JMIR Publications

ABSTRACT

Background:

To accurately communicate the severity of self-experienced violence to a person not present at the event is crucial for proper treatment of the victim and a reasonable penalty of the perpetrator. Recent data shows that humans have biases in this communication, where the severity of psychological violence is underestimated and physical violence is overestimated by the persons reading texts compared to the writers experiencing the interpersonal violence. Furthermore, the severity of psychological violence is less accurately communicated than physical violence. Recent advances in computational language models provide opportunities for automated evaluation of the severity of narrations of violence.

Objective:

We investigate whether these biases can be removed with computational algorithms trained to measure the severity of violence.

Methods:

The data analyzed in this study is taken from Sikström et al (2021) and were collected in two phases using the Prolific Academic website for online recruiting. The aim of this study was to investigate whether a computerized language model could remove the biases in communication of severity of violence found in Sikström et al (2021). This was accomplished by first quantifying the text data to word embedding, i.e. a vector describing the meaning of a text, and then using machine learning to map the embedding to a scale of severity of violence.

Results:

The results show that the computational model mitigates the accuracy bias and removes the calibration biases.

Conclusions:

Our results suggest that computational models can be used for debiasing severity evaluations of violence. These findings may have application in legal context, prioritizing of resources in the society and how violent events are presented in media.