An adapted version of T. Berndtsson's abstract "Letter Word Text Taxonomy" illustration / CC-BY 4.0
HLT, Best Practice

More HLT FTW: A Quick Recap of the SELMA Project

Open source human language technology SELMA has come to an end. Well, it actually ended in early spring, but there was an extra time with demos/discussions, a very successful official review, and lots of other important things – which is why we're providing a final blog post only now. Here's a recap of the "stream learning for multilingual knowledge transfer" effort from a DW perspective.

Let's start with the basics: SELMA was an HLT project focused on developing AI-powered speech and text technologies to assist media monitors and journalists. Coordinated by DW, it officially ran from January 2021 to March 2024 and also cooperated with other projects like Speech Brain

As a fairly complex and ambitious research and innovation action (RIA), SELMA dealt with a broad range of HLT topics that included benchmarking, diarization, diversity monitoring, speech synthesis, and semi-automated podcasting. For more detailed (academic) info on this, we suggest you head over to the project pages listing the SELMA publications and its code/data output.

Another great achievement of the RIA: It helped refine the plain X media localization platform into a mature product that recently hit the mark of 1000 users. Furthermore, technology developed in SELMA was applied in the monitoring tool Monitio and several DW prototypes.

Here's a video demonstrating DW Speaker, DW Summarizer, and our experimental podcast creator (the latter is explained in detail in this post)

Insights and lessons learned

Regarding the coordination and managing process of SELMA, here's six important things we came to realize:

  • a small consortium can achieve a lot–with the right partners

  • developing and immediately applying/testing software is a lot of fun

  • publishing data for researchers and/or for the public comes with many challenges (e.g. the constant risk of IP infringements)

  • it's a must to have expert ethical board when starting a big data project

  • one should always be prepared for technological disruptions (like the introduction of ChatGPT)

  • direct, positive feedback from end users (e.g. journalists in DW newsrooms) is an amazing thing

What's next?

SELMA may be over, but other HLT projects are underway, and the usage of plain X and Monitio is in full swing (don't hesitate to ask for a demo). So what are we going to do next? Well, there are user comments to structure and exploit, models and modules to update, new or enhanced engines to test, UX and UI concepts to tweak. And of course, we'd love to bring (more) DW prototypes to the production level.

Kay Macquarrie
Alexander Plaum