A Named Entity Recognition Method for Tibetan based on SWF-Cascade-BiLSTM-CRF Mode

Tibetan named entity recognition is a key task in Tibetan informatization processing. In this paper, we propose the Cascade-BiLSTM-CRF model that fuses Tibetan word vectors and syllable vectors (SWF). The cascade technique can separate entity boundary delineation and entity class judgment for Tibetan named entity recognition to reduce the training time. In this paper, this paper uses a corpus of $5.23\times 10^{4}$ manually annotated sentences and experiments show that the training time of Cascade-BiLSTM-CRF model is reduced by 28% compared with BiLSTM-CRF model. The fusion of word vectors and syllable vectors is performed on this basis to construct the SWF-Cascade-BiLSTM-CRF Mode model, which not only improves the F1 value by 1.7 percentage points, but also reduces the training time by 18%.