Statistical language models with structural elements

Statistical n-gram language models, which can be seen as a variant of finite state grammars with transition probabilities written on the arcs of the graph, have been applied successfully in speech technology. We will present the new theoretical concept of structural elements and outline an automatic method to improve n-gram language models by distributing those structural elements in word sequences. In our model, n-grams that have not occurred in the training will be interpolated by ngrams containing structural elements. We give a detailed description of the algorithm and present first results of a system trained on a small corpus which consists of spontaneous speech dialogues.