Using machine learning to assess rape reports: “Signaling” words about victims' credibility that predict investigative and prosecutorial outcomes

Purpose: The second of two articles from a larger study whose aim was to teach a computer to detect innuendo (or signaling) about a victim's credibility in incident reports of rape. This study explored if the words expressed or not expressed, intentionally or not, influenced case progression and outcomes. Methods: We employed machine learning, specifically text classification, to identify predictive phrases. Sample consisted of 5638 incident reports of rape with a sexual assault kit from a U.S., urban Midwestern jurisdiction. Results: As hypothesized, predictive phrases were different in cases that stalled earlier. Cases not recommended for prosecution lacked detail and more heavily mentioned: (in)actions of victims, actions that stall cases, and procedural words. Reports where victims were not believed or unfounded were similarly vague, procedural, and terse. Cases recommended for prosecution predictively mentioned suspects and the rape statute. Conclusions: We taught a computer to detect signaling via words that were predictive of case progression and outcomes. Negative signals about a victim's credibility often presented as unqualified statements of “fact” or observations or procedural words, indicating a focus on the process vs. victim or suspect. Implications and recommendations are provided, including how unqualified doubts about victims' credibility have substantial public safety consequences.