Methodology

The monitor uses data from Twitter (currenlty, X), a social network that used to provide free access to all tweets made by public users through an API. Only anonymized public data was accesed (private profiles are not available), so no personal information was disclosed in the making of the monitor. Data was collected and stored in real time every hour for the period 1 March 2022-14 March 2023.

The monitor took a sample of 180 active Twitter accounts of women with leadership in opinion and public visibility, and more than 3,000 followers:
•    69 journalists
•    41 politicians
•    27 artists
•    27 communicators 
•    16 social leaders

For the automatic and scaled detection of digital violence in networks, a neural network model (machine learning) was trained using local data and annotators. 

In particular, a model called BETO was used, a BERT-type transformer trained by a team from the Computer Science Department of the University of Chile on a large corpus of Spanish texts (*). The BETO model was trained with 7,697 tweets classified in digital violence and digital non-violence, annotated by 6 local annotators. Thus, the model learned to classify new, unpublished and previously unseen tweets. After six successive rounds of annotation and training, the final model obtained an accuracy of 79.7% on the test sample, and a macro F1 of 76%. 

Once the violent tweets were detected, they were classified into different categories using a list of 350 insults. The list of insults was built from one used in the Argentitian Project (**) and updated based on the inspection of  9,000 noted tweets of Uruguayan origin.