Co-training using prosodic, lexical and morphological information for automatic sentence segmentation of Turkish spoken language

Sentence segmentation of speech aims to detect sentence boundaries in a stream of words output by the speech recognizer. Sentence segmentation is a preliminary step toward speech understanding. It is of particular importance for speech related applications, as most of the further processing steps; such as parsing, machine translation, and information extraction, assume the presence of sentence boundaries. Typically, statistical methods require a huge amount of manually labeled data, which is time and labor consuming process to prepare. In this work, novel multiview semi-supervised learning strategies for the solution of sentence segmentation problem are proposed. The aim of this work is to nd e ective semi-supervised machine learning strategies when only a small set of sentence boundary labeled data is available. This work proposes three-view co-training and committee-based strategies incorporating with the agreement, disagreement and self-combined strategies using lexical, morphological and prosodic information, and investigates the performance of the proposed learning strategies against the baseline, self-training, and co-training. The experimental results show that the proposed learning strategies highly improve the sentence segmentation problem since data sets can be represented by three redundantly su cient and disjoint feature sets.