Performance Evaluation of WordNet-based Semantic Relatedness Measures for Word Prediction in Conversational Speech

The recognition of conversational speech is a hard problem. Semantic relatedness measures can improve speech recognition performance when using contextual information, as Demetriou [5] has shown. The standard n-gram approach in language modeling for speech recognition cannot cope with long distance dependencies [4]. Therefore J. Bellegarda [2] proposed combining n-gram language models, which are effective for predicting local dependencies, with latent semantic analysis for long distance dependencies. WordNetbased semantic relatedness measures can be used for word prediction using long distance dependencies, as in these examples from our experiments: