Unsupervised learning for syntactic disambiguation

We present a methodology framework for syntactic disambiguation in natural language texts. The method takes advantage of an existing manually compiled non-probabilistic and non-lexicalized grammar, and turns it into a probabilistic lexicalized grammar by automatically learning a kind of subcategorization frames or selectional preferences for all words observed in the training corpus. The dictionary of subcategorization frames or selectional preferences obtained in the training process can be subsequently used for syntactic disambiguation of new unseen texts. The learning process is unsupervised and requires no manual markup. The learning algorithm proposed in this paper can take advantage of any existing disambiguation method, including linguistically motivated methods of filtering or weighting competing alternative parse trees or syntactic relations, thus allowing for integration of linguistic knowledge and unsupervised machine learning.