ryosuke-yamaki/hol-ccg-roberta-large

Hol-CCG

Official pretrained checkpoint for Holographic CCG Parsing (Yamaki, Taniguchi, and Mochihashi; ACL 2023), released with the official implementation at ryosuke-yamaki/hol-ccg.

Hol-CCG formulates CCG parsing as recursive composition in a continuous vector space, using holographic embeddings (circular correlation) as an explicit compositional operator over word and phrase vectors. This checkpoint uses real-valued normalization (the default configuration of the official implementation) with a fine-tuned roberta-large encoder.

Performance on CCGbank

Split	LF1	Supertagging Acc.
dev (WSJ 00)	92.73	96.49
test (WSJ 23)	92.69	96.57

Model family

Model	Normalization	Encoder	Test LF1
ryosuke-yamaki/hol-ccg-roberta-large (this model)	real	roberta-large	92.69
ryosuke-yamaki/hol-ccg-roberta-large-complex	complex	roberta-large	92.10
ryosuke-yamaki/hol-ccg-roberta-base	real	roberta-base	91.98
ryosuke-yamaki/hol-ccg-roberta-base-complex	complex	roberta-base	91.56

Usage

Install the official implementation with the [hub] extra, then load the model by repository ID:

from holccg import Parser

parser = Parser.from_pretrained("ryosuke-yamaki/hol-ccg-roberta-large", device="cpu")

parsed = parser.parse_tokens(
    ["The", "dog", "barks", "."], index=1, sentence_id="sentence.1"
)
print(parsed.status.value)  # "success"
print(parsed.auto)          # CCGbank .auto derivation string

Use device="cuda" for GPU inference. For the holccg-parse CLI, download the bundle and pass its directory:

hf download ryosuke-yamaki/hol-ccg-roberta-large --local-dir holccg-roberta-large
holccg-parse --model_dir holccg-roberta-large --input_raw_path in.raw --output_auto_path out.auto

Training data and licence

The model was trained on CCGbank v1.1 (LDC2005T13), WSJ sections 02-21. The released files contain model weights, category-label inventories, and rule-frequency statistics derived from CCGbank annotations; no CCGbank text is included. Reproducing training or evaluation requires your own licensed copy of CCGbank. The model weights are released under the MIT licence.

Citation

@inproceedings{yamaki-etal-2023-holographic,
    title = "Holographic {CCG} Parsing",
    author = "Yamaki, Ryosuke  and
      Taniguchi, Tadahiro  and
      Mochihashi, Daichi",
    editor = "Rogers, Anna  and
      Boyd-Graber, Jordan  and
      Okazaki, Naoaki",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.15/",
    doi = "10.18653/v1/2023.acl-long.15",
    pages = "262--276"
}