Prosodic boundaries prediction in Russian using morphological and syntactic features

Abstract

The paper presents a comparison between three approaches towards prosodic boundary prediction in Russian text, namely a rule-governed method and methods involving statistical classifier and deep learning technique. The methods aim to predict all possible prosodic boundaries in text applying morphological and syntactic information. All used features were described in terms of Universal Dependencies framework by means of SyntaxNet parser. The rule-governed method runs in a bottom-up fashion, using the information about syntax group edges and applying data-driven and hand-written linguistic rules. For machine learning methods, conditional random fields classifier and bidirectional LSTM model were built, with such features as part-of-speech tag, syntactic dependency type, syntactic relation embedding and presence of syntactic link between the current and adjacent words. As experimental material, we used the data of CORPRES corpus, containing over 30 hours of professionally read speech. Used separately, morphological features are slightly superior to syntactic ones, and their combination improves the results. BiLSTM yields the highest F1 measure value of 90.4, as compared to 88.8 for CRF and 83.1 for rule-based method.