Applying Existing Named Entity Taggers at BARR IBEREVAL 2017 Task

We present our experiments applying, off-the-shelf, two existing Named Entity Recognition (NER) taggers for the Biomedical Abbreviation Recognition and Resolution (BARR) task at IberEval 2017. The first system is a Perceptron tagger based on sparse, shallow features whereas the second is a bidirectional Long-Short Term Memory neural network with a sequential conditional random layer above it (LSTM-CRF) and initialized with ngram word embeddings. Due to time constraints, we only managed to submit a run from the Perceptron tagger, although in this paper we will also report results with the LSTM-CRF tagger evaluated on the development data. Results show that both Perceptron and LSTMCRF perform reasonably well for SHORT abbreviations whereas the Perceptron model fails to generalize properly for the LONG entity class.