Artificial Neural Networks for Repairing Language

Spontaneous language contains many discontinuities caused by unusual order, false starts, repairs, repetitions, pauses, etc. Since data-driven artificial neural networks possess an inherent fault tolerance we use this property for dealing with such forms of “sequential noise”. We describe an approach for a flat syntactic and semantic interpretation of ill-formed utterances in spontaneous dialogs using symbolic methods for communication and simple known mappings as well as connectionist methods for unknown mappings. As an example for fault-tolerant flat analysis we describe the use of our syntactic and semantic representation for recovering from repairs in our hybrid approach using real–world spontaneous dialog utterances.