A three-level classification of French tweets in ecological crises

Highlights：

• Automatic crisis management, using natural language processing techniques, is a quite hot research topic.

• Characterization of crisis-related messages according to three dimensionsrelatedness, intentions to act and urgency.

• The first French manually annotated crisis dataset.

• Experiments with binary classification (useful vs. non useful), three classes (non useful vs. urgent vs. non urgent) and multiclass classifications (i.e., intention to act categories) using traditional feature-based machine learning and new deep learning architectures by developing a domain shift approach over pre-trained word embeddings and incorporating different metadata information in a multi-input architecture.