T5-Based Named Entity Recognition for Social Media: A Case Study for Location Extraction

Abstract

Social media platforms have emerged as invaluable sources of real-time information, particularly during emergencies and events. Efficient and accurate extraction of location names from this unstructured text data can significantly enhance response efforts. This study investigates the performance of two models, T5 and SpaCy, for extracting location names from 5554 Indonesian-language tweets related to traffic conditions. The T5 model, leveraging its Transformer architecture and extensive pre-training, achieved a significantly higher accuracy of 95% in training and 93% in testing compared to SpaCy's 45% and 41 % respectively. This disparity highlights T5's superior ability to handle complex language patterns and indirect location references often found in social media text. Conversely, SpaCy's reliance on Convolutional Neural Networks (CNNs) poses limitations in effectively processing diverse location representations and non-local text patterns. The results demonstrate the potential of T5 as a powerful tool for location extraction in social media analysis, with significant implications for improving disaster response and public safety efforts.