The aim of the research project was to describe products that require explanation, such as board games, learning materials or creative children's sets, in an efficient, high-quality and target group-specific manner. Under the leadership of Prof. Dr. Alexander Bartel, the researchers at TTZ Günzburg used innovative methods of Natural Language Processing (NLP) together with their practice partner: Instead of developing their own language model, they combined advanced NLP techniques with sophisticated prompt engineering and an experimentally tuned parameterization. Among other things, this resulted in the new prompting technique “Adaptive Contextual Few-Shot Prompting”, which extends the concept of Few-Shot Prompting. This enabled them to precisely capture the meaning of complex source information and generate new, context-appropriate texts from it. Specifically, this means that the AI processes a wide variety of formats of product descriptions from Hutter Trade - from game instructions and blurb texts to structured product data - and extracts context-relevant information from them to generate an appropriately adapted text.
The focus: ideal adaptation to target groups and e-commerce platforms
During development, particular attention was paid to controllable text production: among other things, the application makes it possible to adapt the style, tonality and linguistic characteristics to different target groups - such as parents, teachers or children. By integrating specially developed components for recognizing emotional moods (so-called sentiment modules), emotional effects can be specifically controlled without compromising information density or brand coherence.
The application also takes into account the platform-specific requirements of digital commerce: the generated texts meet all the formal specifications of modern e-commerce platforms - including length restrictions, language variants and multilingualism. In addition, the keywords relevant to the findability of a product are extracted from the e-commerce platform and integrated into the generation of the texts. This not only ensures that the generated texts formally meet the specified criteria of eCommerce platforms, but also helps to increase the findability of the products on offer in terms of content.
AI application reduces text generation effort by 70 percent
“Thanks to the AI application we developed, we were able to reduce our text generation costs in the eCommerce sector by at least 70%,” summarizes Herrmann Hutter, CEO of Hutter Trade. “This has a direct impact on our efficiency and competitiveness: we gain valuable resources for other tasks and can place our product range on the market faster and more flexibly.”
The TTZ Günzburg project is an example of how scalable, AI-supported text generation can be successfully implemented at the interface between linguistic modeling, user psychology and digital marketing.
Contact
Prof. Dr. Alexander Bartel
