Syntactic Disambiguation by Learning Weighted Government Patterns from a Large Corpus

A method of syntactic disambiguation based on proper prepositional phrase attachment, or, more generally, attachment of the clauses in specific grammatical cases, is described. The research was based on Spanish and Russian material. The data set built and used by the procedure is a kind of a syntactic government patterns dictionary. The algorithm requires a morphological and a syntactic parser and assigns probability weights to the variants built by the parser. No manual markup is required. At the training stage, the procedure works iteratively on a large text corpus, in alternating steps re-estimating the frequencies of individual government patterns and then the weights of the variants. The method is compatible with other methods of estimation of the variants. The data set built by the algorithm is useful for compilation of a combinatory dictionary for human readers. Some generalizations of the method are discussed.