Automatic Detection of Levels of Intimate Partner Violence Against Women with Natural Language Processing Using Machine Learning and Deep Learning Techniques

Abstract

Violence against women continues to claim thousands of lives worldwide each year. The COVID-19 pandemic only aggravated the problem by confining many victims together with their aggressors. When a woman experiences this situation, she usually falls into denial, justifies the aggressive behavior of her partner, or even blames herself for provoking it. The sooner she realizes that she is experiencing intimate partner violence, she can act and prevent her advance in the violence cycle (from psychological violence to physical violence, which could lead to femicide). The work proposes a classifier artificial intelligence model to detect levels of psychological violence against women in written virtual expressions (messages ‘from him to her’ in a couple) to ‘alert her’ about the risk that she runs in that relationship. 5250 records in Spanish were extracted with 4 techniques from 6 different sources. Definition of 5 intimate partner psychological violence levels (0-Low Risk, 1-Emotional Blackmail, 2-Jealousy/Justification, 3-Insults/Humiliations, and 4-Threats/Possessiveness) and the data labeling were supervised by a psychologist expert on the problem. Techniques TF-IDF and Word2Vec were used to get the vectors and were tested five Machine Learning algorithms (SVM, MLP, Random Forest, Logistic Regression, and Naive Bayes) with various combinations of parameters. Too were tested pad sequences with LSTM and Bidirectional LSTM. The best result was 93.45% accuracy and 0.2476 categorical cross-entropy loss, obtained with extensive preprocessing, pad sequences, and LSTM. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.