How we develop products for automated content within the AI & Automation Lab
Since 2020, we’ve been investigating within the AI & Automation Lab of Bayerischer Rundfunk how artificial intelligence and automation might be utilized in public service broadcasting, working on the interface of journalism, computer science and product development.
We have now learned lots from automating stock market reports, the Corona newsletter and basketball reports, and based on this we’ve developed a tool in type of a chart that is meant to assist us and other newsrooms to record vital steps within the workflow.
It’s an summary that we’re repeatedly developing and that may vary and be supplemented depending on the project.
Our Raci chart relies on the 6-phase product development, which we’ve adapted to our needs. Thus far, the AI & Automation Lab is working inside BR with editorial teams, the search engine marketing/distribution department and software development. The teams and their functions are specified by columns. The rows represent the person project phases and work steps:
Our product ideas to this point have emerged from our work on the interface of editorial and technology. Along with the editorial teams, we have a look at the several areas within the journalistic workflow where AI and automation might be used to support.
Good incubators for the launch of our products were hackathons and the AI Fellowship on the London School of Economics, through which we were capable of construct the primary prototypes in a short while.
Nearly all of our projects to this point fall into the next categories:
- Data often provides the idea for automated content. Once you may have created an automation process, different content might be generated from one data source.
In our automated stock market texts, we use data from the Munich Stock Exchange to generate a every day top gainers/losers report of the Dax corporations. We also use the identical workflow to generate an automatic weekly review specifically on Bavarian company data. - As described in our Data Driven Publishing concept, we consider automation as a way of personalization. In our Remix Regional project, we create automated metadata to regionalize existing radio news and possibly offer it to users of our radio web and app services.
- We’re also working with our colleagues within the investigative teams to develop ideas for AI-assisted research with large amounts of information.
Within the case of the Xinjiang Police Files, for instance, we were capable of provide an open source translation model due to our own AI infrastructure. We translated sensitive text files from Mandarin into English under great time pressure on several machines in parallel. - One other use case is the AI-supported automation of labor routines.
Especially for repetitive tasks, equivalent to analysing hundreds of user comments, an AI model is usually a very helpful assistance system for the editorial team.
Before we implement an idea and move on to the subsequent phase, we check along with the and the whether there’s enough interest for the automated content and whether it pays off by way of BR’s public service mission. Also automated content is barely successful if there’s enough demand for it.
Within the exploration phase we condense the thought and create an idea:
- A crucial query initially is: ? For a lot of challenges in a newsroom there are already existing solutions.
- As well as, we evaluate our resources and people of the teams involved: automation at all times has to do with working over existing workflows, for which we want who can design and evaluate a latest system along with their existing jobs.
- The work with the editorial team already starts within the exploration phase: First, we take a to grasp exactly how a product can best be integrated into the every day routine. A have a look at the editorial workflow helps to implement the tool at the suitable time and in the popular way (directly within the CMS, via Microsoft Teams, via email).
- Our previous automation formats are based on data that’s converted to text. Subsequently, we research the : Is the info published reliably? And the way quickly? Should the info be published only after agency reports, for instance, the query arises whether the event of automated content pays off.
Once we’ve condensed the concept, we conduct an initial to raised understand whether the planned format is of interest or — within the case of internal automation — would facilitate the work of the editorial team.
With the product definition, the goals and requirements of the product are defined intimately. This includes the functionalities of the product, the goal group, the intended effect and the success criteria.
- On this phase, all teams involved define the concrete and which (Key Performance Indicators) might be used to measure it.
The goals, for instance saved time, have to be defined as a KPI and made measurable. - One other vital step is to define the . Here we’re guided by the agile development (minimum viable product), by which central functions are defined and an prolonged range of functions is described on this basis. (see Development)
Automation at all times has an impact on existing workflows. Subsequently, a is mandatory through the product definition, as that is the one technique to successfully integrate the product.
We start the event of the system by creating an . This helps us to quickly test the of the product and provides all teams involved a primary impression of it.
In accordance with the motto , we test if the purely conceptual idea also passes the primary practical test and meets the fundamental expectations of all colleagues involved. Several teams are already participating within the creation of the MVP:
- For the automation of content, we currently implement systems by which we use different wordings depending on the info basis. This manner we will be certain that that data just isn’t hallucinated (sometimes the case with generative AI) and the editorial team has full control over the style and structure of the text from the start.
- We also try to indicate first results quickly when simplifying work processes: In our project “What’s there, what’s missing”, by which we use an AI model to investigate user comments in response to inquiries to the editorial team, the MVP already showed at an early stage which sort of comments the AI will detect. Nonetheless, AI models need numerous training data, which is why we partly artificially enhance data for the MVP (data augmentation) so as to quickly give a primary impression. In the midst of development, the synthetic data might be replaced by real data.
- Throughout the development of the product, we work agile: alternate with from the editorial team and an prolonged user test.
Before the system is used for the primary time, we arrange a with all colleagues involved and publish an in regards to the latest product.
Within the AI & Automation Lab, we use our own infrastructure for the products we develop.
This permits us to make use of technologies flexibly, but at the identical time we’re chargeable for operation and maintenance. In an effort to save resources, we discontinue products which can be now not needed as early as possible and, even within the case of latest developments, we repeatedly check whether the respective product contributes to BR’s goals.
: https://miro.com/app/board/uXjVPoxqSsU=/#tpicker-content
The chart is meant as a development guide for automation products within the newsroom and might be used and adapted for other projects. Click on the title to repeat the Miro-Board or use the share icon to download it as a pdf.
https://newsproduct.org/product-kit/understanding-product-development-lifecycle