Natural Language Model for Automatic Identification of Intimate Partner Violence Reports from Twitter [article]

Mohammed Ali Al-Garadi, Sangmi Kim, Yuting Guo, Elise Warren, Yuan-Chi Yang, Sahithi Lakamana, Abeed Sarker
2021 medRxiv   pre-print
Intimate partner violence (IPV) is a preventable public health issue that affects millions of people worldwide. Approximately one in four women are estimated to be or have been victims of severe violence at some point in their lives, irrespective of their age, ethnicity, and economic status. Victims often report IPV experiences on social media, and automatic detection of such reports via machine learning may enable the proactive and targeted distribution of support and/or interventions for
more » ... in need. Methods We collected posts from Twitter using a list of keywords related to IPV. We manually reviewed subsets of retrieved posts, and prepared annotation guidelines to categorize tweets into IPV-report or non-IPV-report. We manually annotated a random subset of the collected tweets according to the guidelines, and used them to train and evaluate multiple supervised classification models. For the best classification strategy, we examined the model errors, bias, and trustworthiness through manual and automated content analysis. Results We annotated a total of 6,348 tweets, with inter-annotator agreement (IAA) of 0.86 (Cohen's kappa) among 1,834 double-annotated tweets. The dataset had substantial class imbalance, with only 668 (~11%) tweets representing IPV-reports. The RoBERTa model achieved the best classification performance (accuracy: 95%; IPV-report F1-score 0.76; non-IPV-report F1-score 0.97). Content analysis of the tweets revealed that the RoBERTa model sometimes misclassified as it focused on IPV-irrelevant words or symbols during decision making. Classification outcome and word importance analyses showed that our developed model is not biased toward gender or ethnicity while making classification decisions. Conclusion Our study developed an effective NLP model to identify IPV-reporting tweets automatically and in real time. The developed model can be an essential component for providing proactive social media based intervention and support for victims. It may also be used for population-level surveillance and conducting large-scale cohort studies.
doi:10.1101/2021.11.24.21266793 fatcat:b26gwkoirfclxetkkz37asj2py