Gazetteer Enhanced Named Entity Recognition for Musical User-Generated Content

The vast amount of social information distributed over various electronic media channels, namely User-Generated Content (UGC), has been the fastest growing source of information in recent years. Named Entity Recognition (NER) researches on UGC in specific business areas such as music have potential widespread application prospects, such as user intention analysis and products recommendation. Current neural NER models achieve remarkable performance on standard benchmark tasks. It is challenging to deal with noisy, low-context text and complex, emerging entities in real-world setting. A new approach has been proposed to solve the issues. It learns character-level and token-level knowledge simultaneously, and incorporates gazetteers as external knowledge with them which augment the understanding of the music UGC. Experimental results show that our approach achieves an improvement of 15 points in F1 score compared with the modern baseline.