Detail of publication

Abstract

The presented thesis is devoted to the spoken language understanding task. The thesis pre- sents a new discriminative model for this task. First, the spoken language understanding is described in the context of spoken dialog systems and in relation to an automatic speech recognition. Then the state of the art is presented. The current methods for spoken langu- age understanding are presented as well as methods related to the presented discriminative model. In the following chapter, the goals of the thesis are stated. The main goal is to develop a new discriminative model which is able to process uncertain input in the form of word-based or phoneme-based lattices and generate multiple output semantic hypotheses. One of the subgoals of this thesis is devoted to a research of method for effective combi- nation of statistical and knowledge-based approaches to spoken language understanding. The spoken language understanding is decomposed into three partial models. A concept model assigns the global meaning of the utterance in the form of abstract semantic tree. A semantic entity detection tags the local parts of the meaning with the semantic entities. An alignment model links these two semantic representations and forms a discriminative spoken language understanding model. The concept model is represented by the hierarchi- cal discriminative model which was developed as an extension of a statistical model based on semantic tuple classifiers. The semantic entity detection model performs the search for all occurrences of the semantic entities which are defined by knowledge-based context-free grammars. Then, the description of used data, recognition models, speech decoder, and evaluation methodology is presented. In the part devoted to experimental evaluation the values of specific parameters are selected and justified. The last chapter concludes the thesis and presents the overall performance of the presented method for spoken language understanding. It also describes the fulfilment of all goals of this thesis and presents the possible improvements and applications of the developed model.