Performance-oriented dependency parsing

Dependency parsing has become very popular among researchers from all NLP areas, because dependency representations contain very valuable easy-to-use information. In the last decade a lot of dependency parsers have been developed, each of them somehow special with its own unique characteristics. In the course of this thesis I have developed yet another parser MDParser. In this work I discuss the main properties of dependency parsers and motivate MDParser’s development. I present the state of the art in the field of dependency parsing and discuss the shortcomings of the current developments. To my mind the main problem of the current parsers is that the task of dependency parsing is treated independently of what happens before and after it. Therefore the preprocessing steps and the embedding in applications are neglected. However, in practice parsing is rarely done for the sake of parsing itself, but rather in order to use the results in a follow-up application. Additionally, current parsers are accuracy-oriented and focus only on the quality of the results, neglecting other important properties, especially efficiency. The design of MDParser tries to counter all these drawbacks. The evaluation of some NLP technologies is sometimes as difficult as the task itself. For dependency parsing it was long thought not to be the case, however, some recent works show that the current evaluation possibilities are limited. In this thesis I broadly present and discuss both intrinsic and extrinsic evaluation methodologies for dependency parsing. Both approaches have numerous disadvantages which I demonstrate in my work. The attachment scores, which are the most used metric of the intrinsic evaluation, do not differentiate between different dependency types, are being computed for the same portions of treebanks since many years, and thus often promote overfitting to this particular kind of data, which is especially dangerous because the data contains a certain amount of inconsistencies. The extrinsic evaluation, which eval-