Towards a treebank for all tastes

Syntax is a discipline of many theories, and it is accordingly difficult to build a syntactically annotated corpus that would not put off at least some syntacticians by an alleged or real theoretical bias. Yet despite appearances and focus on slightly different sets of linguistic phenomena the theories strive to describe and explain the same object – a natural language. In fact there is a large pool of implicit wisdom shared by all syntactic theories and a significant overlap of linguistic knowledge can be extracted from all theory-specific formats. Thus a treebank offering different views of syntactic annotation while based on a single core pattern need not be a dream out of touch with reality. In addition to constituency and dependency trees of various shapes, suited to the taste of experts in linguistics, one of the views may be close to the representation of syntactic structure to which Czech students are exposed at the higher elementary and secondary levels. Such a treebank should indeed be useful beyond academic community to other professionals and lay users interested in language and linguistics. Obviously, for most of them the bigger the better, but not at an unbearable decrease in reliability. Yet the largest existing treebanks reach the relatively modest sizes of several million words, an insufficient number for many tasks. The reason is the cost of manual checking needed to improve the error rate of automatic syntactic annotation tools, which still perform much less reliably than part-of-speech taggers. However, to match the size of a balanced POS-tagged corpus, the use of automatic parsing tools without manual checking is inevitable. Building on previous efforts in treebank annotation, especially the Prague Dependency Treebank (PDT) and the NEGRA/TIGER Corpus (Hajič, 2006; Hajič et al., 1998; Skut et al., 1997, i.a.) we want to make a further step towards a large corpus with a reasonably reliable, automatically assigned syntactic annotation. With this aim in mind, we propose an explicitly defined annotation scheme consisting of a linguistically founded, potentially underspecified morphological and syntactic core, complemented by multiple interaction shells, customizable in shape and detail according to the preferences of humans or computer applications, accessible to lay users and satisfying demands of experts at the same time (§2).