📄 Floresta Sintá(c)tica: A treebank for Portuguese

Summary

The paper reviews the first year of work on Floresta Sintá(c)tica, the first open syntactically annotated treebank for Portuguese. The project is run jointly by the VISL group at the University of Southern Denmark and the Computational Processing of Portuguese project. The material is newspaper text from the CETEMPúblico corpus, chosen because it has no copyright problems. First, sentences were segmented manually using linguistic criteria. Fragments without ordinary syntax, such as poems, sports results and address lists, were marked as <sic>. The text was then parsed automatically by PALAVRAS, a parser based on Constraint Grammar (CG). Annotators checked the output in two stages: first the word-level tags, then the constituent trees. A separate PSG-like program builds these trees from the already corrected CG annotation. Because of this split, first-level errors do not reach the trees, and the parser itself is improved alongside the manual revision. The treebank is published on the web with search, download and graphical tree viewing. It is intended to serve as a public benchmark for evaluating syntactic parsers of Portuguese.

Key points

• In the first year, 1,427 sentences were fully revised. This is about 10% of the first million words of the CETEMPúblico corpus, and every annotation level in them was checked by hand.
• Floresta Virgem is the full million words, annotated automatically by PALAVRAS. It has not yet been checked by hand, but the parser has already been tuned to the principles worked out during manual revision.
• During preprocessing, 8–9 thousand new lexemes were added to the PALAVRAS lexicon. This widened morphological coverage and balanced the representation of the Brazilian and European varieties of Portuguese.
• Errors in the newspaper texts themselves were deliberately left uncorrected. For frequent headlines and for colloquial and incomplete constructions, special sentence-type markers were introduced, such as #D for discourse and #E for ellipsis.
• Annotation consistency was checked with a test that compared the annotations of three annotators pairwise. Each disagreement was classified either as an error or as a genuine ambiguity.