Syntactically annotated corpora of Estonian

Syntactically annotated corpora are needed 1) to train and test parsers and various language technological products grammar checkers, information retrievers and extractors, machine translators etc; 2) to check the agreement of existing linguistic theories with the real language usage. The corpora can be annotated on different levels of depth. In shallow syntactically annotated corpora a syntactic function is determined for every wordform; in deep syntactically annotated corpora (treebanks) also the dependency structure is determined for every sentence (graphically represented as a tree). There exists a Constraint Grammar shallow syntactic parser for Estonian, developed by K. Müürisep and T. Puolakainen. To train the parser, we have annotated texts of written Estonian (20 000 words of fiction, 6 000 words of legal text and 10 000 words of newspaper texts). By now we have extended the size of the corpus up to 200 000 words. We have also started to build two versions of Estonian treebank. A parallel corpus of 50 sentences from J.Gaarder's novel "Sophie's world" has been annotated and aligned and a Constraint Grammar plus phrase structure hybrid treebank is being developed, currently consisting of 2400 automatically generated trees, 149 of them manually revised.