texturejc/texture-frames-args

texture-frames · argument-extraction head

The argument-extraction stage of texture-frames, a fast FrameNet semantic-frame 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 microsoft/deberta-v3-large on FrameNet 1.7 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 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

from texture_frames import FrameParser
parser = FrameParser()
for ann in parser.parse("The chef gave food to the customer ."):
    print([(a.role, a.text) for a in ann.arguments])
# [('Donor', 'The chef'), ('Theme', 'food'), ('Recipient', 'to the customer')]

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	DeBERTa-v3 tokenizer with the <t> / </t> markers added

Loading is handled by texture_frames.weights.load_args.

Results

Open-Sesame test split, weighted F1 (non-core FEs = 0.5):

Metric	This head	T5 baseline
Argument F1	0.750	0.753
Speed	single forward pass (~50–60 ms)	3 beam-search passes

Parity with the generative baseline while running ~4× faster. The encoder went 0.628 (flat BIO) → 0.712 (detect-then-classify) → 0.750 (+ WordNet augmentation) across redesigns.

Training

microsoft/deberta-v3-large, AdamW lr 1e-5, warmup 0.06, weight decay 0.01, batch 16, max length 320, bf16, 6 epochs with WordNet synonym augmentation. Data: FrameNet 1.7 (NLTK), Open-Sesame splits.

Licence

Code (the package): MIT. Weights: trained on FrameNet 1.7, which carries its own academic-use terms — review them before redistributing.

Citation

@software{texture_frames,
  author = {Carney, James},
  title  = {texture-frames: a fast DeBERTa encoder FrameNet parser},
  url    = {https://github.com/texturejc/Texture_Frames},
  year   = {2026}
}

Builds on David Chanin's frame-semantic-transformer; thanks to the Berkeley FrameNet and Open-Sesame projects.