parl-german-speechact

Annotation Layer: Speechact

This model is part of GePaDeU, which equips parliamentary debates of the German Bundestag with rich semantic and pragmatic information across multiple annotation layers.

parl-german-speechact is trained to classify a given speech act segemnt enriched with its surrounding context (defined as previous and following sentence) into zero, one or multiple speech act types. Speech acts are discouse units, serving specific pragmatic/rhetorical functions. In our taxonomy, speech acts capture different types of cooperation (e.g., Request, Report) and conflict (e.g., Accusation, Demand) communication in parliamentary debates.

---

🔍 Model Overview

• Task Type: Multi-label, multi-class sequence classification
• Base Model: GBERT large
• Fine-tuning method: full fine-tuning
• Language: German

---

📚 Dataset

Models were trained and evaluated on 250 manually annotated parliamentary speeches from the German Bundestag, ranging from 2017-2021, resulting in 12,947 speech act segments.

Data Splits

• Train: 8,583 instances
• Dev: 1,214 instances
• Test: 3,150 instances

---

🏋️ Model Training

---

📊 Evaluation

---

🚀 How to Use

Please, refer to our GitHub repo for detailed instructions on the required input format and how to run the model.

---

⚠️ Limitations