A Knowledge-Enhanced Neural Sequence Labelling Model for Named Entity Recognition on Noisy User-Generated Contents
Pengcheng Yang, Jun Zhang · 2020
In recent years, named entity recognition (NER) on user-generated content (UGC) has attracted extensive attention due to its potential widespread application prospects, such as breaking news detection and products recommendation. In this paper, we present a novel knowledge-enhanced neural sequence labelling model (KNSLM), which incorporates external entity knowledge to augment the understanding of UGC with character, word and external entity knowledge. The proposed KNSLM model addresses both text denormalization and data sparsity issues, which cannot be solved directly by current sequence labelling methods. We empirically show that this KNSLM model relieves such issues, which obtaining excellent results in UGC dataset on User-generated Text (WNUT-2017). The experimental results show this KNSLM model achieves entity and surface F1 score 55.09% and 53.12%, respectively.