achieve A,If S1 requests A,S2 should accept or reject,If utt not understood or wrong,should repair/correct,etc,Discourse Actor Algorithm,Act on any obligations before doing own goals,If user doesnt take lead,system will take initiative,While conversation not finished,If system has obligations,then address obligations,else if system has turn,then if system has intended conv.Acts,then call NL generator,else if some material is ungrounded,then do grounding,else if some proposal not accepted,then consider proposals,else if high-level goals unsatisfied,then address goals,else release turn to try to end conv,else if no one has turn,then take turn,else if long pause,then take turn,Event-driven Dialogue,Extension of frame-based approaches,Doesnt model or recognize user intention,Models system state,intentions,Uses more complex representation of context to perform semantic interp of user utt,Modular approach:dialogue,task,belief,Based on current belief/task state,select dialogue options:e.g.confirmations,requests,Contextual Interpretation,Interpretation of utt generates contextual functions,E.g.new_for_system(X):confirm(X)or spec(X,Y),repeated_by_user(X):cancel“confirm(X),Inferred_by_system(X):goal confirm(X),Modified_by_user(X):goal repair confirm(X),Negated_by_user(X):repair(X),Combine goals with initiating actions,Include meta-strategy of increasing system control when need repair,Rational Interaction,
点击显示更多内容>>

最新DOC

最新PPT

最新RAR

收藏 下载该资源
网站客服QQ:3392350380
装配图网版权所有
苏ICP备12009002号-6