Exploring LSTM vs. BERT for Event Detection in Social Media Posts

In today's digital world, social media and online platforms generate a lot of unstructured text data. Event detection and classification from such data becomes important for understanding and responding to global events. Classifying text into categories like political events, riots, and disasters plays a important role in public safety, disaster response, and media analysis. This work compares the performance of LSTM and BERT, on a event classification task to categorize social media posts into five event categories: terror, political, disaster, riot, and positive. The results indicate that BERT outperforms LSTM in all the evaluating metrics. LSTM generally delivers more balanced but less accurate results, often with a faster processing time. In cases where precision is crucial, models like BERT are preferred, even if they require higher computational resources. This work tells importance of selecting a model that goes along with the specific demands of task complexity and available computational resources.