We assemble a broad Natural Language Understanding benchmark suite for the German language and consequently evaluate a wide array of existing German-capable models in order to create a better understanding of the current state of German LLMs. Our benchmark consists of 29 different tasks ranging over different types such as document classification, sequence tagging, sentence similarity, and question answering, on which we evaluate 10 different German-pretrained models, thereby charting the landscape of German LLMs. In our comprehensive evaluation we find that encoder models are a good choice for most tasks, but also that the largest encoder model does not necessarily perform best for all tasks. We make our benchmark suite and a leaderboard publically available at https://supergleber.professor-x.de and encourage the community to contribute new tasks and evaluate more models on it (https://github.com/LSX-UniWue/SuperGLEBer).

For our submission for Subtask 1, we developed a custom classification head that is designed to be applied atop of a Large Language Model. We reconstructed the hierarchy across multiple fully connected layers, allowing us to incorporate previous foundational decisions in subsequent, more fine-grained layers. To find the best hyperparameters, we conducted a grid-search and to compete in the multilingual setting, we translated all documents to English.

We present an intuitive approach for hallucination detection in LLM outputs that is modeled after how humans would go about this task. We engage several LLM “experts” to independently assess whether a response is hallucinated. For this we select recent and popular LLMs smaller than 7B parameters. By analyzing the log probabilities for tokens that signal a positive or negative judgment, we can determine the likelihood of hallucination. Additionally, we enhance the performance of our “experts” by automatically refining their prompts using the recently introduced OPRO framework. Furthermore, we ensemble the replies of the different experts in a uniform or weighted manner, which builds a quorum from the expert replies. Overall this leads to accuracy improvements of up to 10.6 p.p. compared to the challenge baseline. We show that a Zephyr 3B model is well suited for the task. Our approach can be applied in the model-agnostic and model-aware subtasks without modification and is flexible and easily extendable to related tasks.

In this paper, we describe our approach to the clickbait spoiling task of SemEval 2023.The core idea behind our system is to leverage pre-trained models capable of Question Answering (QA) to extract the spoiler from article texts based on the clickbait title without any task-specific training. Since oftentimes, these titles are not phrased as questions, we automatically rephrase the clickbait titles as questions in order to better suit the pretraining task of the QA-capable models. Also, to fit as much relevant context into the model’s limited input size as possible, we propose to reorder the sentences by their relevance using a semantic similarity model. Finally, we evaluate QA as well as text generation models (via prompting) to extract the spoiler from the text. Based on the validation data, our final model selects each of these components depending on the spoiler type and achieves satisfactory zero-shot results. The ideas described in this paper can easily be applied in fine-tuning settings.

Structured Sentiment Analysis is the task of extracting sentiment tuples in a graph structure commonly from review texts. We adapt the Aspect-Based Sentiment Analysis pointer network BARTABSA to model this tuple extraction as a sequence prediction task and extend their output grammar to account for the increased complexity of Structured Sentiment Analysis. To predict structured sentiment tuples in languages other than English we swap BART for a multilingual mT5 and introduce a novel Output Length Regularization to mitigate overfitting to common target sequence lengths, thereby improving the performance of the model by up to 70%. We evaluate our approach on seven datasets in five languages including a zero shot crosslingual setting.