ghotriw/deberta-v3-base-biaffine-dep-pos-en-ewt

ModernBiaffineParser — microsoft/deberta-v3-base

Biaffine dependency parser + joint UPOS tagger trained on Universal Dependencies English Web Treebank (EWT).

Encoder: microsoft/deberta-v3-base (frozen weights not included — loaded from HuggingFace at runtime) Custom head: biaffine_head.safetensors — word projection, arc/rel/POS MLPs and biaffine layers Labels: 53 DEPREL labels · 19 UPOS tags Score convention: s_arc[dep, head], s_rel[dep, head, rel]

Metrics (EWT, decode: Eisner (projective MST))

Split	LAS	UPOS	UCM	LCM
dev	93.57%	98.12%	75.46%	66.37%
test	93.52%	98.10%	76.75%	67.79%

ONNX / production use

model.fp16.onnx — fp16 model for CPU inference via ONNX Runtime (~357 MB). Inputs: subwords [B, W, 20] int64. Outputs: s_arc [B,W,W], s_rel [B,W,W,R], word_repr [B,W,768], s_pos [B,W,P].

# download only inference artifacts
hf download ghotriw/deberta-v3-base-biaffine-dep-pos-en \
  model.fp16.onnx vocabs.json idiom_classifier.json tokenizer.json \
  --local-dir ./model

hf download ghotriw/deberta-v3-base-biaffine-dep-pos-en \
  lexicon.json phrasal-verbs.json \
  --local-dir ./dic

vocabs.json — DEPREL and UPOS vocabularies (str→int dicts). idiom_classifier.json — linear idiom head (W·mean(word_repr)+b, BCE). lexicon.json + phrasal-verbs.json — candidate lexicons for idiom/phrasal-verb detection.

Input format

The model expects a word-level subword grid [B, W, fix_len=20] (int64), where each word is independently tokenised with the encoder's sentencepiece tokeniser and padded/truncated to 20 subword slots. Position 0 is a synthetic ROOT word whose only subword is [CLS] (id 1).

Vocabularies

config.json contains rel_vocab (str→int) and pos_vocab (str→int). Index 0 is the <pad> / ROOT slot and should be ignored in evaluation.