Zero-Shot Award Criteria extraction via Large Language Models from German Procurement Data from Switzerland

Public procurement serves as a model for sustainable practices (Sönnichsen and Clement, 2020). Recent legislation in Switzerland mandates considerations of economic, environmental, and social responsibility in public spending, including within the realm of public procurement. To assess the extent to which these legislative measures have influenced public procurement practices, one may examine Award Criteria (ACs) based on which procuring entities determine the most suitable bidder. This paper demonstrates the potential of Natural Language Processing (NLP) for extracting ACs from Swiss calls for tenders (CFTs), specifically those in German. We evaluate the efficacy of a German Large Language Model (LLM) in executing four tasks with a single zero-shot prompt: (1) Text Classification (TC), determining whether a call for tenders (CFT) includes ACs; (2) Named Entity Recognition (NER), identifying ACs and other related named entities; (3) Relation Extraction (RE), elucidating relationships between named entity instances; and (4) Formatting, compiling the information into a structured JSON format. We evaluate our approach on a set of 167 annotated CFTs. This approach facilitates the automated monitoring and evaluation of ACs overtime regarding sustainability. Both our code and the annotated dataset are publicly available: https://github.com/kapllan/GATE-CH.