Model Card: AIxcellent Vibes' Model for Candy Speech Detection

Model Details

• Model Type: Transformer-based encoder (XLM-RoBERTa-Large)
• Developed by: Christian Rene Thelen, Patrick Gustav Blaneck, Tobias Bornheim, Niklas Grieger, Stephan Bialonski (FH Aachen, RWTH Aachen, ORDIX AG, Utrecht University)
• Paper: AIxcellent Vibes at GermEval 2025 Shared Task on Candy Speech Detection: Improving Model Performance by Span-Level Training
• Base Model: XLM-RoBERTa-Large (Conneau et al., 2020)
• Fine-tuning Objective: Detection of candy speech (positive/supportive language) in German YouTube comments.

Model Description

• Binary level: Classify whether a comment contains candy speech.
• Span level: Detect the exact spans and categories of candy speech within comments, using a BIO tagging scheme across 10 categories (positive feedback, compliment, affection declaration, encouragement, gratitude, agreement, ambiguous, implicit, group membership, sympathy).

Intended Uses

• Research: Analysis of positive/supportive communication in German social media.
• Applications: Social media analytics, conversational AI safety (mitigating sycophancy), computational social science.
• Not for: Deployments without fairness/robustness testing on out-of-domain data.

Performance

• Dataset: 46k German YouTube comments, annotated with candy speech spans.
• Training Data Split: 37,057 comments (train), 9,229 (test).
• Shared Task Results:
  • Subtask 1 (binary detection): Positive F1 = 0.891 (ranked 1st)
  • Subtask 2 (span detection): Strict F1 = 0.631 (ranked 1st)

Training Procedure

• Architecture: XLM-RoBERTa-Large + linear classification layer (BIO tagging, 21 labels including “O”).
• Optimizer: AdamW
• Learning Rate: Peak 2e-5 with linear decay and warmup (500 steps).
• Epochs: 20 (with early stopping).
• Batch Size: 32
• Regularization: Dropout (0.1), weight decay (0.01), gradient clipping (L2 norm 1.0).
• Postprocessing: BIO tag correction and subword alignment.

Limitations

• Domain Specificity: Trained only on German YouTube comments; performance may degrade on other platforms, genres, or languages.
• Overlapping Spans: Cannot handle overlapping spans, as they were rare (<2%) in the training data.
• Biases: May reflect biases present in the dataset (e.g., demographic skews in YouTube communities).
• Generalization: Needs evaluation before deployment in real-world moderation systems.

Ethical Considerations

• Positive speech detection is less studied than toxic speech, but automatic labeling of “supportiveness” may reinforce cultural biases about what counts as “positive.”
• Must be complemented with human-in-the-loop moderation to avoid misuse.

Citation

@inproceedings{thelen-etal-2025-aixcellent,
    title = "{AI}xcellent Vibes at {G}erm{E}val 2025 Shared Task on Candy Speech Detection: Improving Model Performance by Span-Level Training",
    author = "Thelen, Christian Rene  and
      Blaneck, Patrick Gustav  and
      Bornheim, Tobias  and
      Grieger, Niklas  and
      Bialonski, Stephan",
    editor = "Wartena, Christian  and
      Heid, Ulrich",
    booktitle = "Proceedings of the 21st Conference on Natural Language Processing (KONVENS 2025): Workshops",
    month = sep,
    year = "2025",
    address = "Hannover, Germany",
    publisher = "HsH Applied Academics",
    url = "https://aclanthology.org/2025.konvens-2.33/",
    pages = "398--403"
}