In pace with the success of corpus-based approaches to theoretical and computational linguistics, the collocation of corpora has evolved into a research activity in its own. As the currently available corpora either lack annotation depth or closure, more data will be annotated in the future, preferably with minimal human intervention. This paper tries to approach the problem of treebank development from a logic-based learning perspective, applying several alternative forms of inference in order to assess their potential for automatically generalizing from a seed corpus annotated by hand to a corpus of POS annotated sentences, in order to automatically produce syntactic annotations that are good enough to use as training material for a parser. We shall show that syntactic annotations can be created automatically in large quantity via deductive and abductive explanation-based learning (EBL). Although these automatically created structures are not statistically representative with respect to many quantitative aspects of the treebank, the annotations may provide useful qualitative and quantitative data which might be extracted and reinvested into a parser. We shall compare the benefits and investments of automatically created structures to that of human-annotated structures and suggest some possible strategies how EBL approaches can be combined with manual annotation.