Is a Pretrained Model the Answer to Situational Awareness Detection on Social Media?

Abstract

Social media can be valuable for extracting information about an event or incident on the ground. However, the vast amount of content shared, and the linguistic variants of languages used on social media make it challenging to identify important situational awareness content to aid in decision-making for first responders. In this study, we assess whether pretrained models can be used to address the aforementioned challenges on social media. Various pretrained models, including static word embedding (such as Word2Vec and GloVe) and contextualized word embedding (such as DistilBERT) are studied in detail. According to our findings, a vanilla DistilBERT pretrained language model is insufficient to identify situation awareness information. Fine-tuning by using datasets of various event types and vocabulary extension is essential to adapt a DistilBERT model for real-world situational awareness detection.