We have described the underlying principles, the implemented architecture, and the evaluation of a new screening approach for learning the analysis of spoken language. This work makes a number of original contributions to the fields of artificial intelligence and advances the state of the art in several perspectives: From the perspective of symbolic and connectionist design we argue for a hybrid solution, where connectionist networks are used wherever they are useful but symbolic processing is used for control and higher level analysis. Furthermore, we have shown that recurrent networks provided better syntactic and semantic prediction results than 1-5 grams. From the perspective of connectionist networks alone, we have demonstrated that connectionist networks can in fact be used in real-world spoken-language analysis. From the perspective of natural language processing we argue that hybrid system design is advantageous for integrating speech and language since lower speech-related processing is supported by fault-tolerant learning in connectionist networks and higher processing and control is supported by symbolic knowledge structures. In general, these properties support parallel rather than sequential, learned rather than coded, fault-tolerant rather than strict processing of spoken language.
The main result of this paper is that learned flat representations support robust processing of spoken language better than in-depth structured representations and that connectionist networks provide a fault-tolerance to reach this robustness. Due to the noise in spontaneous language (interjections, pauses, repairs, repetitions, false starts, ungrammaticalities, and also additional false word hypotheses by a speech recognizer) complex structured possibly recursive representations often cannot be computed using standard symbolic representations like context-free parsers. On the other hand, there are tasks like information extraction from of spoken language which may not need an in-depth structured representation. We believe our hybrid connectionist techniques have considerable potential for such tasks, for instance for information extraction in restricted but noisy spoken-language domains. While an in-depth understanding like inferencing for story interpretation needs complex structured representations, a shallow understanding for instance for information extraction in noisy speech language environments will benefit from flat, robust and learned representations.