KlarTextCoders at StaGE: Automatic Statement Annotations for German Easy Language

Akhilesh Kakolu Ramarao, Wiebke Petersen, Anna Sophia Stein, Emma Stein, Hanxin Xia

Abstract: This paper presents our approach to the GermEval 2024 task, "Statement Segmentation in German Easy Language (StaGE)," addressing both subtasks: predicting the number of statements and identifying statement spans. We introduce a novel method integrating partof-speech information with pre-trained BERT models, achieving leading performance in both the subtasks. For the statement count prediction (subtask 1), our model achieved a precision of 0.65, a recall of 0.68, and an F1-score of 0.65, with a Mean Absolute Error (MAE) of 0.36 and Mean Squared Error (MSE) of 0.43. For statement span annotation (subtask 2), we adapted our BERT model (used for subtask 1) to perform token-level classification, achieving a chrF score of 0.36 and a Jaccard similarity of 0.29. We also detail our exploration of alternative approaches to the shared task, including a rule-based system, LLMs, and traditional machine learning models. These machine learning models used a comprehensive feature set, combining Abstract Meaning Representation (AMR) features to capture deep semantic structures, part-of-speech (POS) tags for syntactic information, and other linguistic features.