Analyzing the structure of legal sentences in legal document is an important phase to build a knowledge management system in Legal Engineering. This paper proposes a new approach to recognize logical parts in Vietnamese legal documents based on a statistic machine learning method -Conditional Random Fields. Beside linguistic features such as word features, part of speech features, we use semantic features of logical parts such as trigger features and ontology features to improve the result of the annotation system. Experiments were conducted in a Vietnamese Business Law data set and obtained 78.12% at precision and 68.72% at recall measure. Compare to state-of-the-art systems, it improves the result for recognizing some logical parts. Keywords-legal text mining; semantic annotation; machine learning; conditional random field; named entities recognitionI.978-1-4799-8044-4/15/$31.00