Parallel treebanks, which comprise paired source-target parse trees aligned at sub-sentential level, could be useful for many applications, particularly data-driven machine translation. In this paper, we focus on how translational divergences are captured within a parallel treebank using a fully automatic statistical tree-to-tree aligner. We observe that while the algorithm performs well at the phrase level, performance on lexical-level alignments is compromised by an inappropriate bias towards coverage rather than precision. This preference for high precision rather than broad coverage in terms of expressing translational divergences through tree-alignment stands in direct opposition to the situation for SMT word-alignment models. We suggest that this has implications not only for tree-alignment itself but also for the broader area of induction of syntaxaware models for SMT.