How can you depend on machines to do the heavy lifting when you also need a lightweight approach?

The Court Administration of Latvia process over 30,000 case documents each year. Legal documents regarding individual court disputes are anonymized and then published for reference by the public. This was a time-consuming manual process. Court secretaries would draft documents and upload the files. Then someone else had to review each one to anonymize key information before publication.

Data showed it took about eight seconds to anonymize a single sentence. So, the average document took between 15-20 minutes to process manually. This equated to approximately 1000 work days per year. The Court Administration of Latvia not only wanted to make this anonymization process faster and free up people’s time. They also wanted to maintain the reliability of the information given to the public.

However, the anonymization of sensitive data is more complex than simply replacing text. In a typical court dispute, there are multiple elements involved. Each of these – like names, license plate numbers, addresses and account numbers – requires a consistent and contextual identifier. Meaning someone identified as Person A or an account labelled Account 1 must remain as Person A or Account 1 every time they are referenced in the published court document. The Court Administration of Latvia wanted to make this a lighter touch, streamlined approach that would improve productivity.

One potential solution would have been to use standard machine learning. To label the individual identifiers. To establish correlations between information within the documents. And to apply different rules to sets of information. Yet machine learning on its own is typically used to analyze huge amounts of general data. In this case, the Court Administration of Latvia needed to apply anonymization principles based on legal rules to small amounts of high-quality data that would change from case to case.

That’s why we proposed incorporating machine teaching as well.

The Court Administration of Latvia needed a faster and future-proofed approach to anonymizing court documents for publication. By optimizing the entire anonymization process, we improved productivity among employees. We also created an intuitive system that the courts could continue to use long after our engagement was over.

Machine training provided the highly efficient approach the Court Administration of Latvia required. Through containerization, we created a lightweight and scalable application that could be used on any IT infrastructure. And while much of the heavy lifting is now performed by the customized machine learning and machine teaching application, the new process still involves human intervention so there is no loss of control or quality assurance.

This engagement involved collaboration between our team and the court’s administration. With individuals from both organizations focused on labeling and teaching the system to recognize entities.

The final application was custom-made for the Latvian legal system’s rules. The machine learning element recognizes and replaces specific information while preserving contextual meaning. The machine teaching element helps the courts label and anonymize new information.

As part of our commitment to leaving lasting capabilities, we also educated the court administration about language models and the methodology used for machine teaching. The court administration’s role changed from standard users to trainers with a deeper understanding of the technology and how to control the system. Change agents, including those from our team, played a pivotal role. By facilitating this knowledge transfer and helping the court administration build the necessary skills to apply machine learning and machine teaching to this critical workflow.