Domain-Level Reasoning for Spoken Dialogue Systems

In view of recent advances of technology, such as the miniaturisation of increasingly powerful computing devices with a parallel reduction of costs, as well as the increased performance of specialised algorithmic solutions such as speech recognition, the prospect of a daily use of speech dialogue technology in a variety of situations becomes more realistic than ever [1]. In the automotive infotainment environment, for instance, product-level progress has been made that enables an interaction design to move away from spelling-based entry of parts of navigation destinations to a natural input of complete addresses. At the same time the vocabulary size has been increased.

This chapter presents an introduction to the fundamental concepts and developments that shaped the field of dialogue systems research, and thus presents the basis of our work. Speech act theory, together with the plan-based notion of rational agency, has been the predominant framework of most theoretical approaches to both human-human and human-computer dialogue and is still highly influential today. However, conversely to this process that may be called “top down,” as it starts at the high level of intentions, in the dialogue systems community a “bottom up” movement has become popular. It can be characterised as aiming to build actual systems, especially using speech, that can realise specific dialogues at the surface level without a major theoretical or computational overhead. This has led to the development of dialogue scripting approaches, such as VoiceXML, among others.

This chapter provides an introduction to First-Order Logic (FOL) in terms of general concepts and notations as well as in terms of its use as a modelling and inference device. As such, FOL is the basis of both our proposed domain modelling approach (cf. Chapter 4) and our implemented reasoning engine CIDRE (cf. Chapter 5).

In this chapter, we describe our approach to a logic-based modelling of application domains. The modelling concerns structure of tasks that a user may want to achieve using a dialogue systems.We refer to these models¹ as domain theories. The task of the domain modelling can be broadly divided in two areas: The modelling of static information, i.e. structural information that does not change over time, and, on the other hand, the dynamic information which concerns time-dependent processes and changes in the states of objects. Of course, the two kinds of models are tightly interconnected in a typical application domain. Individual objects on the domain level, such as a geographical location or an appointment, are referred to as entities. Concerning the static information, the modelling of classes of entities can be distinguished from the modelling of their instances in a domain. The first problem is often referred to as the construction of a classification scheme, or ontology, while the latter has resemblance to representing items in a data base. Concerning the dynamic models, one of the most important aspects is the formalisation of so-called fluents and fluent change events. This approach is inspired by Allen and Ferguson’s Interval Temporal Logic (cf. Section 3.5). However, it is substantially adapted and extended in order to be useful in combination with our reasoning engine CIDRE (which is described in Chapter 5). We mainly defend the usefulness of our modelling approach by describing how important concepts from different domains are generalised into modules to form a library of reusable domain theories (cf. Section 4.4). In particular, we discuss the modelling of the TRAINS domain on the basis of the concepts introduced.

In this section, we discuss the design and implementation of a prototype that features the integration of our logic-based reasoning approach into a VoiceXML-based dialogue system and architecture. More precisely, a sample domain and dialogue application has been implemented and integrated with the reasoning engine which deals with the domain-level requirements and inferences arising from the interaction.

The construction of the VoiceXML-based prototype (cf. Chapter 6) illustrates some of the benefits, but also some of the drawbacks of relying on the concepts that a dialogue scripting language provides. Some problems of the integration in that prototype are technical, for instance, the need to interrupt the form interpretation algorithm. Other problems are more fundamental. In particular, the form-based representation of a dialogue does not directly provide an adequate representation of some aspects of our requirements-inferences dialogue management approach. As a consequence this prototype does not take full advantage of the capabilities the reasoning engine offers.

In this chapter, we describe the implementation of a revised prototype that has been developed on the basis of the information state-based dialogue management approach described in the last chapter. This chapter includes the description of a basic infrastructure for realising the integration between the reasoning engine CIDRE, the prototype dialogue manager, and components of the linguistic front-end.

We have presented a novel architecture integrating domain-level reasoning with dialogue management in spoken language dialogue systems. The approach consists of an adaptation and extension of the conventional SLDS architecture in terms of a dialogue manager that takes advantage of a proposed interactive reasoning engine. The task of the reasoner is to process the domain-level information in a requirements-inferences approach. This means that with regard to the domain level the interaction between the user and system is characterised by the user asserting or retracting requirements and the system providing inferences. By relying on this approach, various dialogue activities can be realised, such as hypothetical reasoning and conflict resolution. In particular, both of these activities correspond to choosing adequate sets of user requirements. Using our approach the dialogue manager is able to present inferences and relate them to the particular user requirements the inference is based on. Thus, the basis for an explanation and interactive conflict resolution is provided. We argue that these functionalities should be an essential part of a flexible and user-friendly SLDS. Since the reasoning engine manages the domain-level information, it can also be considered an interface between the dialogue manager and application domain functionality.

Description logic (DL) is a restricted form of first-order logic that has been popular for representing knowledge in various application areas, such as engineering or bioinformatics. In particular, description logic is connected to the term “semantic web” which is an effort to establish standard knowledge representation formalisms and algorithms (e.g. the Web Ontology Language, OWL, [102]) to enable a distributed and automated processing of information published on the web.