English Split and Rephrase Rewriter
This T5 model rewrites one complex English sentence into two shorter,
meaning-related sentences. The two output parts are separated by <SEP>.
Usage
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
model_id = "soundstarrain/english-split-and-rephrase-rewriter"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
text = "The service stores the file and returns a warning when validation fails."
inputs = tokenizer("split: " + text, return_tensors="pt", truncation=True)
outputs = model.generate(**inputs, max_new_tokens=192, num_beams=1, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))
The expected output format is:
The service stores the file. <SEP> A warning is returned when validation fails.
The companion GitHub repository adds recursive processing, protected-span handling, and lightweight output checks.
Project repositories
| Repository | Purpose |
|---|---|
| Inference code | Combines both models in a recursive pipeline and provides the CLI |
| Gate model | Predicts whether another split is needed |
| Rewriter model | This model; rewrites one sentence into two related parts |
Intended use
Use this model to generate a two-part English split-and-rephrase candidate inside a sentence preprocessing pipeline.
Evaluation
On the held-out BiSECT test split, the model produced two non-empty output
parts for 100% of 527 examples. Mean token-F1 against the reference
rewrites was 0.7500. These are automatic BiSECT test metrics. Human
evaluation and out-of-domain performance are not measured.
Limitations
The reported evaluation covers the BiSECT test split. Output quality can vary with genre, sentence length, terminology, and the amount of rewriting needed.
Training data
The checkpoint was fine-tuned on 200,000 filtered English split-and-rephrase
pairs derived from BiSECT. The training corpus is not included in this
repository. The model card records GEM/BiSECT and the original paper for
attribution.
License and attribution
Apache-2.0 applies to the released model files and metadata. The checkpoint is
fine-tuned from google-t5/t5-base. The training corpus is not redistributed
here. See NOTICE.md for attribution.
Citation
Kim, J., Maddela, M., Kriz, R., Xu, W., and Callison-Burch, C. “BiSECT: Learning to Split and Rephrase Sentences with Bitexts.” EMNLP 2021.