Performance analysis of different BERT implementation for event burst detection from social media text

Abstract

The language models play very important role in natural language processing (NLP) tasks. To understand natural languages, the learning models are required to be trained on large corpus. This requires a lot of time and computing resources. The detection of information like events, and locations from text is an important NLP task. As events detection is to be done in real-time so that immediate actions can be taken, hence we need efficient decision-making models. The pertained models like bi-directional encoders representation from transformers (BERT) gaining popularity to solve NLP problems. As BERT based models are pre-trained on large language corpus it requires very less time to adapt for domain specific NLP task. Different implementations of BERT have been proposed to enhance efficiency and applicability of the base model. The selection of right implementation is essential for overall performance of NLP based system. This work presents the comparative insights of five widely used BERT implementations named as BERT-base, BERT-large, Distill BERT, Robust BERT approach (RoBERTa-base) and RoBERT-large for event detection from the text extracted from social media streams. The results show that Distill-BERT model outperforms on basis of performance metric like precision, recall, and F1-score while the fastest to train also.