Improved prediction of stress in out-of-vocabulary words

I report on an improved method of stress assignment, based on Coleman (1993): the improvements arise from the addition of probabilities to the rules, enabling the most probable pattern of stress assignment to be estimated in a principled manner. Rule probabilities are estimated by using the dictionary as a training corpus. As with rule-based systems in general, although the method is far from perfect, the improvements it brings are particularly useful when it comes to the pronunciation of out-of-vocabulary words