Neural Modeling for Named Entities and Morphology (NEMO2)

Named Entity Recognition (NER) is a fundamental NLP task, commonly formulated as classification over sequence of tokens. Morphologically-Rich Languages (MRLs) pose challenge to this basic formulation, the boundaries Entities do not necessarily coincide with token boundaries, rather, they respect morphological boundaries. To address NER in MRLs we then need answer two questions, namely, what are units be labeled, and how can these detected classified realistic settings, i.e., where no gold morphology available. We empirically investigate questions on novel benchmark, parallel tokenlevel morpheme-level annotations, which develop for Modern Hebrew, morphologically rich-and-ambiguous language. Our results show that explicitly modeling leads improved performance, hybrid architecture, precedes prunes decomposition, greatly outperforms standard pipeline, decomposition strictly NER, setting new performance bar both Hebrew tasks.