Artificial Neural Networks for Automatic Knowledge Acquisition in Multiple Real-World Language Domains

In this paper we describe a new approach for learning spontaneous language for multiple domains using artificial neural networks. This approach is based on a novel use of flat syntactic and semantic representations, fault-tolerant processing of noisy spontaneous language, and learning of individual domain-dependent subtasks. This approach has been implemented in our parallel and incremental architecture SCREEN (Symbolic Connectionist Robust EnterprisE for Natural language) which we have based on a careful selection and interaction of symbolic modules and artificial neural networks. We present the learned syntactic and semantic categorization and we examine the potential for increasing the portability by focusing on multiple corpora and domains. We claim that the general properties of learning, fault tolerance, and flat representations as implemented in SCREEN have the potential to increase the portability of neural network-based systems for spontaneous language analysis.