texturejc/texture-frames-de-args

texture-frames-de · argument-extraction head

The argument-extraction stage of texture-frames-de, a German frame-semantic parser. Given a sentence with a marked trigger and its frame, it finds the spans that fill the frame's roles (frame elements) and labels each.

It fine-tunes deepset/gbert-large on the SALSA 2.0 corpus with a detect-then-classify design — two heads on one backbone, a single forward pass:

• Head A — span detection: a role-agnostic 3-class BIO tagger (O/B/I), "is this token part of an argument?". Dense signal, arbitrary-length spans.
• Head B — role classification: for each detected span, pool its tokens (start ⊕ end ⊕ mean) and classify into only the current frame's frame elements (plus a NULL reject class), masked via the bundled lexicon.

The input carries the predicate marker and the frame's FE menu ({frame} [FE1; FE2; …] : … <t> {trigger} </t> …). A NULL-bias at inference sets the precision/recall operating point.

This is one of three stages. Use it through the package rather than alone.

Usage

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

from texture_frames_de import FrameParser
parser = FrameParser()   # downloads this + the frame head on first use
for ann in parser.parse("Die Polizei verhaftete den Verdächtigen ."):
    print([(a.role, a.text) for a in ann.arguments])
# [('Authorities', 'Die Polizei'), ('Suspect', 'den Verdächtigen')]

Lower null_bias (default 2.0) for higher argument recall: FrameParser(null_bias=0.0).

Files

File	What
args2_model.pt	model state_dict (backbone + detection + role heads)
role2id.json	{role name → id} label map (incl. <NULL>) + base_model
tokenizer files	gbert-large tokenizer with the <t> / </t> markers added

The custom head (Args2Model) is defined in the package; loading is handled by texture_frames_de.weights.load_args.

Results

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

Metric	Value
Weighted F1 (non-core FEs = 0.5)	0.844 (P 0.884 / R 0.808, NULL-bias 2.0)
Speed	~17 ms/example (single forward pass)

Not directly comparable to the English texture-frames args head: SALSA role spans are syntactic constituents (clean boundaries), which flatters exact-span F1 relative to FrameNet's looser character spans. Discontinuous role spans (13.8% of gold, from German verb brackets / extraposition) are represented and scored as their enclosing span. 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, 4 sampled NULL negative spans/example. Data: SALSA 2.0, 80/10/10 split by sentence id. 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.