Australian Human Rights Commission adopts machine learning

The Australian Human Rights Commission was drowning in a sea of duplicates, tangled in nested folders and perplexed by lost documents. Funding shortfalls and other challenges saw the Commission unable to implement an EDRMS solution that was viable.

An EDRMS solution implemented for the Agency with just over 150 staff in February 2019 now employs RecordPoint’s Records365 and SharePoint Online along with Machine Learning technologies to classify records without the need for staff input.

Researching options, the Commission headed by Ron McLay, Chief Information Officer and Ryan McConville, Information Manager, incorporated the Department of Finance’s study into the failings of the traditional EDRMS.

In particular, the report suggested that records management should be automated, rather than a being a manual task for public servants. Inspired by the report, the Commission set out to implement a fully automated EDRMS, using artificial intelligence (AI) and machine learning. This would form the basis of RADICAL (Record And Document Innovation & Capture - Artificial Learning).

The Commission avoided customization and add-ons for RecordPoint and SharePoint Online, focusing on configuration instead. A common problem agencies have experienced is in the customization and use of third-party add-ins to suit existing or outdated business processes. This often resulted in systems that were difficult to use, inefficient and unreliable, and hard to upgrade with user uptake suffering accordingly.

Harnessing the native functionality of RecordPoint and SharePoint translated to improved business processes. The Commission also incorporated simple navigation for easy browsing of records, supported by a powerful search feature in Records365.

Before RADICAL, the Commission managed its corporate records in paper files and electronic file shares. The paper file was considered the primary file, while electronic copies of those files were kept for ease of reference and sharing.

The process of creating and sentencing paper files was time consuming and relied on staff members with limited experience, and often no interest in records management, to make accurate decisions about the retention and disposal of valuable corporate records.

The Commission’s approach was ‘configuration over customization’ as recommended by the DTA, focused on human-centered design. Staff were consulted extensively on current needs and pain points. When possible, native Records365 and SharePoint functionality was preserved, limiting the need for end user training and burdensome change management.

Records classification involves categorizing records by function and activity as set out in the Administrative Functions Disposal Authority (AFDA Express).

Traditionally, the classification process has been performed manually by records officers. The manual element of classification can be time consuming, can lead to inaccuracy and can be disruptive to staff. Previous methodologies to automate records classification uses rules trees that classify records based on their metadata and saved location. However, rules trees need to  be built and maintained by experienced records officers and rely on end users to apply accurate metadata and save to specific locations.

Leveraging AI in this process solves many of these problems by combining a minimal rules tree with a machine learning model. If a record cannot be categorized by a rule, the machine learning model classifies the record based on its contents. This system eliminates the need to maintain complex rules trees, the reliance on metadata and record location.

The RADICAL project team worked with RecordPoint’s AI developers to create a statistical model that can classify records against AFDA Express and the Commission’s agency-specific records disposal authority.

The statistical model is developed by taking a set of records that have been manually classified and applying Natural Language Processing techniques to normalize the document content into vectors. The model is then trained using algorithms.

After an initial training period, the RADICAL statistical model can categorize individual records with an accuracy of 80%. The Commission expects this accuracy will increase over time. RADICAL also re-categorizes records each time they are edited, ensuring the classification is always current.

Although the machine learning model will initially work in conjunction with a rules tree, as the accuracy of the model increases the rules will be gradually removed and the Commission will rely solely on machine learning to manage their corporate records.

A ‘greenfield’ implementation was a strategic advantage to the design and change management process. RADICAL presented an opportunity to use an advanced AI-driven platform to deliver an easy, modern and powerful platform without staff preconceptions and complex data migration.

Compiling a training dataset for the machine learning model proved to be another challenge. In order to provide a learning dataset for Record365’s learning algorithm, a minimum of 1,000 electronic records was needed, individually classified against each class in AFDA Express and the Commission’s agency specific authority.

As the Commission’s primary files are paper, no electronic records had been previously classified. While a problem at first, this was also an advantage as the Commission could ensure the dataset had been classified accurately and consistently, in turn increasing the quality and accuracy of their model.

Another challenge was that the model accurately classified records according to their content but cannot yet factor in the context in which they are created. For example, a legal advice about a procurement process will be classified under the AFDA Procurement function rather than Legal Services, as its contents primarily concern procurement. Additionally, where the model identifies the correct function, identifying the correct disposal class often requires a subjective assessment beyond the capability of the Commission’s current model.

However, in working with the RecordPoint team, the Commission was able to address these scenarios and have several promising options to test.

The Commission is also looking for opportunities to further leverage Microsoft’s machine learning technology to improve the accessibility of its records. For example, the Commission has been testing machine learning services to transcribe audio-visual records to text, and to translate some publications to Easy English.