Issues in Spoken Dialogue Systems for Human- Computer Interaction

1. Introduction

Ideal communication model for Human-Computer Interaction (HCI) can be derived from human –human interaction which includes both verbal and non-verbal components. Non-verbal communication includes sign languages, facial expressions, gestures, emotions, lip movement, etc., while the main component of verbal communication is a natural language utterances. Understanding human-human interaction process and deriving a computational model of it is an enigmatic problem and achieving an interaction with computers that can be called close to human–like, if not similar, is still too far to achieve. However, achieving human-like interaction has been a desired goal of human computer interaction and the fascinating idea of using natural languages for interacting with computers has long been a research topic of Artificial Intelligence (AI) and HCI researchers. Systems, like ELIZA, which uses natural language interface, have been developed as early as in 60s. Science fictions are full of wonderful pieces of fantasies about spontaneous human-like conversations with computers. The idea of developing “talking computers” (Samuel et al., 2010) has been haunting researchers in AI and speech technology for past few decades. However, spoken language as a means of communication with computers has become a reality only in recent past owing to the rapid advancements in computing, speech and language technology. Several research prototypes as well as commercial applications are now available that use spoken language communication. Interfaces that use voice as input for controlling appliances, creating documents, searching in an existing database, are already in place. However, systems equipped with such voice-based interfaces have limited capability in that they do not engage the user in a natural conversation. They accept speech inputs, process them and perform some actions or report an error. Research efforts are continuing to develop spoken dialogue systems which “provide an interface between a user and computer-based application in many cycles in a relatively natural manner” (McTear, 2002). Figure 1 shows a wide range of tasks of varying complexity suitable for dialogue-based interfaces. However, supporting such interactions brings additional complexities, such as interpreting the user input to get involved in a conversation, handling uncertainty, knowing its state, confirming, asking to repeat, etc.

Figure 1.

Example dialogue applications with increasing order of complexity

In order to have more effective interaction dialogue systems must integrate speech with other natural modalities such as facial expressions, gestures, eye movements, etc. This has lead to the idea of Embodied Conversational Agents (ECA) (Traum & Rickel, 2002) either in the form of animated humanoid avatars or, as talking heads. These agents communicate with the user to plan a task that an agent is able to carry out independently, e.g., book travel tickets, prepare diet chart to improve poor dietary habits, assist in planning investment, etc. Rich models have been developed along this line. However, the core of ECA work is not the use of human language technology but the use of facial expressions and gestures in order to bring emotions and politeness in conversation. Research efforts are being made focusing on individual modality as well as to integrate various modalities and to bring emotions in conversation to achieve truly human-like interaction. More recently, particularly the techniques used in the COMPANIONS project (http://www.companions-project.org) has led a group of researchers to believe that emotion and politeness is more dependent on the language than its originator realized, and the best place to locate them may be speech and language phenomenon instead of facial expressions and gestures (Wilks et al., 2011). This has made speech and language processing issues the core concern of dialogue systems. Important challenges in the development of multimodal interface have been discussed in Griol et al. (2011).