Large Language Models (LLMs) are highly capable, state-of-the-art technologies and widely used as text classifiers for various NLP tasks, including sentiment analysis, topic classification, legal document analysis, etc. In this paper, we present a systematic analysis of the performance of LLMs as text classifiers using five German datasets from social media across 13 different tasks. We investigate zero- (ZSC) and few-shot classification (FSC) approaches with multiple LLMs and provide a comparative analysis with fine-tuned models based on Llama-3.2, EuroLLM, Teuken and BübleLM. We concentrate on investigating the limits of LLMs and on accurately describing our findings and overall challenges.

Text-generative large language models (LLMs) offer promising possibilities for terminology work, including term extraction, definition creation and assessment of concept relations. This study examines the performance of ChatGPT, Perplexity and Microsoft CoPilot for conducting terminology work in the field of the Austrian and British higher education systems using strategic prompting frameworks. Despite efforts to refine prompts by specifying language variety and system context, the LLM outputs failed to reliably differentiate between the Austrian and German systems and fabricated terms. Factors such as the distribution of German-language training data, potential pivot translation via English and the lack of transparency in LLM training further complicated evaluation. Additionally, output variability across identical prompts highlights the unpredictability of LLM-generated terminology. The study underscores the importance of human expertise in evaluating LLM outputs, as inconsistencies may undermine the reliability of terminology derived from such models. Without domain-specific knowledge (encompassing both subject-matter expertise and familiarity with terminology principles) as well as LLM literacy, users are unable to critically assess the quality of LLM outputs in terminological contexts. Rather than indiscriminately applying LLMs to all aspects of terminology work, it is crucial to assess their suitability for specific tasks.

This paper proposes an error classification framework for a comprehensive analysis of the output that large language models (LLMs) generate in a few-shot named entity recognition (NER) task in a specialised domain. The framework should be seen as an exploratory analysis complementary to established performance metrics for NER classifiers, such as F1 score, as it accounts for outcomes possible in a few-shot, LLM-based NER task. By categorising and assessing incorrect named entity predictions quantitatively, the paper shows how the proposed error classification could support a deeper cross-model and cross-prompt performance comparison, alongside a roadmap for a guided qualitative error analysis.

This paper evaluates GPT models as out-of-the-box Abstract Meaning Representation (AMR) parsers using prompt-based strategies, including 0-shot, few-shot, Chain-of-Thought (CoT), and a two-step approach in which core arguments and non-core roles are handled separately. Our results show that GPT-3.5 and GPT-4o fall well short of state-of-the-art parsers, with a maximum Smatch score of 60 using GPT-4o in a 5-shot setting. While CoT prompting provides some interpretability, it does not improve performance. We further conduct fine-grained evaluations, revealing GPT’s limited ability to handle AMR-specific linguistic structures and complex semantic roles. Our findings suggest that, despite recent advances, GPT models are not yet suitable as standalone AMR parsers.

In this work we investigate whether large language models (LLMs) ‘understand’ bridging relations and can use this knowledge effectively. We present the results obtained from two tasks: generation of texts containing bridging and filling in missing bridging spans. We show that in most of the cases LLMs fail to generate bridging in a reliable way.

Extended metaphor is the use of multiple metaphoric words that express the same domain mapping. Although it would provide valuable insight for computational metaphor processing, detecting extended metaphor has been rather neglected. We fill this gap by providing a series of zero- and few-shot experiments on the detection of all linguistic metaphors and specifically on extended metaphors with LLaMa and GPT models. We find that no model was able to achieve satisfactory performance on either task, and that LLaMa in particular showed problematic overgeneralization tendencies. Moreover, our error analysis showed that LLaMa is not sufficiently able to construct the domain mappings relevant for metaphor understanding.

Contemporary research in social sciences increasingly utilizes state-of-the-art generative language models to annotate or generate content. While these models achieve benchmark-leading performance on common language tasks, their application to novel out-of domain tasks remains insufficiently explored. To address this gap, we investigate how personalized language models align with human responses on the Moral Foundation Theory Questionnaire. We adapt open-source generative language models to different political personas and repeatedly survey these models to generate synthetic data sets where model-persona combinations define our sub-populations. Our analysis reveals that models produce inconsistent results across multiple repetitions, yielding high response variance. Furthermore, the alignment between synthetic data and corresponding human data from psychological studies shows a weak correlation, with conservative persona-prompted models particularly failing to align with actual conservative populations. These results suggest that language models struggle to coherently represent ideologies through in-context prompting due to their alignment process. Thus, using language models to simulate social interactions requires measurable improvements in in-context optimization or parameter manipulation to align with psychological and sociological stereotypes properly.