Paraphrase Generation

This guide details how to configure and run controllable paraphrase generation with VietQuill.

How Controllable Generation Works

VietQuill allows you to steer how the model paraphrases sentences using three control knobs (values from 0 to 100, rounded to nearest 5):

Parameter	What It Controls	Lower Value (0 - 40)	Higher Value (70 - 100)
lexical	Vocabulary overlap	Replaces many words with synonyms	Retains original wording & key terms
syntactic	Sentence structure	Restructures clauses and grammar	Keeps original sentence pattern
semantic	Meaning preservation	More creative / flexible rewriting	Strictly faithful to original meaning

VietQuill automatically attaches these constraints as a prefix before passing to the model:

SEM_<semantic> SYN_<syntactic> LEX_<lexical> : <your_text>

Example: SEM_90 SYN_85 LEX_60 : Hôm nay trời đẹp quá, mình muốn đi dạo công viên.

---

Generation Methods

1. Single Text Paraphrasing: paraphrase

from vietquill import EnsembleModelForParaphraseGeneration

model = EnsembleModelForParaphraseGeneration(device="cuda")

text = "Chính phủ đang triển khai nhiều chính sách hỗ trợ doanh nghiệp nhỏ và vừa."

results = model.paraphrase(
    text,
    lexical=60,
    syntactic=80,
    semantic=95,
    num_candidates=3,
    num_beams=5
)

for r in results:
    print("-", r)

2. Batch Paraphrasing: paraphrases

For processing large datasets or multiple sentences efficiently on GPU:

sentences = [
    "Hôm nay trời đẹp quá, mình muốn đi dạo công viên.",
    "Thủ đô của nước Pháp là thành phố nào?",
    "Học máy là một nhánh quan trọng của trí tuệ nhân tạo.",
    "Làm thế nào để học lập trình Python nhanh nhất?"
]

# Batch inference
batch_results = model.paraphrases(
    sentences,
    batch_size=8,
    num_candidates=2
)

for orig, cands in zip(sentences, batch_results):
    print(f"Original: {orig}")
    for c in cands:
        print(f"  -> {c}")

---

Style Presets

Instead of manually tuning percentages, you can use ParaphraseStyle:

Style Preset	Lexical	Syntactic	Semantic	Decoding	Best For
ParaphraseStyle.CONSERVATIVE	80	90	95	Beam search	Proofreading, minor rephrasing, high precision
ParaphraseStyle.BALANCED	65	85	85	Sampling, top_p=0.9, temperature=0.7, oversample=4	General-purpose paraphrasing
ParaphraseStyle.DIVERSE	40	60	85	Sampling, top_p=0.95, temperature=1.0, oversample=6, min_semantic=80	Data augmentation, creative rewriting

The control values alone change the output of the Ume checkpoints very little, so each style also sets decoding defaults. Arguments you pass explicitly (num_beams, top_p, do_sample, ...) always override the preset. Sampled styles give different outputs on each call; seed with torch.manual_seed for reproducibility.

DIVERSE can drift in meaning. Supply a quality estimator to drop candidates whose estimated semantic score is below min_semantic (the best-scoring candidate is kept if none pass). The estimator is never loaded implicitly; pass it per call or set model.estimator once:

from vietquill import EnsembleModelForParaphraseQualityEstimation

model.estimator = EnsembleModelForParaphraseQualityEstimation()
diverse = model.paraphrase(text, style=ParaphraseStyle.DIVERSE)  # filtered, min_semantic=80
strict = model.paraphrase(text, style=ParaphraseStyle.DIVERSE, min_semantic=90)

from vietquill import EnsembleModelForParaphraseGeneration, ParaphraseStyle

model = EnsembleModelForParaphraseGeneration()

text = "Dữ liệu lớn đóng vai trò then chốt trong chuyển đổi số."

conservative = model.paraphrase(text, style=ParaphraseStyle.CONSERVATIVE)
balanced = model.paraphrase(text, style="balanced") # Strings are also accepted
diverse = model.paraphrase(text, style=ParaphraseStyle.DIVERSE)

---

Advanced Decoding Parameters

paraphrase and paraphrases forward arbitrary keyword arguments to Hugging Face transformers.GenerationConfig:

• max_length (int): Maximum generated sequence length (default: 128).
• num_beams (int): Number of beams for beam search (default: 5). With do_sample=True, num_beams=1 draws independent samples.
• no_repeat_ngram_size (int): Prevent repetition of n-grams (default: 2).
• do_sample (bool): Sample instead of beam search (default: False).
• oversample (int): Extra sequences generated beyond num_candidates for de-duplication and filtering (default: 2).
• estimator / min_semantic: Optional semantic filtering, see Style Presets above.
• temperature (float): Sampling temperature (if do_sample=True).
• top_p (float): Nucleus sampling cutoff.

---

Single-Model Paraphrasing: AutoModelForParaphraseGeneration

If you are loading a standalone single checkpoint from Hugging Face Hub or a local path without sentence/question routing, use AutoModelForParaphraseGeneration:

from vietquill import AutoModelForParaphraseGeneration

# Loads directly from repository root
model = AutoModelForParaphraseGeneration(
    hub_id="ngwgsang/vietquill-vit5-base-sentence-tsubaki",
    device="cuda"
)

result = model.paraphrase(
    "Dữ liệu lớn đóng vai trò then chốt trong chuyển đổi số.",
    lexical=70,
    syntactic=80,
    semantic=90
)
print(result)