deepquillapp/gector-grammar

DeepQuill grammar checker (GECToR)

A GECToR token tagger (Omelianchuk et al., 2020) for grammar checking in fiction, exported to ONNX for the DeepQuill writing app. One forward pass labels each word $KEEP, $REPLACE_x, $APPEND_x, $DELETE, $TRANSFORM_VERB_* or $TRANSFORM_AGREEMENT_*.

Trained by DeepQuill with tools/gec-train in the DeepQuill repository, using only data and code that allow commercial use. No BEA-2019 data (Lang-8, NUCLE, FCE, W&I+LOCNESS) and no published GECToR checkpoint were used.

Files

File
model.onnx	weights stored fp16, cast to fp32 at load; inputs input_ids, attention_mask [1,T]; outputs label_probs [1,T,L] (softmax) and error_probs [1,T]
tokenizer.json	RoBERTa byte-level BPE; encode ["$START", *words] as pre-tokenized input
labels.json	{labels, keep_label, max_length}; the label head is labels minus the trailing <PAD>
verb-form-vocab.txt	verb-form table for $TRANSFORM_VERB_* tags

Read each word's tag from its first sub-token. DeepQuill flags a sentence only when a word's error probability reaches 0.6, and applies a tag only when its probability reaches 0.6. It drops $TRANSFORM_CASE_* tags.

Evaluation

On a 109k-word novel at threshold 0.6, the model catches 6 of 8 known errors, leaves all 12 known traps (dialect, fragments, names) alone, and raises 14.4 flags per 1,000 sentences.

Training data and attribution

What	Source	Licence
Base encoder	roberta-base (Facebook AI)	MIT
General errors	C4_200M (Stahlberg & Kumar, 2021, Google), via liweili/c4_200m	edits CC BY 4.0; text ODC-BY
Underlying text	C4 (Raffel et al., 2020)	ODC-BY
Fiction	Project Gutenberg, via sedthh/gutenberg_english	US public domain
Training code	gotutiyan/gector @ a4a342e	MIT
Verb forms	grammarly/gector data/verb-form-vocab.txt @ 3d41d28	Apache-2.0

• Felix Stahlberg and Shankar Kumar. 2021. Synthetic Data Generation for Grammatical Error Correction with Tagged Corruption Models. BEA workshop. Dataset: https://github.com/google-research-datasets/C4_200M-synthetic-dataset-for-grammatical-error-correction (CC BY 4.0).
• Colin Raffel et al. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. JMLR. C4 is released under ODC-BY.
• Kostiantyn Omelianchuk et al. 2020. GECToR – Grammatical Error Correction: Tag, Not Rewrite. BEA workshop.