This is the report of Day 2 of the IPTC Autumn 2018 Meeting in Toronto. See the report from Day 1 and the report from Day 3. All the presentations are available to IPTC members in the IPTC Members Only Zone.

Day 2 of the IPTC Autumn 2018 Meeting in Toronto was a deep dive into search and classification. Many of our members are working hard to make their content accessible quickly and easily to their customers, and user expectations are higher than ever, so search is a key part of what they do. 

First up we had Diego Ceccarelli from Bloomberg talking through their search architecture. Users of Bloomberg terminals have very high expectations that they will see stories straight away: They have 16m queries and 2m new stories and news items per day, with requirements for a median query response time of less than 200ms and for new items to be available in search results in less than 100ms. And as Diego says, “with huge flexibility comes huge complexity.” For example, because customers expect to see the freshest content straight away, the system has no caching at all!

To achieve this, the Bloomberg team use Apache Solr – in fact they have 3 members of staff dedicated to working on Solr full-time, and have contributed a huge amount of code back to the project, including their machine-learning-based “learning to rank” module which can be trained to rank a set of search results in a nuanced way. Bloomberg also worked with an agency to develop open source code used to monitor a stream of incoming stories against queries, used for alerting. Other topics Diego raised included clustering of search results, balancing relevance and timeliness, crowdsourcing data to train ranking systems, combining permissions into search results, and more – a great talk!

Our heads already reeling with all the information we learned from Bloomberg, we then heard from another search legend, Boerge Svingen, one of the founders of FAST Search in Norway and now Director of Engineering at the New York Times. He spoke about how NYT re-architected their search platform to be based around Apache Kafka, a “distributed log streaming” platform that keeps a record of every article ever published on the Times (since 1851!) and can replay all of them to feed a new search node in around half an hour. The platform is so successful that it is used to feed the “headless CMS” (see yesterday’s report) based on GraphQL which is used to render pages on nytimes.com for all types of devices. Boerge and his team use Protocol Buffers as their schema to keep everything light and fast. More information in Boerge’s slide deck, available to IPTC members.

Next up was Chad Schorr talking about search at Associated Press, discussing their Elastic implementation on Amazon Web Services. Using a devops approach based on “immutable infrastructure” meant that the architecture is now very solid and well-tested. Chad was very open and spoke about issues and problems AP had while they were implementing the project and we had a great discussion about how other organisations can avoid the same problems.

Then Robert Schmidt-Nia from DPA talked about their implementation of a content repository (in effect another “headless CMS”!) based on serialising NewsML-G2 into JSON using a serverless architecture based on Amazon Lambda functions, AWS S3 for storage, SQS queues and Elasticsearch. Robert told of how the entire project was built in three months with one and a half developers, and ended up with only 500 lines of code! It can now be used to provide services to DPA customers that could not be provided before, including subsets of content based on metadata such as all Olympics content.

Next, Solveig Vikene and Roger Bystrøm from Norway’s news agency NTB spoke about and gave a live demo of their new image archive search product. They demonstrated how photographers can pre-enter metadata so that they can send their photos to the wire a few seconds after taking them on the camera. Some functions like global metadata search and replace and a feature-rich query builder made their system look very impressive.

Veronika Zielinska from Associated Press spoke about AP’s rule-based text classification systems, showing the complexity of auto-tagging content (down to disambiguating between two US Republican Congressmen both called Mike Rogers!) and the subtlety of AP’s terms (distinguishing between “violent crime” events versus the social issue of “domestic violence”) therefore the necessity of manually creating, and maintaining, a rules-based system.

Stuart Myles then took us on a tour through AP’s automated image classification activities, looking at whether commercial tools are yet up to the task of classifying news content, the value of assembling good training sets but the difficulties in doing so, and the benefits of starting with a relatively small taxonomy that is easier for machine learning systems to understand.

Dave Compton talked us through Thomson Reuters Knowledge Items used by the OpenCalais classifier and how they use the PermID system to unify concepts across their databases of people, organisations, financial instruments and much more. Dave described how Knowledge Items are represented as NewsML-G2 Knowledge Items, and are mapped to Media Topics where possible.

On that subject, Jennifer Parrucci of the New York Times, and chair of the IPTC NewsCodes Working Group, gave an update on the latest activities of the group, including the ongoing Media Topic definitions review, adding new Media Topic terms after suggestions by the Swedish media industry, and work with schema.org team on mapping between schema.org and Media Topics terms.

As you can see, it was a very busy day!