A Multilingual Method for Clause Splitting

This paper addresses the clause splitting problem and proposes a multilingual method for detecting clause boundaries in unrestricted texts. The method combines language independent machine learning techniques with language specific rules in order to take the first step in building the hierarchical structure of sentences. The results of a machine learning algorithm, trained on an annotated corpus, are processed by a rule-based module which deals with clause boundaries not included in the learning process. Formal indicators of coordination and subordination, together with verb type information (finite or non-finite) are used for identifying clause boundaries. The method was evaluated on Romanian and English and the F-measure for clause start detection is 95% for Romanian and 92% for English.