EURO 2025 Leeds
Abstract Submission

2445. From Raw Data To Real Impact: Unlocking Efficiencies for Prisoner Rehabilitation with Data Science, ML and OR

Invited abstract in session WA-34: Advancements of OR-analytics in statistics, machine learning and data science 6 , stream Advancements of OR-analytics in statistics, machine learning and data science.

Wednesday, 8:30-10:00
Room: Michael Sadler LG10

Authors (first author is the speaker)

1. Roger Brown
Ministry of Justice, UK Government

Abstract

In prisons, keyworkers work closely with offenders, aiding rehabilitation & to reduce reoffending once released. Improving quality of keywork (KW) sessions is crucial for these objectives.
National Keywork team manually reads & scores a sample of prisoner KW case notes monthly, aggregating the scores producing prison KW Quality Assessment (QA) metrics. This approach is resource-intensive, difficult to sustain with results only statistically valid for annual accountability. Also, manual scoring introduces subjectivity & potential human error.
We've developed a Machine Learning approach to automate QA of these sessions using innovative approaches & advanced Natural Language Processing (NLP) methods
- Train Gradient Boosting model
- Convert text to Longformer
- Use Named Entity Recognition removing names
- Investigate Claude AI models as alternate approach.
Given the subjective nature of the scoring process, traditional machine learning performance metrics were insufficient. We employed Krippendorff’s alpha, a statistical measure of observer agreement, to evaluate model performance.
We’ve significantly scaled number of KW sessions QAd each month. Before, 2,500 sessions were manually scored monthly; new algorithm scores all 80,000+ sessions in fraction of the time & cost, 24 hrs. They’re providing more complete & timely information for prison teams to act upon. This algorithm helps prison leaders improve KW quality, crucial for rehabilitation & reducing reoffending.

Keywords

Status: accepted

Back to the list of papers