Dear dida-follower,

Unfortunately, we are rarely allowed to report on the content of machine learning projects with our customers. We are all the more pleased that this is different in a project with idealo (website for finding best product offers) for the extraction of information from product descriptions and that we can go into detail about the machine learning models used. Since the neural language model BERT plays a major role in this project, we are also seamlessly continuing the topic of NLP and neural language models from the last newsletter. Click here for the idealo Information Extraction Case Study (in English).

Today, neural language models are often the first choice when it comes to natural language processing. We have summarised why this is so and what advantages they offer in our article "Neural language models: AI for self-service, ticket systems & Co." in the eBook "Artificial Intelligence in Customer Service" together with AI-Spektrum. Of course, the advantages of these models do not only apply in the service sector, but everywhere where search, FAQs, automatic question answering or the like are concerned. Appropriately, here again is the link to our webinar on neural language models.

While Google's BERT language model, which has been tried and tested in practice for a long time, was used in the idealo project, the newer and much larger GPT-3 from OpenAI has been causing quite a stir since last year. A good reason for us to discuss it in more detail in two blog articles. In "GPT-3 and Beyond - Part 1: Basic Recipe" we describe how GPT-3 works, why some consider it dangerous and which similar models you can try for free. In "Part 2: Shortcomings and remedies" we show in which situations and for which tasks GPT-3 does not work well, which approaches might help then and what alternatives exist.

The general success factors of machine learning projects were recently the subject of our webinar "Real added value from ML projects - our success factors", in which we share some of our practical experiences. Recording and slides of the webinar are available in German.

For those waiting for new content on computer vision and remote sensing we have had our webinar "ML for Remote Sensing: analysing satellite data automatically", in which we gave an overview of available satellite data, machine learning methods for processing it automatically, and practical use cases in a corporate context. Recording and slides of the webinar are available in German.

If you are interested in the topics of our latest webinars, especially the remote sensing webinar, but would prefer to discuss them in English, please contact Moritz Weiten.

As always, we welcome feedback and topic requests for next time. And if you would like to discuss a specific topic with an experts, you are cordially invited to arrange a free Tech Lunch or ML Expert Talk with us.

Best regards
Philipp Jackmuth and Lorenz Richter