Speech and Human-Machine Dialog
Description
Speech and Human-Machine Dialog focuses on the dialog management component of a spoken language dialog system. Spoken language dialog systems provide a natural interface between humans and computers. These systems are of special interest for interactive applications, and they integrate several technologies including speech recognition, natural language understanding, dialog management and speech synthesis.
Due to the conjunction of several factors throughout the past few years, humans are significantly changing their behavior vis-à-vis machines. In particular, the use of speech technologies will become normal in the professional domain, and in everyday life. The performance of speech recognition components has also significantly improved. This book includes various examples that illustrate the different functionalities of the dialog model in a representative application for train travel information retrieval (train time tables, prices and ticket reservation).
Speech and Human-Machine Dialog is designed for a professional audience, composed of researchers and practitioners in industry. This book is also suitable as a secondary text for graduate-level students in computer science and engineering.
More details
Other editions
Additional editions
Content
1. Introduction
Nobody was able to foresee, 50 years ago, that the interaction between humans and machines would become increasingly sophisticated (high-level programming languages, multimedia graphic interfaces, etc.) and that such a huge number of people would use human-machine interfaces in the professional domain and also in their private lives. However, the main problem of dialog with computers relies in the difference between the formal languages, created to control the machines, and the natural language, used and understood by humans. This chapter is devoted to the way of how to fill this gap between the two types of languages. We will examine for this purpose a particular example of the current work on spoken natural language understanding.
Since for the most part, natural language research has its roots in symbolic system approaches, modeling of language understanding is often motivated by capturing cognitive processes, thus, integrating theories from linguistics and psychology. These cognitive models, however, are mainly established on the basis of written texts and often implemented using hand-crafted rules. Cognitive models presume the syntactic correctness of a sentence and in doing so, ignore spontaneous speech effects. The problem of ellipsis in spontaneous dialogs was analyzed by Morell (1988), but only few implementations deal with this issue in practice. Minor work has been dedicated to methods for recovery of interpretations in which parses are incomplete. (For example the utterance how much time does it take in New York for limousine service could be interpreted as the time either necessary to get a limousine at the airport or the transportation time between the airport and downtown New York.) Various analyses (Chapanis, 1979) considered spontaneous speech effects, including disfluencies, e.g., hesitations, repeated words and repairs or false starts, which are common in normal speech, as afternoon flight from from Denver to San Francisco. Only a few research prototype systems, e.g., CMUPHOENIX (Ward, 1994), take these effects into account. The ability to cope with spontaneous speech is crucial for the design of systems in real world applications. The following sections, taken from (Minker et al., 1999), introduce the case grammar formalism, used by the system L'ATIS (Bennacef et al., 1994), as well as the train travel-based systems MASK (Gauvain et al., 1997) and ARISE (Lamel et al., 1998). Another example of a case grammar-based implementation is the CMU-PHOENIX parser (Ward, 1994).
2. Case Grammar Formalism
In the domain of spoken language information retrieval, spontaneous effects in speech are very important. These include false starts, repetitions and ill-formed utterances. Thus, it would be improvident to base the semantic extraction exclusively on a syntactic analysis of the input utterance. Parsing failures due to ungrammatical syntactic constructs may be reduced if those phrases containing important semantic information could be extracted whilst ignoring the non-essential or redundant parts of the input utterance. Restarts and repeats frequently occur between the phrases. Poorly syntactical constructs often consist of well-formed phrases which are semantically correct.
One approach to extracting semantic information is based on case frames. A frame is a data structure, a type of knowledge representation in artificial intelligence (Minsky, 1975). It is a cluster of facts and objects that describe some typical object or situation, together with specific inference strategies for reasoning about the situation (Allen, 1988). A frame may be thought of as a network of nodes and relations. The top levels of a frame are fixed, and represent facts that are always true about the supposed situation. The lower levels have terminals or slots that need to be filled-in by specific instances of data. Each terminal can specify conditions its assignments must meet. The assignments themselves are usually smaller sub-frames. Collections of related frames are linked together into frame systems.
The original concept of a case frame as described by Fillmore (1968) is based on a set of universally applicable cases. They express the relationship between a verb and the related syntactic groups. Fillmore's cases correspond in fact to the Latin declensions: nominative, accusative and instrumental. Bruce (1975) extended the Fillmore theory to any concept-based system and defined an appropriate semantic grammar, whose formalism is given in Figure 2.1.
The case grammar uses in fact the stereotypical data structure of frames (Minsky, 1975). However, in order to fill in the frame slots, the notion of syntax (Fillmore, 1968) is added in the form of local marker-constraint relations. In the example query
