favorproject/staged-kidnapping-nlp-detection

NLP classification of staged versus genuine kidnapping complaint narratives in Nigeria # Detecting Staged and Fraudulent Kidnapping Reports Using Natural Language Processing Code and data for the paper "Detecting Staged and Fraudulent Kidnapping Reports Using Natural Language Processing: Evidence from Nigerian Police Case Narratives" by Ugwu, Akor Peter (Department of Computer Science, Benson Idahosa University, Benin City, Nigeria). ## The problem Staged kidnapping reports are a growing burden on Nigerian law enforcement. Individuals fake their own abduction to extort ransom from their families. Exposed cases in Edo, Ondo, Delta, and Lagos States between 2025 and 2026 each consumed days of investigative effort before the deception collapsed, and police commissioners have warned that false kidnap claims divert vital security resources away from genuine victims. This project trains text classifiers to flag likely staged reports at the point of complaint intake, before investigators commit field resources. ## Key results - Best models: Logistic Regression and Linear SVM. Cross-validation accuracy of 94.2 and 93.3 percent respectively, held-out test accuracy of 88.9 percent, F1-score of 0.895, ROC-AUC of 0.988. - Recall of 0.944 on the staged class: the system caught 17 of 18 staged reports in the test set. - Staged reports are marked by terms describing bound-victim videos, bank transfer demands, online ransom appeals, financial pressure, lone departure under a private pretext, and use of the victim's own telephone line. - Genuine reports are marked by terms describing armed men, gunshots, highway interception, witnesses, multiple victims, and forest movement. - The dominant error is the false staged flag on genuine lone-victim abductions, which fixes the tool's role as a triage aid under human review, not an automatic decision system. ## Files | File | Description | |---|---| | `build_corpus.py` | Builds the corpus of 120 anonymised complaint narratives (60 staged, 60 genuine) modelled on publicly reported Nigerian police cases resolved between …