A Comparative Study of a Domain-Specific SLM and a General-Purpose LLM for Title Generation
The increasing adoption of artificial intelligence in journalism has created demand for language models that can support editorial workflows while complying with legal, ethical, and organizational requirements. In particular, concerns regarding copyrighted training data, transparency, and operational control have motivated interest in domain-specific small language models (SLMs) that can be trained entirely on licensed content. This study investigates the extent to which a custom domain-specific SLM is suitable for title generation in a journalistic environment when compared with the general-purpose large language model GPT-4o. The project wa
Resource information
- Authors
- Duelli, Lorenz, Mauczka, Andreas, Redl, Christoph
- Research area
- Artificial intelligence
- Publication date
- 2026
- DOI
- 10.5281/zenodo.20742268
- Upstream last modified
- Sep 14, 2026
- License
- Creative Commons Attribution 4.0 International
- Last record update
- Nov 6, 2026
About this resource
The increasing adoption of artificial intelligence in journalism has created demand for language models that can support editorial workflows while complying with legal, ethical, and organizational requirements. In particular, concerns regarding copyrighted training data, transparency, and operational control have motivated interest in domain-specific small language models (SLMs) that can be trained entirely on licensed content. This study investigates the extent to which a custom domain-specific SLM is suitable for title generation in a journalistic environment when compared with the general-purpose large language model GPT-4o. The project was realized in cooperation with the Austria Presse Agentur (APA). Its primary objective is to provide a local alternative to external LLMs such as GPT-4o. The model was trained exclusively on data from the APA news archive to ensure that only licensed material was used. The development process involved selecting a tokenizer, model architecture, and training parameters. Both automated and human evaluations were conducted to assess differences between title suggestions generated by the SLM, GPT-4o, and the original titlesof the source texts. The BERTScore and ROUGE-L indicate that the titles generated by the SLM are closer to the original reference titles. However, the examples generated by the LLM are preferred by human evaluators. Further improvements are needed in the training data, the model configuration and the training approach.