The frame-classification stage of texture-frames-de, a German frame-semantic parser. Given a sentence with a marked trigger, it predicts which of 1,027 frames the trigger evokes.

It fine-tunes deepset/gbert-large on the SALSA 2.0 corpus and uses marker-token pooling: the trigger is wrapped in entity markers (… <t> kündigte </t> …) and the frame representation is the concatenation of the two marker tokens' hidden states (not [CLS]), focusing the classifier on the predicate. A single forward pass — no beam search.

At inference the logits are soft-masked toward the trigger lemma's candidate frames (via a bundled SALSA lexicon + simplemma lemmatization), so a confident non-candidate can still win while golds outside the top candidate are recovered.

The custom head (FrameMarkerModel) is defined in the package; loading is handled by texture_frames_de.weights.load_frame.

Test split (held-out 10% of SALSA sentences), operating point picked on dev:

Not directly comparable to the English texture-frames frame head — different corpus, label space (1,027 vs 1,221), and splits. Read as a strong standalone German result.

deepset/gbert-large, 5 epochs, AdamW lr 1e-5, warmup 0.06, weight decay 0.01, batch 16, max length 320, bf16. Data: SALSA 2.0, 80/10/10 split by sentence id (train 30,089 / dev 3,787 / test 3,729 frame instances). See the repo for the training notebook.

Code (the package): MIT. Weights: for non-commercial research use. They are trained on SALSA, layered on TIGER — both academic / non-commercial licences, with SALSA additionally restricting commercial use of derived data. Review the SALSA and TIGER licence terms before any commercial use or redistribution. The corpus itself is not distributed here and must be obtained under licence.

Builds on David Chanin's frame-semantic-transformer and its encoder rearchitecture texture-frames; thanks to the SALSA and TIGER projects and to deepset for gbert-large.