AbstractsComputer Science

The objective of this thesis was text-based prediction of phrasal prominence. Improving natural sounding speech synthesis motivated the task, because phrasal prominence, which depicts the relative saliency of words within a phrase, is a natural part of spoken language. Following the majority of previous research, prominence is predicted on binary level derived from a symbolic representation of pitch movements. In practice, new classifiers and new models from different fields of natural language processing were explored. Applicability of spatial and graph-based language models was tested by proposing such features as word vectors, a high-dimensional vector-space representation, and DegExt, a keyword weighting method. Support vector machines (SVMs) were used due to their widespread suitability to supervised classification tasks with high-dimensional continuous-valued input. Linear inner product and non-linear radial basis function (RBF) were used as kernels. Furthermore, hidden Markov support vector machines (HM-SVMs) were evaluated to investigate benefits of sequential classification. The experiments on the widely used Boston University Radio News Corpus (BURNC) were succesful in two major ways: Firstly, the non-linear support vector machine along with the best performing features achieved similar performance than the previous state-of-the-art approach reported by Rangarajan et al. [RNB06]. Secondly, newly proposed features based on word vectors moderately outperformed part-of-speech tags, which has been inevitably the best performing feature throughout the research of text-based prominence prediction.