Innovations in Natural Language Processing
In a world where every part of society and almost every human experience has been impacted by digital tools and services, so too has policing seen a huge shift in the last ten years.
Now the pace is accelerating. With the advent of increasing types of, and use of digital tools, digital channels and generative AI comes the inevitable increase in digital evidence. Digital forensic science is evolving accordingly.
The NPCC Digital Forensic Science Strategy1 outlines that “Digital forensic (DF) science - examining digital evidence to support investigations and prosecutions - was once niche but is now very much mainstream.
Over 90% of all crime is recognised as having a digital element, and society’s accelerating use of technology means the critical role DF science plays will only grow.
We have developed this strategy to address the huge opportunities and corresponding challenges this presents for policing.
The challenge for digital forensics is that it is evidence-focused. Its primary goal is to process and review digital data in a forensically sound manner to produce evidence that can be used in court.
The timeline for retrieving evidence is much slower compared to digital investigation which is intelligence or inquiry-focused and used to generate leads and investigative direction.
To address the challenge of an exponential increase in digital material needing forensic examinations to ensure best evidence, research is being stress-tested by academia and policing to accelerate the use of AI in policing. Specifically, one area of AI, NLP (Natural Language Processing) offers exciting opportunities.
Natural Language Processing in Policing
Natural Language Processing (NLP) represents a transformative branch of artificial intelligence that enables machines to interpret, analyse and generate human language in meaningful ways. At its core, NLP encompasses a comprehensive set of algorithms designed to understand, manage and process human language through various techniques including text classification, clustering, summarisation and machine translation.
Within the context of policing, NLP serves as a critical bridge between the vast volumes of unstructured textual data generated through digital communications and the computational analysis required for effective law enforcement. The technology uses sophisticated methods such as Named Entity Recognition (NER)2 to identify and categorise essential information within crime reports, distinguishing between people, organisations, vehicles and locations - enabling more refined crime grouping and analysis. This proves particularly valuable when searching and interpreting the exponential growth in digital data that is so central to modern criminal investigations.
The practical applications of NLP in policing are extensive and continuously evolving, with law enforcement agencies adopting these technologies across administrative duties, forensic investigations, crime data analysis, and the documentation of criminal activities. NLP tools can quickly analyse vast quantities of text-based data - ranging from interview transcripts and witness statements to online communications and social media posts. According to Europol’s AI and Policing report3, they are particularly useful in time-critical situations such as abductions, hostage scenarios, or investigations into child abuse and exploitation. Automated translation systems can also help with cross-border collaboration by breaking down language barriers, whilst text summarisation capabilities allow investigators to extract crucial details from extensive police reports without sacrificing accuracy.
Advanced NLP applications, including those employed within Europol’s Secure Information Exchange Network Application (SIENA), already provide real-time machine translation capabilities that enhance communication amongst law enforcement agencies from 51 countries. It is also notable that NLP-enabled systems can identify predatory communications, thus helping find internet criminals, and prevent online grooming - representing a shift in how law enforcement agencies tackle cybercrime and protect vulnerable populations.
An LLM (large language model) is a specific type of model within NLP that is trained on massive amounts of text data to understand and generate human-like language. Internationally LLMs are being stress-tested by researchers, as demonstrated by these case studies from Korea and Romania.
LLM-Driven Evidence Analysis
Researchers from the Korean National Police Agency4 have demonstrated the practical application of Large Language Models (LLMs) in real-world digital forensics through a case study involving over 140,000 messages from mobile phones seized during drug-related investigations. The study employed three LLM models - GPT-4o, Gemini 1.5, and Claude 3.5 - to analyse anonymised data from actual crime scenes, focusing on distinguishing between explicitly drug-related communications and metaphorical uses of drug-related terminology.
Through various prompt engineering techniques, the research team achieved remarkable results. GPT-4o demonstrated the highest independent recall and F1 score (a measure of how well the model identifies relevant information without missing key details or including incorrect ones), but a system combining all three models achieved exceptional precision with minimal hallucination.
