The BBC has been applying AI algorithms to help unlock its huge content archive, much of which has gone unseen by modern viewers. The technology is accelerating the previously highly labour-intensive process of sifting through over 250,000 TV shows dating back to 1953.
This is far from being an academic exercise for the BBC, given the expectation of a continued real-terms decline in its license fee income (its dominant source of revenue). The immediate priority is to get more content into market, with monetisation likely to come later. The BBC is also pursuing a technology lead in concepts that underpin a new era in content discovery and, potentially, content personalisation.
“Our aim was to unlock the archive for programme makers and that is what we are focusing on now,” said George Wright, Head of Internet Research and Future Services at BBC R&D. Wright was forthright about the achievements so far, not just in searching the archive intelligently for programmes that are likely to appeal to contemporary audiences, but identifying segments within them for novel shows.
“It is a huge breakthrough in terms of unlocking the BBC’s archives for programme makers. As you can imagine, manually looking through hundreds of thousands of items is simply not practical, but using AI we can search programme information much more efficiently. And as you will see in ‘Made by machine: When AI met the archive’, we are also developing techniques to actually search inside the programmes themselves.”
‘Made by machine: When AI met the archive’ is one of a series of experimental programmes the BBC is broadcasting across the nights of September 4-5 on BBC Four to explore themes around AI in media and demonstrate its own progress with AI. The programme collection is themed ‘BBC 4.1 AI TV’. The ‘Made by machine’ show is part-made by AI itself and some of the archive content being shown across the two nights (which BBC Four considers “hidden gems”) was discovered using the new AI techniques.
The intra-content search capability that Wright refers to above applies a combination of AI-enabled techniques to ‘view’ footage and create ‘new’ programme segments. Firstly, the algorithms learn to identify what a scene comprises by recognising details such as type of landscape, objects within it, whether people are present and, if so, what they are wearing. Secondly, the system scans the subtitles of archive programmes and looks for connections between words, topics and themes as it pieces footage together.
Thirdly, the system analyses whether the action is high-energy, with a lot of activity on screen, or low-energy, with not much happening (visually, at least). Then the system tries to create a compilation, moving between high and low energy scenes. Finally, combining all it has learnt, the AI-based solution creates a new piece of content.
The BBC views intra-content search as a path towards object-based broadcasting, where programmes can be assembled automatically at the scene level from multiple sources. You can read a separate story about object-based broadcasting here.
The BBC clearly has an eye on the future with this work but the archive selection it is creating will be exploited by programme makers almost immediately.
So far, content selection does not take account of individual user preferences. At this stage the algorithms operate at a generic level, determining (in the immediate application) what BBC Four audiences as a whole might like from the entire BBC archive, using the channel’s previous schedules and programme attributes as the guide. It then ranks programmes in order of relevance and hands them over to programme makers for the final decision.
Effectively, the AI system relies on the acceptance that past BBC 4 schedules are an accurate guide to what audiences want today, so humans still retain editorial control. “It is one way of giving the AI a sense of what makes a good BBC Four programme, giving our expert schedulers a selection of programmes they might never have found,” says Wright. “Then they can make the final decision on what audiences might like.”
The BBC is not alone in employing AI to help make novel programmes from existing scenes. At the Wimbledon tennis championships in July, IBM applied its Watson AI platform to create highlights packages by analysing scenes on the basis of facial and emotion recognition, taking account of movements such as fist pumping, arm raising and hand shaking, as well as audio cues like shouting at the umpire. It also took account of player motion and match statistics.
This made it possible to create highlights packages that were available minutes after the day’s live events concluded, instead of hours. Once again, human schedulers had the final say. It meant producers could distribute many more highlights packages within the precious time window when they are valuable. They achieved a scale of output that would not be possible, or at least not be affordable, with human editors alone.