Combining black-box taggers and parsers for modern standard Arabic

A number of trainable dependency parsers have been presented in the literature. These parsers require tagged input: this may potentially cause a problem, because taggers are not in general 100% accurate, and any errors in tagging are likely to lead to errors in the output of the parsers. The current paper investigates the relationship between tagging errors and parsing errors. The investigation is carried out on Arabic text, using specific taggers and parsers, but the lessons that can be learned are applicable to other languages and other tools of the same kind.