Between Metaphors, Language Models, and Structured Data: CRC 1475 at LREC 2026

The conference is regarded worldwide as one of the most important platforms for language resources, language modeling, and the evaluation of modern AI systems. Several researchers from the CRC 1475 “Metaphors of Religion” were also represented at the conference and the accompanying workshops. The focus was on questions that are also central to the work of the SFB: How can metaphorical and non-literal language be automatically recognized? What role do structured linguistic data play for modern AI systems? And what are the limitations of current Large Language Models (LLMs)? 

Structured Language Data as the Foundation of Explainable AI

In addition to the main conference, numerous specialized workshops took place. These included the workshop “Structured Linguistic Data and Evaluation (SLiDE).” There, international researchers discussed the importance of high-quality linguistic annotations and structured language resources for the development of transparent and robust AI systems.

During the workshop, Adam Roussel presented the paper “Fast and Flexible Example-based Treebank Search with Vector Symbolic Architectures.” The work explores new approaches to flexible search in complexly annotated linguistic data and combines symbolic linguistic structures with modern AI methods.

Colleagues from the CRC also participated in the poster session as part of the workshop. There, work on specialized language corpora and domain-specific data preparation was presented. The contributions highlighted how important high-quality annotations and sustainable language resources continue to be for computer-assisted language processing.

Metaphors as a Challenge for Large Language Models

Another workshop featuring participation from SFB 1475 was “Learning Non-Literal Expressions with Small Data @ LREC 2026.” The focus there was on metaphorical, ironic, and other non-literal linguistic phenomena—precisely those areas that remain particularly challenging for current language models.

As part of the workshop, Sebastian Reimann presented the paper “A Novel Dataset and Three Ways to Approach Automatic Metaphor Detection in German Religious Online Forums,” co-authored with Tatjana Scheffler. The work investigates the automatic detection of metaphorical language in German-language religious online forums and introduces a new dataset as well as various methodological approaches for this purpose. The central question is how complex semantic phenomena such as metaphors can be reliably analyzed using computer-aided methods—particularly in specialized domains and with limited data sets. The paper builds directly on the research of SFB 1475, which focuses on metaphorical meaning-making in religious communication and combines hermeneutic methods with computer-aided techniques.

The fact that metaphorical language remains a central test case for AI systems was also evident in many other conference presentations. It became clear time and again that while large language models are powerful tools, they still face limitations when it comes to interpreting implicit, context-dependent, or figurative meanings.

International Networking Between Linguistics and AI Research

LREC 2026 also highlighted how closely basic linguistic research, data annotation, and modern AI development are now intertwined. Topics such as the following were particularly well represented:

• structured language resources
• explainable AI
• semantic annotation
• metaphorical and non-literal language
• sustainable research data
• evaluation of generative language models

For the CRC 1475, the conference thus offered not only an international platform for presenting its own research but also important impetus for further work on computer-assisted methods for analyzing religious imagery.

Further information:
LREC 2026
Workshop on Structured Linguistic Data and Evaluation (SLiDE)
Workshop Learning Non-Literal Expressions with Small Data