Research on Korean Named Entity Recognition Based on Syllable-Morpheme Fusion

In order to deal with the problems of unclear boundary delimitation and low accuracy rate of Korean named entity recognition, this paper proposes a Korean syllable word fusion named entity recognition model based on Transformer's. Firstly, word embedding is performed for syllables and morphemes separately by the BERT pre-training model.

And secondly, two different vector fusion methods are used to fuse syllable vectors and word vectors, i.e., a simple vector splicing method and a heuristic fusion method that takes into account the connection and difference between the two vectors, and the fused vectors are used as the input of the SWT-NER model. As well the F1 value in the Korean named entity recognition dataset published by KLUE reached 88.03%, which is about 3 to 4 percentage points higher than the single granularity experiment.