Abstrakt
The paper presents two experiments with ensemble parsing, in which we obtain a 1.4% improvement of UAS compared to the best parser. We use five parsers: MateParser, TurboParser, Parsito, MaltParser a MSTParser, and the data of the analytical layer of Prague Dependency Treebank (1.5 million tokens).
We split training data into 10 data-splits and run a 10-fold cross-validation scheme with each of the five parsers. In this way, we obtain large parsed data to experiment with.
In one experiment, we calculate partial accuracy rates of each parser according to a list of parameters, which we then use as weights in a combination of parsers using an algorithm for finding the maximum spanning tree. In the other experiment, we calculate success rates for agreements of parsers (e.g.
Mate+MST vs. Turbo+Malt), and use these rates in another combination of parsers.
Both experiments achieve an UAS above 90.0% (1.4% higher than TurboParser), the experiment with accuracy rates achieves better LAS.