texturejc/texture-frames-de-frame

texture-frames-de · frame-classification head

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.

This is one of three stages. Use it through the package rather than alone; the pipeline handles trigger detection and argument extraction around it.

Usage

pip install git+https://github.com/texturejc/texture-frames-de

from texture_frames_de import FrameParser
parser = FrameParser()   # downloads this + the args head on first use
for ann in parser.parse("Die Polizei verhaftete den Verdächtigen am Bahnhof ."):
    print(ann.frame, "|", ann.trigger)
# Arrest | verhaftete

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.

Files

File	What
frame2_model.pt	model state_dict (backbone + marker-pooling classifier)
frame2id.json	{frame name → id} label map + base_model
tokenizer files	gbert-large tokenizer with the <t> / </t> markers added

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

Results

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

Metric	Value
Frame accuracy	0.9045 (candidate bias 4.0)
Candidate-coverage ceiling	0.984
Speed	~16 ms/example (single forward pass)

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.

Training

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.

Licence

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.

Citation

@software{texture_frames_de,
  author = {Carney, James},
  title  = {texture-frames-de: a German frame-semantic parser (gbert / SALSA)},
  url    = {https://github.com/texturejc/texture-frames-de},
  year   = {2026}
}

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.