Analyzing and Integrating Multilingual Dependency Parsers
Speaker: Ryan McDonald , Google
In recent years there has been an abundant amount of work on dependency parsing. This work has led to the production of a number of language independent parsing systems as well as a significant jump in state-of-the-art parsing accuracy. Almost every recent model can be classified into one of two paradigms. This first, called transition-based parsing, parameterizes models over actions in abstract state machines. The second, called graph-based parsing, parameterizes models over substructures of dependency graphs. These models are fundamentally different. However, recent empirical evaluations have shown that parsing accuracy between the models is not statistically significant. In this talk I will analyze the empirical properties of each model beyond simply accuracy and show that the models do indeed have marked differences and that these differences align with theoretical expectations. I will then discuss methods for integrating the two models and show that simple techniques yield large improvements. An error analysis of the integrated models reveals many interesting properties.