Topic modelling enriched LSTM models for the detection of novel and emerging named entities from social media
Patrick Jansson, Shuhua Liu · 2017
Named entity recognition techniques have achieved impressive performance on well-known datasets of canonical texts. However, the detection of emerging and rare entity from user generated noisy text such as tweets, online reviews and forum discussions still remains a challenging task. In this paper, we report our study on the detection of unusual, previously unseen entities from social media. Based on the WNUT2017 shared task datasets, we explore an approach that combines LDA topic modelling with LSTM deep learning on word level and character level embeddings. The LDA topic modelling generates topic representation for each post which is used as a feature for each word in the post. The deep learning components consist of two-layer bidirectional LSTM and a CRF output layer. A large amount of experiments was conducted to understand the effects of the different modelling components and improve system performance. Our latest results reached a best performance level of F1 value at 45.30 on entities and 43.86 on surface forms, a significant improvement over the WNUT2017 best performer with F1 scored at 41.86 on entities and 40.24 on surface forms.