Vicomtech at DA-VINCIS: Detection of Aggressive and Violent Incidents from Social Media in Spanish
Authors: Naiara Perez Miguel
Date: 20.09.2022
Abstract
This paper describes the participation of the Vicomtech NLP team in the DA-VINCIS shared task. This shared task is focused on mentions of violent events in Spanish tweets, and proposes two subtasks: first, detecting whether a violent incident is mentioned in a tweet; and, second, determining which type of violent event is being mentioned. We participated in this shared task with multiple systems built on Transformer-based models, which we fine-tuned on different versions of the provided data. Among others, we explored the impact of automatic data augmentation and relabelling. Further, we tested masking keywords during training as a means to avoid the models from overfitting these recurrent expressions. Our systems ranked in 2nd place in both tasks, with F1-scores of 77.32 and 52.86 respectively.
BIB_text
title = {Vicomtech at DA-VINCIS: Detection of Aggressive and Violent Incidents from Social Media in Spanish},
keywds = {
Deep Learning, Transformers, Text Classification, Spanish, Online Social Networks
}
abstract = {
This paper describes the participation of the Vicomtech NLP team in the DA-VINCIS shared task. This shared task is focused on mentions of violent events in Spanish tweets, and proposes two subtasks: first, detecting whether a violent incident is mentioned in a tweet; and, second, determining which type of violent event is being mentioned. We participated in this shared task with multiple systems built on Transformer-based models, which we fine-tuned on different versions of the provided data. Among others, we explored the impact of automatic data augmentation and relabelling. Further, we tested masking keywords during training as a means to avoid the models from overfitting these recurrent expressions. Our systems ranked in 2nd place in both tasks, with F1-scores of 77.32 and 52.86 respectively.
}
date = {2022-09-20},
}
