Published March 7, 2022 | Version v1
Conference paper Open

Transformer-based Methods with #Entities for Detecting Emergency Events on Social Media

  • 1. University of La Rochelle
  • 2. Sorbonne University, STIH/CERES

Description

This paper summarizes the participation of the L3i laboratory of the University of La Rochelle in the TREC Incident Streams 2021. This track aimed at identifying critical information present in social media by categorizing and prioritizing tweets in disaster situation to assist emergency service operators. For both classifying tweets by information type, and ranking tweets by criticality, we proposed a multitask and multilabel learning approach based on representing the tweet text and the event types with pre-trained language models, and by highlighting entities and hashtags. We also experimented with bag of words representation and classical machine learning methods for the prioritization task. We conclude that our multitask approach, while it can take advantage from both tasks, achieved the best performance in comparison with different proposed ensembles. Our submissions obtained top performance for the prioritization task, and higher than the median for the information type classification task.

Files

TREC_2021___Incident_Track___5_November___no_limit___Transformer_based_Methods_for_Detecting_Emergency_Events_on_Social_Media.pdf

Additional details

Funding

NewsEye – NewsEye: A Digital Investigator for Historical Newspapers 770299
European Commission