310 likes | 514 Views
Mixed-initiative Dialog Management. Reporter Chun-Feng Liao NCCU Department of Computer Science. Papers. Ramakrishnan, N., et al., “ Mixed-Initiative Interaction = Mixed Computation ” , Proc. ACM SIGPLAN Workshop PEPM ’ 02, January 2002.
E N D
Mixed-initiative Dialog Management Reporter Chun-Feng Liao NCCU Department of Computer Science
Papers • Ramakrishnan, N., et al., “Mixed-Initiative Interaction = Mixed Computation”, Proc. ACM SIGPLAN Workshop PEPM’02, January 2002. • Eric Nyberg et al. "DialogXML: Extending VoiceXML for Dynamic Dialog Management“ Proceedings of the Human Language Technology Conference ,2002.
Agenda • Introduction • Mixed-initiative interaction • Partial Evaluation • VoiceXML Dialog Management • DialogXML Dialog Management • Conclusion • Future Work
Introduction • 定義與說明Mixed-initiative Dialog System 相關術語. • 說明VoiceXML interpreter中是使用 Partial Evaluation技術來處理對話機制的FIA演算法 (Form Interpretation Algorithm)。 • 以DialogXML來解決使用VoiceXml難以處理複雜的對話機制的問題。
What is Mixed Initiative • 老公:Hi,告訴你一個天大的消息… • 老婆:我也有一個天大的消息要告訴你… • 老公:那我們誰先講 ? Initiative :在此指對話狀態中的「主動權」。 以上的對話二人都嘗試拿到主動權,但若一方沒有讓出,對話無法進行,所以要商議誰先掌握主動。
Mixed Initiative (2) • 傳統的語音系統設計模式中,對話完全由系統主導,所以是System Initiative 。 • 假設A與系統進行對話,若A與系統輪流掌握對話主動權,就是Mixed Initiative。
System-initiative System-initiative : 都是系統在主導對話。
Using Partial-evaluation method to Model Dialog Systems • Partial-evaluation原本是程式語言領域中,用來做Program最佳化的技術。 • 假設一個Function 有二個參數,而我們事先知道其中一個,利用已知的參數先執行程式的一部份。
假設 • 本論文所探討的Dialog形態限定如下 : • Dialogs as task-oriented , to fill a set of slots. • 共二位參與者,其中一個是電腦,另一個是人。 • Mix-initiative arises from unsolicited reporting . ( 紅色的terms下面幾張投影片會解釋 )
Software Models of Dialog-based Interactive Systems • [Allen 01] • Finite-state machines • Slot-and-filler • Frame-based • Planning • Agent-based programming J. Allen, D. Byron, M. Dzikovska, G. Ferguson,L. Galescu, and A. Stent. “ Towards Conversational Human-Computer Interaction ” . AI Magazine, 2001.
Tiers of Mixed-initiative Interaction • [Allen 99] Four levels of mixed-initiative . • Unsolicited Reporting • Sub dialog initiation • Fixed subtask initiative • Negotiated mixed initiative J.F. Allen, C.I. Guinn, and E. Horvitz. “Mixed-Initiative Interaction “.IEEE Intelligent Systems, Vol. 14(5):pages 14--23, Sep-Oct 1999.
Unsolicited Reporting • I – Initiative • R – Response • 小寫s – System • 小寫c – Caller 客戶爭奪主導權
Unsolicited Reporting 客戶爭奪主導權
範例說明 • 某家的pizza的訂pizza語音系統,對話流程如下: • 接通後先跟caller問好 • 詢問size、topping與crust • comfirm
Partial Evaluation • 其實系統的目的是得到這三個值。 • 前後次序及方式則不介意。 • 使用者可以一句填二個值,系統會再詢問未填的值。 • 如果是熟客的話可能一次三個值就全可填滿。 Form size crust topping
Dialog Analysis • 整個對話流程可化成如下的狀態圖 • 問題變成 : 如何從d1走到confirm,並且每一個點恰好都經過一次? • 如何寫程式處理?
將適合PE的部份抽出 • 將slot-filling的部份抽出成Partial evaluation layer。 • 上層使用一般機制處理即可,下層使用PE實作會更有效率。 使用PE技術
PE Programming Model PE 傳統做法
VoiceXML Dialog Management Architecture • Vxml documents defines a finite state machine and describes a sequence of interactions . • Dialog in Vxml documents involves a sequence of forms and menus . • Forms specify a set of slots to be filled by user input .
VoiceXML Form-level Grammar • The core of interpreter is FIA (Form Interpretation Algorithm) • Programmer may provide form-level grammar to make mixed-initiative interaction . • All dialogs are Partially Evaluated .
VoiceXML Dialog 機制的問題 • Dialog彼此互相獨立,不保留上一Dialog的狀態(Stateless),在智慧型Agent的用途上,有些Agent需要trace之前的對話狀態。 • 需要使用大量的Forms來表達複雜的對話,對程式員來說很麻煩。 • 相較於Philips SpeechMania HDDL等既存的Dialog Modeling語言,缺乏許多重要機制。(to be found out…)
VoiceXML 缺乏state機制 • Natural dialog has a notion of state • Dialog designers often think in terms of state-transition networks. • VoiceXML does not explicitly model states and transitions among states . VoiceXML的Form無法輕易表達上面這個流程 !
使用DialogXML加強VoiceXML的Dialog功能 • DialogXML將對話狀態以state machine的方式來表達,再自動轉成VoiceXML。
DialogXML System Architecture DialogXML文件 Browser
Vxml與Dxml的比較 • VoiceXML有傳統程式語言looping與conditional control的特色。而DialogXML採用Dialog 設計人員較習慣的「declarative」方式(類似XSLT) • Vxml使用FIA來實作,而Dxml則是透過Transform機制實作。 • Dxml可將複雜的對話描述簡化。
Conclusion • Mix-initiative provides a more natural and personalized mode of interaction. • Partial Evaluation can be usefully viewed as a programming model for realizing mixed-initiative dialogs. • VoiceXML is lack of some advance dialog functionalities , DialogXML slove this problem by programming transformation.
Future Work • 原系統是使用OpenVXI來修改,目前只提供文字界面,未整合真正的語音(Speech Recognition)功能。 • 增加information-seeking功能