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.

This paper introduces our submission for the SemEval 2022 Task 8: Multilingual News Article Similarity. The task of the competition consisted of the development of a model, capable of determining the similarity between pairs of multilingual news articles. To address this challenge, we evaluated the Word Mover’s Distance in conjunction with word embeddings from ConceptNet Numberbatch and term frequencies of WorldLex, as well the Sentence Mover’s Distance based on sentence embeddings generated by pretrained transformer models of Sentence-BERT. To facilitate the comparison of multilingual articles with Sentence-BERT models, we deployed a Neural Machine Translation system. All our models achieve stable results in multilingual similarity estimation without learning parameters.

We present a system that creates pair-wise cosine and arccosine sentence similarity matrices using multilingual sentence embeddings obtained from pre-trained SBERT and Universal Sentence Encoder (USE) models respectively. For each news article sentence, it searches the most similar sentence from the other article and computes an average score. Further, a convolutional neural network calculates a total similarity score for the article pairs on these matrices. Finally, a random forest regressor merges the previous results to a final score that can optionally be extended with a publishing date score.

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.

This paper introduces the novel task of scene segmentation on narrative texts and provides an annotated corpus, a discussion of the linguistic and narrative properties of the task and baseline experiments towards automatic solutions. A scene here is a segment of the text where time and discourse time are more or less equal, the narration focuses on one action and location and character constellations stay the same. The corpus we describe consists of German-language dime novels (550k tokens) that have been annotated in parallel, achieving an inter-annotator agreement of gamma = 0.7. Baseline experiments using BERT achieve an F1 score of 24%, showing that the task is very challenging. An automatic scene segmentation paves the way towards processing longer narrative texts like tales or novels by breaking them down into smaller, coherent and meaningful parts, which is an important stepping stone towards the reconstruction of plot in Computational Literary Studies but also can serve to improve tasks like coreference resolution.

Humans frequently are able to read and interpret emotions of others by directly taking verbal and non-verbal signals in human-to-human communication into account or to infer or even experience emotions from mediated stories. For computers, however, emotion recognition is a complex problem: Thoughts and feelings are the roots of many behavioural responses and they are deeply entangled with neurophysiological changes within humans. As such, emotions are very subjective, often are expressed in a subtle manner, and are highly depending on context. For example, machine learning approaches for text-based sentiment analysis often rely on incorporating sentiment lexicons or language models to capture the contextual meaning. This paper explores if and how we further can enhance sentiment analysis using biofeedback of humans which are experiencing emotions while reading texts. Specifically, we record the heart rate and brain waves of readers that are presented with short texts which have been annotated with the emotions they induce. We use these physiological signals to improve the performance of a lexicon-based sentiment classifier. We find that the combination of several biosignals can improve the ability of a text-based classifier to detect the presence of a sentiment in a text on a per-sentence level.

Whenever researchers write a paper, the same question occurs: “Where to submit?” In this work, we introduce WTS, an open and interpretable NLP system that recommends conferences and journals to researchers based on the title, abstract, and/or keywords of a given paper. We adapt the TextCNN architecture and automatically analyze its predictions using the Integrated Gradients method to highlight words and phrases that led to the recommendation of a scientific venue. We train and test our method on publications from the fields of artificial intelligence (AI) and medicine, both derived from the Semantic Scholar dataset. WTS achieves an Accuracy@5 of approximately 83% for AI papers and 95% in the field of medicine. It is open source and available for testing on https://wheretosubmit.ml.

This paper describes our system for the SemEval 2019 Task 4 on hyperpartisan news detection. We build on an existing deep learning approach for sentence classification based on a Convolutional Neural Network. Modifying the original model with additional layers to increase its expressiveness and finally building an ensemble of multiple versions of the model, we obtain an accuracy of 67.52% and an F1 score of 73.78% on the main test dataset. We also report on additional experiments incorporating handcrafted features into the CNN and using it as a feature extractor for a linear SVM.

In this paper we describe our system for SemEval-2018 Task 7 on classification of semantic relations in scientific literature for clean (subtask 1.1) and noisy data (subtask 1.2). We compare two models for classification, a C-LSTM which utilizes only word embeddings and an SVM that also takes handcrafted features into account. To adapt to the domain of science we train word embeddings on scientific papers collected from arXiv.org. The hand-crafted features consist of lexical features to model the semantic relations as well as the entities between which the relation holds. Classification of Relations using Embeddings (ClaiRE) achieved an F1 score of 74.89% for the first subtask and 78.39% for the second.