Research project P4S/251/KG4J (Research action P4S)
The National Institute of Criminalistics and Criminology (NICC) plays a scientific role in Belgium’s criminal justice system. First, it offers forensic expertise in domains such as DNA, toxicology and drugs. Second, it engages in criminological research. For instance, it studies recidivism and criminal careers. In doing so, it deals with a wealth of data. Yet a lot of this data is fragmented or underused. Valuable insights remain hidden.
Proposal
The Knowledge Graphs for Justice (KG4J) project aims to address this. How? By building and exploiting data analysis systems using Graph Theory, Knowledge Graphs and Artificial Intelligence (AI) such as Large Language Models. And by defining a policy of data and AI governance.
What is Graph Theory?
Graphs are a natural way to represent connected data. Entities become nodes connected by relationships. Think of a metro map. Each station is a node, each track a relationship. What’s the shortest ride between two stations? What is the most connected station? Graph Theory (GT) answers such questions.
What is a Knowledge Graph?
Adding semantic information — definitions of meaning — turns graphs into Knowledge Graphs (KGs). They are machine readable, interpretable and queryable. And modern AI systems can leverage them to answer questions.
What is a Large Language Model?
A LLM is the engine behind modern conversational AI agents. LLMs have been trained at predicting human discourse based on large quantities of texts. They can talk with us and help perform other tasks. Like programming.
What we will do using this technology
The KG4J project will apply GT, KGs and LLMs to selected use cases. Unlocking the data that is already there, providing new capabilities and results.
Criminological research
The NICC already used a graph to examine recidivism and criminal careers through data sources that were disconnected. The goal is to process the graph to discover new patterns. And evolve it into a KG backed by AI. This allows advanced queries, and easier exploration of criminal trajectories.
Forensic expertise
The NICC usually processes forensic data in a tactical way. With no follow-up analytics to connect the dots. For instance, the DNA databanks match DNA profiles from criminal cases to the ones they hold. And stop there. Turn that data into a graph, and you can reveal co-offending patterns and uncover criminal networks. Connecting evidence supports both ongoing investigations and long-term policy making.
Unstructured data
Transcripts of interviews, expert notes, recordings, the amount of unstructured data is endless. Converting this into a KG is possible through AI. And has become much easier with LLMs. That KG can then answer questions about the interviews, notes, recordings, etc.
Objectives
Build a Knowledge Graph Infrastructure (KGI) with both criminological and forensic science data. Convert a graph into a full KG, running on a server. Queryable with an intuitive interface.
Prove use cases of the KGI. Apply the KGI to compute statistics on recidivism and criminal careers. Use GT to discover novel patterns. Discover usages in forensic domains.
Develop and apply methods to process unstructured data. Build pipelines to extract KGs from text. Perform discourse analysis of interview transcripts or reconstruct criminal trajectories from public records. Equip the KGs systems with LLM’s talkative abilities.
Define a model of data and AI governance. What did we learn? How can the NICC adopt AI in a legal and ethical way aligned with its institutional values? Define principles and rules for data and AI governance.
Conclusion
The KG4J project enables the NICC to deliver stronger forensic and criminological intelligence. It connects the fragmented data. It leverages underused data available today. It builds capacity in new areas through GT, KGs, LLMs. It also develops trustworthy data and AI practices in the service of justice.
