Representing Overlaps in Sequence Labeling Tasks with a Novel Tagging Scheme: Bigappy-Unicrossy

Abstract

AbstractMultiword expression (MWE) identification can be handled by using sequence tagging approach accompanied with stochastic models and variants of IOB tagging scheme. In this paper, we introduce a new tagging scheme called bigappy-unicrossy to rise to the challenge of overlapping MWEs. The bigappy-unicrossy tagging scheme is compared with the two other well-known tagging schemes which are IOB2 and gappy 1-level in the verbal multiword expression (VMWE) identification task using bidirectional Long Short-Term Memory model with a Conditional Random Field layer on top (bidirectional LSTM-CRF). Both the bigappy-unicrossy and the gappy 1-level tagging schemes outperform the IOB2 tagging scheme. The bigappy-unicrossy tagging scheme competes with the gappy 1-level tagging scheme. We believe that our tagging scheme will show better performance on corpora with higher frequency of overlapping cases.KeywordsIOB tagging schemeMultiword expressionsGappy 1-level tagging schemeBigappy-unicrossy tagging schemeLong Short-Term Memory