Chapter 1A FRAMEWORK FOR DEVELOPINGCONVERSATIONAL USER INTERFACESJames Glass, Eugene Weinstein, Scott Cyphers, Joseph PolifroniMIT Computer Science and Artificial Intelligence Laboratory,Cambridge, MA 02139, USA{glass, ecoder, cyphers, joe}@csail.mit.eduGrace ChungCorporation for National Research Initiatives,Reston, VA, [email protected] NakanoNTT Corporation,Atsugi, [email protected] In this work we report our efforts to facilitate the creation of mixed-initiativeconversational interfaces for novice and experienced developers of human lan-guage technology. Our focus has been on a framework that allows developers toeasily specify the basic concepts of their applications, and rapidly prototype con-versational interfaces for a variety of configurations. In this paper we describethe current knowledge representation, the compilation processes for speech un-derstanding, generation, and dialogue turn management, as well as the user in-terfaces created for novice users and more experienced developers. Finally, wereport our experiences with several user groups in which developers used thisframework to prototype a variety of conversational interfaces.Keywords: Conversational interaction, spoken dialogue systems21. IntroductionIn recent years, many sophisticated conversational interfaces have been de-veloped that enable fluent, spoken communication between humans and ma-chines. Such systems are developed by speech and language experts, and re-quire significant effort over a sustained period to achieve good performance.For this reason, non-experts must overcome a significant hurdle to use humanlanguage technologies (HLTs) for their own applications. To address this issue,we have been developing a utility (called SpeechBuilder), which enablesrapid prototyping of spoken dialogue systems by both novice and expert de-velopers. In this paper we motivate the need for this research, describe ourapproach and progress, and describe several experiments we have performedwith novice users creating their own speech-based interfaces.In the following section, we briefly provide additional background on thecurrent state of directed and mixed-initiative dialogue interaction, and moti-vate the need for mechanisms to facilitate the development of mixed-initiativeconversational interface prototypes. We then describe the approach that wehave taken for our research in this area, and give an overview of the user inter-face we have created. We then describe the speech understanding, generation,and dialogue framework used, and describe several experiments we have con-ducted with different groups of users. Finally we compare our research torelated work, and describe our ongoing research in this area.1.1 BackgroundAlthough all spoken dialogue systems can be considered conversational tosome degree, they may be differentiated by the degree with which the systemmaintains control of the conversation, and the inherent amount of flexibilityprovided to the user to ask for a) what they want, b) in the way they want to askfor it, and c) when they want to ask it. In the most conservative approach, thecomputer takes complete control of the interaction. These directed-dialogueapplications typically require that the user answer a set of prescribed ques-tions, much like the touch-tone implementation of interactive voice responsesystems. Since the user’s options are restricted, completion of such transac-tions is easier to attain, and it is therefore not surprising that such systems havebeen the first to be successfully deployed on a wide scale [1–3].An alternative approach to human-computer interaction is based on the ideaof mixed-initiative dialogue between the user and the machine. This approachemploys a more flexible dialogue strategy that allows both the user and themachine to participate actively to solve a problem interactively using a conver-sational paradigm. Systems which are built with the mixed-initiative paradigmmust typically process more complex queries than their directed-dialogue coun-terparts [4], and are inherently more difficult to design and deploy. For thisA Framework for Developing Conversational User Interfaces 3reason, the majority of these kinds of systems remain under development inresearch laboratories [5–9], although some are beginning to be deployed pub-licly as well [10].1.2 MotivationAlthough mixed-initiative conversational interfaces are a natural and effi-cient means of communication, there are two fundamental technical barrierswhich limit their widespread use. First, it is difficult to configure the HLT re-quired to create a prototype system, and second, performance optimization istypically an iterative process that is application specific, and not fully auto-mated. Creating a robust, mixed-initiative conversational interface for a newapplication area currently requires a tremendous amount of effort from speechand language experts. The development of speech recognition and languageunderstanding technologies is mostly domain and language specific, requiringa large amount of annotated training data. Dialogue management is typicallyalso fine-tuned for the application, often with domain-dependent functional-ity. System development proceeds iteratively, with prototypes being used tocollect data that can then be used for system development, training, and eval-uation. This iterative process is crucial to achieve good performance. For ex-ample, the initial prototype of a mixed-initiative weather information systemtrained from several thousand utterances collected from a simulated “wizard-of-oz” scenario saw its error rates more than triple when it was first deployedover the telephone to a wide user population [11]. As utterances were contin-uously collected, the performance slowly improved to the point where it ulti-mately exceeded the original laboratory performance. However, this level ofperformance was only achieved through continuous data collection and systemrefinement over a period of time.For conversational interfaces to become as ubiquitous as the telephone, re-searchers must make it easier for developers to create systems that learn andimprove their performance automatically. However, there are many hurdles toeven allowing developers to create an initial prototype. For example, we mustaddress the problems of producing a conversational system in a new domainand language given at most a small amount of domain-specific training data.To achieve this
View Full Document
Unlocking...