![[Adobe Stock]](https://sp-ao.shortpixel.ai/client/to_webp,q_glossy,ret_img,w_750,h_375/https://www.psypost.org/wp-content/uploads/2023/12/artificial-intelligence-computer-chip-750x375.jpg)
In a new study published in Scientific Reports, researchers have utilized artificial intelligence to distinguish between lethal and non-lethal violence against women in intimate relationships. This innovative approach has shown promising results in predicting the likelihood of severe outcomes in such relationships.
For years, the legal and criminology fields have grappled with the challenge of understanding and predicting intimate partner violence. Traditional methods have relied heavily on the analysis of criminal records and behavioral patterns, but these have often fallen short in accurately predicting the escalation from non-lethal to lethal violence. Recognizing the need for more sophisticated tools, researchers turned to artificial intelligence (AI) for solutions. AI, with its ability to process vast amounts of data and identify complex patterns, presented a unique opportunity to delve deeper into the nuances of intimate partner violence.
“Our scientific paper addresses the prediction of intimate partner femicide based on data available in legal documents and using Artificial Intelligence (AI) for three main reasons,” explained study author Esperanza Garcia-Vergara, a PhD student at Loyola University in Seville. “First, intimate partner femicide is considered the leading cause of violent death in women. Despite scientific advances in the identification of intimate partner femicide risk factors and risk assessment instruments, intimate partner femicide still occurs, highlighting the need for further research into this phenomenon.”
“Second, we focus on legal documents as an officially corroborated source for legal aspects as well as criminal behaviors, crucial in research on intimate partner femicide where the data cannot be extracted from victims’ interviews. Third, we use AI to efficiently analyse large data in legal documents, identifying patterns that could contribute to early detection and implementation of effective preventive interventions.”
The study involved analyzing a total of 491 legal cases from Spain, focusing on instances of violence against women by their male partners. The cases were meticulously selected from a larger pool of 8,481 cases identified from the Vlex legal database, a comprehensive source of legal documents. Only final court judgments were considered for the study, excluding provisional resolutions to focus on confirmed and concluded cases.
The research team employed natural language processing (NLP) techniques, a subset of AI that focuses on interpreting human language, to sift through and extract relevant information from these documents. This technology enabled the researchers to efficiently process the textual data, identifying and categorizing pertinent information related to cases of intimate partner violence.
For a thorough analysis, 33 independent variables were identified and grouped into three categories. The first group comprised past criminal behaviors and sanctions, detailing the perpetrators’ general and intimate partner violence-specific criminal history and the nature and duration of sanctions imposed. The second group encompassed environmental and situational factors, such as the location and timing of the incidents, presence of people, and circumstances like disputes. The final group included variables directly relating to the violence characteristics, like its frequency, escalation, type (physical or psychological), and the victims’ coping strategies.
These variables were then analyzed using machine learning algorithms to identify specific elements in legal texts associated with intimate partner femicide.
The research revealed that a more accurate prediction could be achieved by considering a broad range of variables. Initially, when only criminal history and sanctions were analyzed, the accuracy of predictions was limited. However, as the researchers included more variables related to the circumstances and nature of the violence, their ability to distinguish between lethal and non-lethal cases improved markedly.
Specifically, when the analysis focused solely on past criminal behaviors and sanctions, the IBk algorithm led the way in detecting non-lethal cases with a detection rate of 54.27%. However, this performance was surpassed when environmental and situational factors were included, with the RandomForest algorithm reaching a detection rate of 74.16% for non-lethal cases. Remarkably, the inclusion of an even broader set of variables pushed the RandomForest’s detection rate to an impressive 87.04% for non-lethal violence.
In the realm of lethal violence detection, the early stages of the study saw the RandomForest algorithm achieving a 68.49% detection rate. This was eclipsed in later analyses by the BayesNet algorithm, which recorded a 70.56% detection rate after the addition of environmental and situational variables, and an even higher 82.36% with all variables considered.
“In terms of the relevance of our study for the general public, we would say that it provides an optimistic perspective on intimate partner femicide prediction and prevention. On the one hand, our study underlines that information present in legal documents is crucial for the identification of intimate partner femicide cases. This finding underlines the importance of approaching intimate partner femicide detection also from legal context, highlighting the need to train legal professionals in the identification of factors associated with intimate partner femicide and thus contribute to the detection and prevention of the crime.”
“In addition, although our study focuses on intimate partner femicide, it also highlights the potential of applying the same innovative approach to other crime typologies. On the other hand, our study emphasizes the importance of considering multiple factors to enhance the detection of intimate partner femicide. The findings indicate that by taking into account a wide range of diverse factors (such as past criminal behaviors, sanctions, characteristics of violence, as well as environment and situation in which crime occurs) it is possible to identify a greater number of intimate partner femicide cases.”
While the study marks a significant advance in using AI for crime prediction, it is not without limitations. Firstly, the findings are based on Spanish legal cases, which may limit their applicability in different cultural or legal contexts. Moreover, the study’s focus on past incidents means that it could not track the long-term progression of non-lethal violence cases, some of which might escalate into lethal violence. Future research could address this gap by conducting longitudinal studies that follow cases over time, providing valuable insights into how intimate partner violence evolves.
“There are important elements that still require attention by the scientific community, and I would like to highlight one of them. Nowadays there are some cases of violence against women by intimate partners considered low risk for intimate partner femicide, but result in intimate partner femicide. What impedes the correct detection of the real risk in these cases? What differentiates these cases from correctly predicted cases? Further study of intimate partner femicide is needed, with special focus on this issue.”
“To address this, it is essential to adopt a comprehensive approach to the study of intimate partner femicide, recognizing the existence of different profiles within this phenomenon. Each profile may be associated with specific factors influencing intimate partner femicide. The variability in circumstances and characteristics of these situations requires a thorough exploration and understanding of the different elements that contribute to intimate partner femicide. By doing so, we can enhance our prediction and prevention strategies across a broader spectrum of cases.”
The study, “Artificial intelligence extracts key insights from legal documents to predict intimate partner femicide“, was authored by Esperanza Garcia-Vergara, Nerea Almeda, Francisco Fernández-Navarro, and David Becerra-Alonso.
