Pull your treebank up by its own bootstraps
Abstract
We analyze the performance of recent neural syntactic parsers in the task of bootstrapping a treebank, i.e. training and analyzing iteratively in order to enhance speed and quality of the human syntactic analysis. By conducting an extensive and heuristically guided search in the vast grid of options (parser, embedding, configuration, epochs, batch size, size of training set, annotation scheme, language, evaluation method…), we determine the best performing parser configurations: UDify and Trankit share the podium depending on the size of the training set. We also show how these results are integrated into the annotation tool ArboratorGrew, and we propose some preliminary measures that allow predicting the quality of the parse for a new language.
Origin : Publisher files allowed on an open archive