Although parsing performances have greatly improved in the last years, grammar inference from treebanks for morphologically rich lan- guages, especially from small treebanks, is still a challenging task. In this paper we in- vestigate how state-of-the-art parsing perfor- mances can be achieved on Spanish, a lan- guage with a rich verbal morphology, with a non-lexicalized parser trained on a treebank containing only around 2,800 trees. We rely on accurate part-of-speech tagging and data- driven lemmatization in order to cope with lexical data sparseness. Providing state-of- the-art results on Spanish, our methodology is applicable to other languages.