language	en
license	mit
tags	multiword-expressions mwe token-classification deberta nlp
datasets	yusuke196/CoAM
metrics	f1
pipeline_tag	token-classification

Binary MWE Detection (DLT+lo)

DeBERTa-v3-large fine-tuned for multiword expression identification using binary token-level classification. Handles both continuous and discontinuous MWEs.

Paper: "Binary Token-Level Classification with DeBERTa for All-Type MWE Identification" (EACL 2026 Findings)

Results

Model	Overall F1	Continuous F1	Discontinuous F1
DLT+lo (this)	69.8%	72.1%	53.8%
Qwen-72B	57.8%	—	—

Installation

pip install transformers torch spacy networkx
python -m spacy download en_core_web_lg

Usage

from transformers import AutoModel

model = AutoModel.from_pretrained("DiegoRossini/mwe-detection-deberta", trust_remote_code=True)

# Continuous MWE
mwes = model.detect("They made up their minds.")
print(mwes)  # ['made up']

# Discontinuous MWE
mwes = model.detect("I ran into an old friend yesterday.")
print(mwes)  # ['ran into']

# With detailed output
mwes = model.detect("He kicked the bucket last night.", return_details=True)

Thresholds

Default thresholds (start, end, inside): (0.5, 0.6, 0.2)

Adjust based on your precision/recall needs:

• Lower thresholds → more MWEs detected (higher recall)
• Higher thresholds → fewer but more confident MWEs (higher precision)

Training

Trained on CoAM with:

• Encoder: DeBERTa-v3-large
• Linguistic features: NP chunking, dependency distances
• Data augmentation: 30% oversampling

Code: github.com/DiegoRossini/binary-mwe-detection

Citation

@inproceedings{rossini2026binary,
    title = "Binary Token-Level Classification with {DeBERTa} for All-Type {MWE} Identification",
    author = "Rossini, Diego and van der Plas, Lonneke",
    booktitle = "Findings of EACL 2026",
    year = "2026"
}