Our submission to the W-NUT Named Entity Recognition in Twitter task closely follows the approach detailed by Cherry and Guo (2015), who use a discriminative, semi-Markov tagger, augmented with multiple word representations. We enhance this approach with updated gazetteers, and with infused phrase em-beddings that have been adapted to better predict the gazetteer membership of each phrase. Our system achieves a typed F1 of 44.7, resulting in a third-place finish, despite training only on the official training set. A post-competition analysis indicates that also training on the provided development data improves our performance to 54.2 F1.
展开▼
