Government looks into digitising cultural collections with AI

The government has published a report looking into the feasibility of digitising cultural collections in England with emerging technologies such as artificial intelligence (AI).

The report, commissioned by the Department for Digital, Culture, Media and Sport (DCMS), was carried out by Collections Trust. The investigation is also intended to complement the ongoing digitisation efforts of the National Archives Taskforce.

Technical aspects of the government’s ambitions to access and search all cultural collections are investigated in the report, and a recommended framework for mapping and connecting cultural collections is outlined, along with possibilities and limitations around new tools for making the material searchable across organisations.

The suggested architecture outlined in the report involves bringing data together from a wide range of institutions such as museums, archives and libraries. It would then use tools and services to process, clean and enhance data, which would be made available in various ways for uses that could be limited by licensing restrictions that the contributing institutions would specify.

A prototype based on this approach, with sample data from various sources, was developed with modules aimed to demonstrate the possibilities and flexibility of various AI services in the area of content analysis and enhancement. A processing pipeline was also created to allow the AI enhancement services to be applied to the test data.

According to the report, results from the trial demonstrated that the prototype could provide users with a single point of access to various types of cultural data from many institutions.

The prototype also shows that data from various institutions and under different cataloguing standards could be ingested through various technical means, and institutions would not have to make customisations to their contributions to the platform.

However, the pitfalls of AI services were also demonstrated in the trial and, according to the report, had “mixed success”.

While elements such as places were often recognised and enhanced with more information, there were enough errors in classification and identification to prompt the report to recommend that machine-generated enhancements should be clearly identified as auto-generated and kept separate from original source data.

This situation, the report stated, underscores the need to train AI services with substantial amounts of data relevant to cultural heritage collections, rather than focusing on the inherent shortcomings of the technologies.

A number of risks were also identified in the investigation, ranging from duplication of records to lack of long-term commitment to use, maintenance and support of the digitisation approach investigated in the trial.