Exakt och effektiv icke-projektiv dependensparsning

Tidsperiod: 2008-07-01 till 2010-03-31

Projektledare: Marco Kuhlmann

Finansiär: Vetenskapsrådet

Bidragstyp: Projektbidrag

Budget: 1 186 500 SEK

The overall goal of the proposed research project is to develop accurate and efficient methods for non-projective, data-driven dependency parsing. The purpose of parsing is the automatic analysis of the syntactic structure of natural language. Data-driven dependency parsing is distinguished from other approaches in that it employs syntactic representations based on word-to-word dependencies, and makes essential use of empirical data. Most of the previous work on dependency parsing has been confined to representations that satisfy a structural constraint called projectivity. While this constraint facilitates efficient computations, it does so at the cost of accuracy, as it rules out many syntactic analyses that occur in the data. Recent work has proposed various relaxations of the projectivity constraint, to identify classes of dependency structures that strike a better balance between accuracy and efficiency. In this project, we will develop and evaluate a data-driven parsing framework for such "mildly" non-projective dependency structures. Our project is the first comprehensive investigation of parsing methods based on these structures. Its outcome is of major signifcance for the field of data-driven dependency parsing, and for natural language processing in general.