Disaster-Twitter-XLM-RoBERTa-AL

This is a multilingual Twitter-XLM-RoBERTa-base model fine-tuned for the identification of disaster-related tweets. It was trained using a two-step procedure. First, we fine-tuned the model with 179,391 labelled tweets from CrisisLex in English, Spanish, German, French and Italian. Subsequently, the model was fine-tuned further using data from the 2021 Ahr Valley flood in Germany and the 2023 Chile forest fires using a greedy coreset active learning approach.

• Paper: Active Learning for Identifying Disaster-Related Tweets: A Comparison with Keyword Filtering and Generic Fine-Tuning

Labels

The model classifies short texts using either one of the following two labels:

• LABEL_0: NOT disaster-related
• LABEL_1: Disaster-related

Example Pipeline

from transformers import pipeline
MODEL_NAME = 'hannybal/disaster-twitter-xlm-roberta-al'
classifier = pipeline('text-classification', model=MODEL_NAME, tokenizer='cardiffnlp/twitter-xlm-roberta-base')
classifier('I can see fire and smoke from the nearby fire!')

Output:

[{'label': 'LABEL_0', 'score': 0.9967854022979736}]

Full Classification Example

from transformers import AutoModelForSequenceClassification
from transformers import AutoTokenizer, AutoConfig
import numpy as np
from scipy.special import softmax

def preprocess(text: str) -> str:
    """Pre-process texts by replacing usernames and links with placeholders.
    """
    new_text: list[str] = []
    for t in text.split(" "):
        t: str = '@user' if t.startswith('@') and len(t) > 1 else t
        t = 'http' if t.startswith('http') else t
        new_text.append(t)
    return " ".join(new_text)

MODEL_NAME = 'hannybal/disaster-twitter-xlm-roberta-al'

tokenizer = AutoTokenizer.from_pretrained('cardiffnlp/twitter-xlm-roberta-base')
config = AutoConfig.from_pretrained(MODEL_NAME)

# example classification
text = "Das ist alles, was von meinem Keller noch übrig ist... #flood #ahr @ Bad Neuenahr-Ahrweiler https://t.co/C68fBaKZWR"
text = preprocess(text)
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
scores = output[0][0].detach().numpy()
scores = softmax(scores)

# print labels and their respective scores
ranking = np.argsort(scores)
ranking = ranking[::-1]
for i in range(scores.shape[0]):
    l = config.id2label[ranking[i]]
    s = scores[ranking[i]]
    print(f"{i+1}) {l} {np.round(float(s), 4)}")

Output:

1) LABEL_1 0.9999
2) LABEL_0 0.0001

Reference

@inproceedings{Hanny.2024a,
  title = {Active {{Learning}} for~{{Identifying Disaster-Related Tweets}}: {{A Comparison}} with~{{Keyword Filtering}} and~{{Generic Fine-Tuning}}},
  shorttitle = {Active {{Learning}} for~{{Identifying Disaster-Related Tweets}}},
  booktitle = {Intelligent {{Systems}} and {{Applications}}},
  author = {Hanny, David and Schmidt, Sebastian and Resch, Bernd},
  editor = {Arai, Kohei},
  year = {2024},
  pages = {126--142},
  publisher = {Springer Nature Switzerland},
  address = {Cham},
  doi = {10.1007/978-3-031-66428-1_8},
  isbn = {978-3-031-66428-1},
  langid = {english}
}

Acknowledgements

This work has received funding from the European Commission - European Union under HORIZON EUROPE (HORIZON Research and Innovation Actions) as part of the TEMA project (grant agreement 101093003; HORIZON-CL4-2022-DATA-01-01). This work has also received funding from the Austrian Federal Ministry for Climate Action, Environment, Energy, Mobility, Innovation and Technology (BMK) project GeoSHARING (Grant Number 878652).