Speakers
Description
This article presents an application evaluation of the use of large language models (LLMs) for the semantic classification of lexicographical content in three German dialect dictionary projects. Building on earlier experiments with LLM-based semantic classification that showed hit rates of over 80%, the study investigates whether generative AI can support the assignment of meanings to controlled semantic categories while preserving editorial oversight. Evaluation data from user ratings, surveys, and interviews indicate that the tool produces useful classifications in a majority of cases, reduces manual research effort, and contributes to greater consistency in semantic annotation. At the same time, the results reveal recurring limitations, especially in cases involving semantic granularity, context-dependent meanings, and definitions where modifiers are overemphasized. The paper argues that LLMs are most valuable not as replacements for expert lexicographical work, but as components of hybrid workflows that combine automated suggestion with human validation.