EuroBERT-610m-term-sense-balanced

T-EBERT: German IR Term Sense Disambiguation

Fine-tuned EuroBERT-610M on balanced data for distribution-robust German IR term sense disambiguation.

Model Description

This is a distribution-robust variant of T-EBERT, trained on artificially balanced data (50% IR sense, 50% colloquial) to learn semantic features without exploiting class frequency priors.

Key Differences from Standard T-EBERT

Performance

• F1 Score: 0.822 on balanced test set
• Training distribution: 50% IR sense, 50% colloquial (artificially balanced)

Intended Use

Primary use cases:

• Research on distribution robustness in NLP
• Corpora with unknown or varying class distributions
• Domain transfer scenarios
• Applications requiring symmetric per-class performance
• Benchmarking against distribution-agnostic baselines

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "pdjohn/T-EBERT-term-sense-balanced"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Classify a sentence
sentence = "Die internationale Norm verbietet den Einsatz von Gewalt."
inputs = tokenizer(sentence, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    prediction = torch.argmax(outputs.logits, dim=1).item()

# Interpretation
labels = {0: "Colloquial", 1: "IR Sense"}
print(f"Prediction: {labels[prediction]}")

Training Details

Training Data

• Corpus: Same as T-EBERT, but with balanced sampling
• Size: 456 sentences (training) + 114 sentences (test)
• Distribution: 50% IR sense, 50% colloquial (artificially balanced via undersampling)
• Terms: Norm, Kooperation, Regime, Integration, Intervention, Entspannung

Training Procedure

• Base model: EuroBERT-610M
• Fine-tuning method: LoRA (Low-Rank Adaptation)
  • Rank (r): 8
  • Alpha: 16
  • Target modules: q_proj, k_proj, v_proj
  • Bias training: Enabled (bias="all") - critical for balanced learning
  • Dropout: 0.05
• Epochs: 5
• Batch size: 8
• Learning rate: 1e-4

Limitations

• Lower peak performance: 0.82 vs 0.92 F1
• Smaller training set: 456 vs 934 samples (undersampling)
• see T-EBERT for best performance

Acknowledgements

This work was conducted as part of the Tracing International Institutions and Behavior (TIIB) project at TU Darmstadt.

Affiliation: Chair of Transnational Governance, Department of Political Science, Technische Universität Darmstadt

We thank the TIIB project team for their support and the manual annotation of the training corpus.