DGS-Fabeln-1-SE

# DGS-Fabeln-1-SE This is the DGS-Fabeln-1-SE dataset: DGS (German Sign Language) fairy tales 1 with sentiment estimation. This dataset complements the already-published DGS-Fabeln-1 dataset (Data: https://zenodo.org/records/12686707. Paper: https://aclanthology.org/2024.lrec-main.434/) Their conjunction makes a parallel corpus between: 1. From DGS-Fabeln-1: German sign language (DGS) video segments (average ~10 seconds each) of a deaf native narrating classic German fairy tales. 2. From DGS-Fabeln-1: for each segment, corresponding text in simplified German written language. 3. From DGS-Fabeln-1-SE: for each segment, their valence on three levels (negative/neutral/positive), estimated from automated LLMs-based text analysis. 4. From DGS-Fabeln-1-SE: for each segment, raw landmarks and blendshape animation data extracted from MediaPipe 5. From DGS-Fabeln-1-SE: for each segment, a feature vector computed from the raw MediaPipe data (including, mean, SDs, velocities, accellerations, distance between landmarks, ...) This DOI for this dataset is: 10.5281/zenodo.18879038 ## Data structure The DGS-Fabeln-1-SE dataset consists of the following tables: * `DGS-Fabeln-1-SE-Labels.csv` - These are the sentiment/valence labels, i.e., our ground truth for predictions. The table contains both the labels extracted by fours LLMs and their aggregation by majority voting. * Header: Story, id, text_original, Sentiments-GPT5, Multi-GPT5, Sentiments-Perplexity, Multi-Perplexity, Sentiments-Mistral, Multi-Mistral, Sentiments-GPTOSS20B, Multi-GPTOSS20B, Sentiments-Aggregated, Multi-Aggregated * 13 columns * 574 samples * `DGS-Fabeln-1-SE-MotionFeatures.csv` - segment-level features computed from the raw MediaPipe data (e.g., mean, SDs, velocities, accellerations, distance between landmarks, ...). * Header: Story, id, 396x features * 398 columns * 571 samples (there are three missing videos in the original DGS-Fabeln-1 dataset) * 7x `DGS-Fabeln-1-SE-MediaPipe-<tale>.csv` - For each of the seven fairy tales (1-DHUDI, 2-FrauHolle, ...) a file containing the data directly extracted from MediaPipe and interpolated for missing frames, one frame per row. * Header: Story, id, frame, 296 features * 299 columns * Number of samples is the number of frames in the corresponding video ## Publication(s) For this dataset: ``` Nunnari, Fabrizio, Siddhant Jain, Patrick Gebhard. 2026. “Sentiment Analysis of German Sign Language Fairy Tales.” Proceedings of the fifteenth biennial Language Resources and Evaluation Conference (LREC 2026) (Palma, Mallorca, Spain), May. ``` For the connected DGS-Fabeln-1 dataset: ``` Nunnari, Fabrizio, Eleftherios Avramidis, Cristina España-Bonet, Marco González, Anna Hennes, and Patrick Gebhard. 2024. “DGS-Fabeln-1: A Multi-Angle Parallel Corpus of Fairy Tales between German Sign Language and German Text.” Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (Torino, Italy), May, 4847–57. https://aclanthology.org/2024.lrec-main.434. ```