This model is a fine-tuned version of google-bert/bert-base-german-cased on the following dataset.

During inference, we apply the aggregation strategy "simple" from the Hugging Face pipeline API (aggregation_strategy="simple"). This ensures that contiguous tokens sharing the same entity label are grouped and returned as a single entity, improving usability in downstream applications.

The model achieves the following results on the evaluation set:

This token classification model is intended for identifying personal nouns in German texts, including gender-neutral forms (i.e. "Kletter:in, Kletter*innen").

The training corpus comprised approximately 150,000 tokens ("", ). Roughly 20 sentences extracted from the DAX-Briefe subset provided the primary representation of gender-fair language, a feature entirely absent from the 1% dataset.
Epicene nouns were excluded from the final training data because attempting to include them as instances of the positive class caused a noticeable decrease in model performance. Consequently, the current iteration of the model may not reliably detect epicene structures.

To mitigate the severe class imbalance between the 'OTHER' and 'P_NOUN' tokens, a downsampling strategy was implemented. Specifically, 70% of the sentences containing exclusively 'OTHER' tokens were removed from the dataset. This adjustment yielded a less skewed distribution, aligning with the proportions reported by Sökefeld et al. (2023). In the finalized dataset, 'P_NOUN' tokens accounted for approximately 4% of the total corpus.

We applied a pre-trained BERT tokenizer to sentences that had already been split into words. Model training followed the default hyperparameters recommended in the Hugging Face token classification tutorial. The corpus was split into training, validation, and test sets, with 80% used for training and 10% each for validation and testing. Model performance was evaluated at the token level on the held-out test set.

The following hyperparameters were used during training: