This work proposes a novel password guessing approach based on the identification, extraction and recombination of meaningful syntactic patterns present in human-chosen passwords. The proposed method exploits the existence of these patterns across user-selected passwords in order to effectively reduce the search space to be explored during the password guessing process. The password guessing scheme follows a two stage strategy. In the first step, a novel algorithm based on machine learning principles, identifies and extracts the syntactic meaningful patterns from a dataset of passwords. Then, in a second stage, these parts-of-passwords previously segmented are recombined in order to generate new statistically relevant password candidates that are used against a blind evaluation set. The experimental results show that this novel approach is able to guess complex passwords usually robust to traditional password guessing techniques.