DNB-AI-Project at the GermEval-2025 LLMs4Subjects Task: KIFSPrompt - Knowledge-Injected Few-Shot Prompting

Maximilian Kähler, Lisa Kluge, Katja Konermann

Keywords: Automatic indexing, Extreme Multi Label Classification, Subject Tagging

Paper Type: System Description Paper

TL;DR: Inverting RAG-approaches to generate subject terms for a libraries cataloguing system.

Track: LLMs4Subjects

Abstract:

This work-in-progress report presents our system, KIFSPrompt, developed for the
second phase of the shared task LLMs4Subjects at GermEval'25. The primary
focus of this phase is the development of energy- and compute-efficient
large language model (LLM) systems for subject tagging in library cataloging
systems. A key challenge in this task is the requirement to select keywords from
a large, normed vocabulary, the Integrated Authority File (GND). Building on
our previous work ([11] and [12]), our
system leverages few-shot prompting, where a generative LLM is presented with a limited
number of examples of texts annotated with subject terms, and prompted to
identify the most relevant subject terms for a new input text.
To ensure alignment with the library's normed subject terms, we employ a
mapping approach based on embedding similarity.
We extend our previous work by incorporating a retrieval stage, which selects
relevant few-shot examples from the training set to create knowledge-injected prompts,
enabling the LLM to provide more specific and accurate keyword suggestions.
This enhancement allows the LLM to adapt to the input text, resulting in
improved performance in a single prompt.

Submission Number: 2