Critical challenges faced by law enforcement agencies operating under strict time constraints - like the 36-hour window for arrest warrant applications in South Korea – are aided by this framework.
The study represents a significant advancement in applying AI to digital forensics, as it provides law enforcement with the tools needed to rapidly identify criminal evidence whilst maintaining the accuracy and reliability required for legal proceedings.
ROXANNE5 Project: Romania’s Big Horizon Initiative
A good example of innovation in the sphere of digital investigations is ROXANNE (Real time netwOrk, teXt and speaker ANalytics for combating orgaNized crimE) project.
Funded under the European Union’s Horizon 2020 programme, ROXANNE developed the Autocrime platform - an all-in-one investigation system designed to process diverse multimodal data for combating cross-border organised crime. The platform addresses the significant challenge faced by law enforcement agencies where 50 to 80 per cent of investigative workload per case is dedicated solely to extracting preliminary information from raw data, like lawfully intercepted telephone conversations. The complexity increases exponentially in multilingual, cross-border cases. Autocrime integrates state-of-the-art components including speaker identification, automatic speech recognition and named entity detection, capable of transcribing audio in multiple languages, whilst building multiple knowledge graphs that capture the interactions of criminal networks. The AI-powered platform processes data from multiple sources - audio/speech, text, video and non-content metadata. Following hands-on sessions with law enforcement agents who found the platform intuitive and highlighted its innovative multi-technology functionalities, Autocrime has proven its capacity to streamline forensic investigations whilst maintaining compliance with strict data regulations.
In the UK, initially funded by the Police Science, Technology, Analysis and Research (STAR) fund, a year-long study conducted by digital forensic experts from the Forensic Capability Network (FCN) and AI experts from the University of Warwick6 had positive results. Analysing synthetic data using customised natural language processing (NLP) models on secure servers, the project explored the use of AI to identify threatening and abusive language related to violence against women and girls (VAWG). The synthetic dataset was realistic and anonymised – based on real-world VAWG communication scenarios, allowing robust model training without privacy concerns. The platform enabled rapid categorisation of messages, delivered contextual summaries, and offered persona profiling based on conversational data. The persona profiling delivered by the platform can provide behavioural insights into individuals involved in VAWG cases, assisting officers in understanding interaction patterns and motivations. Combing through vast amounts of text-based material, the FCN’s models were able to quickly find words and phrases that could be relevant to a criminal investigation, even when the abuse was subtle or used slang, which is more difficult to detect. In one test, the model took just over one minute to identify three aggressive and emotive phrases within 456 messages, around 21 times faster than an average human investigator7.
Importantly, the project also successfully demonstrated that AI could help protect the privacy of VAWG victims, as it was trained to hone in on relevant messages and ignore material irrelevant to the investigation. The research team suggest the technology could be used in the future as an additional tool to flag potential items of interest to investigators, and AI wouldn’t make investigative decisions itself.
More recently, this project has evolved to include a drugs LLM as well as VAWG and is currently being tested within Stafford Digital Forensics Unit.
Another Police STAR-funded research partnership called DRAGON-Spotter explored the use of NLP to investigate online grooming by analysing messages sent to children.
DRAGON-Spotter sits within a wider programme of innovation called DRAGON8 (Developing Resistance Against Grooming Online), which is run by Swansea University. The DRAGON portfolio of research projects improves practices to keep children safe from technology-assisted sexual exploitation and abuse, including online grooming.
The DRAGON-Spotter tool9 integrates linguistics and artificial intelligence to detect online grooming content by identifying the manipulative language tactics often used by predators - from isolating children emotionally to communicating sexual intent both implicitly and explicitly.
This state-of-the-art tool has been developed through extensive collaboration with child safeguarding practitioners, children, lived experience experts and researchers.
