Current and future uses of Semantic Web technologies at the BBC

Yves Raimond, BBC R&D IRFS / @moustaki

Talk in two parts

How are Semantic Web technologies currently used on the BBC website?
What are future uses of Semantic Web technologies at the BBC?

Radio since 1922

TV since 1930

On the Web since 1994

Programme support

1000 to 1500 programmes per day across 70 channels

www.bbc.co.uk/programmes

The web site is the API

schema.org

Using Linked Data (external)

Using Linked Data (internal)

Towards a Linked Data Platform

World Cup 2010

Tagging articles

These automated aggregations for various World Cup 2010-related concepts are driven by journalists tagging articles with web identifiers available in a centralised triple store and denoting people, places, events etc. and sourced from multiple data providers (including Linked Open Data sources). The resulting relationships between news articles and these concepts are then pushed to the central triple store. A benefit of using web identifiers as tags is that they are unambigous, and that we can retrieve more information about these tags when needed. For example when news articles are tagged with places URIs, we can easily retrieve the geolocation of these places, enabling us to plot our articles on a map. Tagging articles with URIs enable us to tackle a wide range of unforeseen use-cases.

Dynamic aggregations

BBC London 2012

The same approach was scaled up and used to drive the London 2012 BBC web site, covering around 250 countries, 300 events, 36 sports, 10,000 athletes and 30 venues. This data is sourced from multiple places, both from Linked Open Data sources and commercial data providers. Once all this data is sourced and we have web identifiers for all these things in our centralised triple store, we can start annotating our content with them, in a similar way as done for the World Cup 2010 web site. We used this mechanism to build automated aggregation pages for each of these things - for each country, sport, athletes and so on, by querying the centralised triple store for all BBC items related to them. The BBC London 2012 web site has been hugely successfull, and has proved the feasibility of this approach, heavily based on Semantic Web technologies, using a triple store, queried through SPARQL, to store relationships between our content and domain knowledge.

BBC Ontologies

Linked Data and the BBC archive

The BBC Archive

Cataloguing the archive

In Our Time archive

BBC Four Army collection

Tagging programmes

Machines + Users

As mentioned before, the process of manual tagging is very time-consuming and would take a considerable time to apply to the entire archive. The problem is compounded by the lack of availability of textual metadata for a portion of the archive, and programmes for which we have no data or worse, wrong data - these ones which will always be left out of such brand or topic-specific slices of the archive. In the rest of the talk, we describe an alternative approach to publishing archives online. Instead of focusing on very thin slices of archive content, we purposefully expose large archives using machine-generated Linked Data describing this archive that can be inaccurate, and we try to involve users in helping us correct that automatically generated data. In particular, there is enough Linked Data available online to help us bootstrap an automated tagging process.

The World Service archive

Machine listening

Automated speech recognition

Automated transcripts

Automated tagging

So we now need to isolate these useful clues, and infer the topics of the programme from them. Most of the existing concept tagging tools are designed to work on text that was manually written, and rely on punctuation, capitalisation, etc. We therefore developed our own tool. This tool first locates all these useful clues from the automated transcripts. It then uses the structure of the DBpedia graph to disambiguate and rank keywords spotted throughout the transcript. For example, if a programme mentions Paris and Tour Eiffeil a lot, we will pick Paris in France, as it is closer in the DBpedia graph to the Tour Eiffeil. If a programme mentions Paris and Texas a lot, we will pick Paris in Texas, as it is closer in the DBpedia graph to the Texas resource. For each programme, we get a ranked list of DBpedia tags, describing what the programme is about.

Example results

Processing archives in the cloud

We now have a process which can derive for each programme within our archive a ranked list of descriptive Linked Data URIs, describing what each programme is about. However processing large archives remains a challenge. Tagging a programme can take around 90 minutes for a 60 minute programme on one CPU, meaning it would take more than 4 years to process the entire World Service archive. We therefore developed a framework to process very large archives using Amazon Web Services. A message queue distributes work between a number of independent `workers', hosted on AWS, and picking up new jobs as soon as they're up and ready to process data. Using AWS gives us a potentially infinite number of such workers, meaning that the only bottleneck to process a large archive is the bandwidth between our content servers and Amazon's servers. In our case it took around two weeks to process the whole archive (70,000 programmes), for a very low and predictable cost.

Bootstrapping search and discovery

Noise

Data validation

Speaker segmentation

Crowd-sourcing speaker names

Propagating speaker names

We are also able to recognise speakers across programmes. Therefore the names added by users can be propagated to other programmes in the archive, detected as featuring the same speaker. For that we use an index for speakers based on Locality-Sensitive Hashing, a way to hash high-dimensional vectors with a collision probability that increases if the distance between two of these high-dimensional vectors decreases. Here we see that Nick Caistor's name (which I've entered in the previous slide) has been propagated to this other programme featuring him. As this was automatically inferred (and therefore can be mistaken some time), we provide a simple interface for users to validate this inference. This means that we can further evaluate and refine our speaker identification algorithm.

Evaluating speaker identification

User activity

Emerging shape of the archive

Visualising the archive

Such large archives of content are holding a significant amount of 'institutional memories'. A large number of topics will be covered in some form by the programmes held in this archive. In particular, the archive may hold programmes that could provide context for current news events. For example a 1983 `Medical Programme' episode on techniques for measles immunisation could help put in context a recent epidemic. This visualisation is trying to tackle exactly that: surfacing archive programmes that relate to current news events. The big blue dots are topics that were discussed on BBC News in the last five minutes. The small dots are programmes within the archive that relate to those topics. The red-er a dot is, the more connected it is - so the more likely it is to be relevant to a particular news event. We are going to present this visualisation in more details at ISWC this year.

http://worldservice.prototyping.bbc.co.uk

Code

ClOud Marketplace for Multimedia Analysis

(COMMA)

BBC News Juicer

"NEWS HACK" - 17th & 18th Oct in Shoreditch Town Hall (London)

@BBC_News_Labs

Thank you!

Photo credits:

http://www.flickr.com/photos/andyarmstrong/4402416306/
http://www.flickr.com/photos/nicecupoftea/8579975238/
http://www.flickr.com/photos/11561957@N06/5202870020/
http://www.flickr.com/photos/hubmedia/2141860216/
http://www.flickr.com/photos/allison_mcdonald/7604871594
http://www.flickr.com/photos/aayars/4072755936/