Exploring Options for Fast Domain Adaptation of Dependency Parsers
Viktor Pekar, Juntao Yu, Mohab Elkaref, Bernd Bohnet · 2014
The paper explores different domain-independent techniques to adapt a dependency parser trained on a general-language corpus to parse web texts (online reviews, newsgroup posts, we-blogs): co-training, word clusters, and a crowd-sourced dictionary. We examine the relative utility of these techniques as well as different ways to put them together to achieve maximum parsing accuracy. While we find that co-training and word clusters produce the most promising results, there is little additive improvement when combining the two techniques, which suggests that in the absence of large grammatical discrepancies between the training and test domains, they address largely the same problem, that of unknown vocabulary, with word clusters being a somewhat more effective solution for it. Our highest results were achieved by a combination of word clusters and co-training, significantly improving on the baseline, by up to 1.67%. Eval-uation of the best configurations on the SANCL-2012 test data (Petrov and McDonald, 2012) showed that they outperform all the shared task submissions that used a single parser to parse test data, averaging the results across all the test sets. 1