Nina Holland, Head of Digital Forensic and NPCC Lead for Vehicle System Forensics, West Mercia Police, said: “Our collaboration with Swansea University marked a significant breakthrough in applying linguistics to detect and interpret online grooming behaviour. This project was a key milestone in demonstrating how AI can be ethically integrated into digital forensic workflows to support child safeguarding. Swansea University worked closely with us to understand our stringent forensic requirements, to ensure data continuity and alignment with accredited methods.”
The Police STAR funding10 allowed DRAGON-Spotter to be tested on operational datasets, processing over 28,000 messages in 10 hours and offering critical benchmarking data for future scalability. The research project, supported by West Mercia Police (WMP) fostered significant knowledge exchange between the DRAGON-Spotter and WMP teams, enhancing mutual understanding of AI deployment in policing contexts and how to effectively integrate into digital forensic workflows for police.
The DRAGON-Spotter tool looks at the whole conversation between the suspected predator and the child and gives scores to show whether grooming is likely. The tool assists law enforcement in several critical ways: triaging and prioritising cases for processing, reducing analyst workload, strengthening existing cases, and identifying new cases that might otherwise go undetected. The tool has 82% sensitivity for detecting sexual language – that is, for every possible available mention of sexual language, DRAGON-Spotter picks 82% of them. This can support evidence forming significantly. Comprehensive user guides are now available to ensure investigators can interpret results effectively and integrate them into their work. Planning to combine and rollout and embed these innovations in Digital Forensic Units (DFU’s) is the next vital step.
1. National Police Chiefs’ Council, National Digital Forensic Science Strategy, 2020, https://www.npcc.police.uk/SysSiteAssets/media/downloads/publications/publications-log/2020/national-digital-forensic-science-strategy.pdf.
2. Europol, AI and policing – The Benefits and Challenges of Artificial Intelligence for Law Enforcement, 2024, https://www.europol.europa.eu/cms/sites/default/files/documents/AI-and-policing.pdf.
3. Europol, AI and policing – The Benefits and Challenges of Artificial Intelligence for Law Enforcement, 2024, https://www.europol.europa.eu/cms/sites/default/files/documents/AI-and-policing.pdf.
4. Kyung-Jong Kim, Chan-Hwi Lee, So-Eun Bae, Ju-Hyun Choi & Wook Kang, Digital forensics in law enforcement: A case study of LLM-driven evidence analysis, Forensic Science International: Digital Investigation, Vol. 54, 2025, Article 301939, https://www.sciencedirect.com/science/article/pii/S2666281725000782.
5. Srikanth Madikeri et al., Autocrime – open multimodal platform for combating organized crime, Forensic Science International: Digital Investigation, Vol. 54, 2025, Article 301937, https://www.sciencedirect.com/science/article/pii/S2666281725000769.
6. University of Warwick, AI Systems for Understanding and Collecting Evidence in VAWG Investigations, Centre for Operational Police Research, 2024, https://warwick.ac.uk/fac/cross_fac/copr/our-research/understanding-and-collecting-evidence/
7. Dr. Gabriele Pergola, Prof. Arshad Jhumka and University of Warwick, AI Systems for Understanding and Collecting Evidence in VAWG Investigations, Centre for Operational Police Research, 2024, https://warwick.ac.uk/fac/cross_fac/copr/our-research/understanding-and-collecting-evidence/
8. Swansea University, Project DRAGON-S, 2025, https://www.swansea.ac.uk/project-dragon-s/.
9. Swansea University, Project DRAGON-S, 2025, https://www.swansea.ac.uk/project-dragon-s/.
10. Swansea University & West Mercia Police, Revolutionising Digital Forensics Child Sexual Exploitation Workflows Through Linguistics/AI Technology (DRAGON-Spotter), 2025, https://science.police.uk/site/assets/files/3795/revolutionising_digital_forensics
_child_sexual_exploitation_workflows_through_linguistics_ai_technology_dragon-spotter.pdf.
