Vi-SRS: Vietnamese Summary Ranking Signals

Vi-SRS is a large-scale, fine-grained preference (ranking) dataset for Vietnamese abstractive summarization. It contains 120,323 (document, accepted_sum, rejected_sum) preference pairs constructed from 2,000 seed articles sampled from VietNews, covering six complementary preference-generation strategies that target four core summary-quality dimensions (factual consistency, coherence, informativeness, relevance) plus two auxiliary signals (generation degeneration and overall quality diversity).

The dataset is introduced and described in detail in our paper:

Preference Learning for Vietnamese Abstractive Summarization via Fine-grained Synthetic Preference Signals Trong Huy Nguyen, Vu-Loi Le, Huy-Cuong Phan, Hu-Tri-Dung Vo, Thai-Son Dang, Thanh-Huong Le Currently under review at Applied Intelligence (2026).

⚠️ Note: This paper is currently under review. The dataset is released ahead of publication to support reproducibility; the citation below will be updated once the paper is accepted.

Dataset Summary

Vi-SRS (short for Vietnamese Summary Ranking Signals) is built using SRS-Recipe, a synthetic preference-construction framework that generates preference pairs without requiring human annotation for every candidate summary. Given a source article, SRS-Recipe produces a preferred (accepted) summary and a less-preferred (rejected) summary that differ along a specific, targeted quality dimension. This provides denser and more interpretable supervision than conventional binary preference datasets, and is intended for training reward models and running preference-optimization algorithms such as DPO, CPO, and KTO for Vietnamese summarization.

• Language: Vietnamese (vi)
• Source articles: 2,000 seed articles sampled from the VietNews training split (stratified by domain and article length)
• Preference pairs: 120,323
• Construction: Fully synthetic (no human annotation per pair); combines LLM-based generation/judging (GPT-3.5/4o-mini, Gemini) with pretrained Vietnamese summarization models

Dataset Structure

Data Instances

Each line in data_ranking.jsonl is a JSON object representing one preference triplet. Example:

{
  "doc_id": 8199,
  "document": "UBND huyện Hàm Thuận Nam, tỉnh Bình Thuận vừa có kết luận thanh tra ...",
  "accepted_sum": "Ông Châu Toàn - Hiệu trưởng trường Tiểu học Hàm Thạnh 1 bị cáo buộc lập khống chứng từ ...",
  "rejected_sum": "Ông Nguyễn Văn An - Hiệu trưởng trường Tiểu học Hàm Tân 2 bị cáo buộc lập khống chứng từ ...",
  "category": "Incorrect Entities"
}

Data Fields

Field	Type	Description
doc_id	int	Identifier of the source article (shared across all preference pairs derived from the same article)
document	string	The source Vietnamese news article
accepted_sum	string	The preferred (higher-quality) candidate summary
rejected_sum	string	The less-preferred (lower-quality) candidate summary
category	string	The SRS-Recipe strategy that generated this preference pair (see below)

Preference Categories

Each pair is tagged with the strategy used to construct it. The six strategies fall into two groups: hard negatives (semantically close to the source but with a targeted quality defect) and easy negatives (largely irrelevant or degenerated).

category	Group	Targeted dimension	# Pairs
Levels of Coverage	Hard negative	Informativeness	54,000
Negative Pool	Easy negative	Degeneration	20,000
Diverse Summarization Models	Hard negative	Overall quality diversity	14,225
Incoherence	Hard negative	Coherence	12,000
Incorrect Entities	Hard negative	Factual consistency	10,098
Random Negative	Easy negative	Relevance	10,000
Total			120,323

The counts above reflect the minimum number of samples drawn under the training configuration used in the paper's experiments, not the total number of distinct candidate pairs available (e.g., Negative Pool and Random Negative candidates are dynamically sampled during training rather than materialized as fixed pairs).

Data Splits

The released file is a single flattened file spanning all 2,000 source articles (no explicit train/validation split column). In the paper, 1,900 articles are used for training and 100 for validation/testing of the reward model; users who need this split can partition by doc_id accordingly.

Dataset Creation

Source Data

Source articles are drawn from VietNews, a Vietnamese news summarization corpus.

Construction Strategies (SRS-Recipe)

1. Diverse Summarization Models — Summaries are generated by four heterogeneous systems (BARTpho, T5-small-vi-summarization, GPT-3.5, Mistral-7B-OpenOrca) plus the reference summary. Pairwise preferences are derived from agreement between two LLM judges (GPT-3.5-turbo, Gemini-1.0-pro), with cyclic preferences removed.
2. Incorrect Entities — A high-quality summary is generated (Gemini-1.5-pro), its entities are identified (GPT-4o-mini) and then perturbed into incorrect entities (GPT-4o-mini), producing summaries that are fluent but factually inconsistent.
3. Incoherence — Two ViT5-base models fine-tuned on VietNews generate summaries using interleaved top-3 sampling / greedy decoding to introduce controlled disfluency while preserving topical overlap.
4. Levels of Coverage — GPT-4o-mini identifies three salient facts per article and generates nine summaries spanning one-fact, two-fact, and three-fact coverage groups, yielding graded informativeness preferences.
5. Negative Pool — Severely degenerated candidates (repeated phrases, repeated sentences, high-temperature top-k sampling) generated per article and dynamically sampled as easy negatives during training.
6. Random Negatives — Summaries belonging to other, unrelated articles are used as easy negatives via a sliding-window sampling scheme.

Full construction details, prompts, and analysis (including a UMAP visualization showing that the six strategies induce distinguishable perturbation directions in embedding space) are provided in the paper.

Intended Use

Vi-SRS is designed for:

• Training reward models for Vietnamese summarization (e.g., the paper's ViSRS-RM)
• Preference optimization of summarization models with DPO, CPO, KTO, or similar algorithms
• Research on synthetic/scalable preference-signal construction for low-resource languages

Limitations

• Preference labels are synthetically constructed, not collected from human annotators; while the paper shows strong alignment with human ratings and LLM judgments, the signals are a proxy for human preference rather than a direct measurement.
• Preference pairs are not independent samples: many pairs share the same source article (doc_id), so care should be taken when splitting the data (e.g., split by doc_id to avoid document leakage across train/val/test).
• The dataset is Vietnamese news-domain text; generalization to other domains or genres is untested.

Citation

The associated paper is currently under review. Please check back for the finalized citation, or cite the preprint/repository in the meantime:

@article{nguyen2026visrs,
  title   = {Preference Learning for Vietnamese Abstractive Summarization via Fine-grained Synthetic Preference Signals},
  author  = {Nguyen, Trong Huy and Le, Vu-Loi and Phan, Huy-Cuong and Vo, Hu-Tri-Dung and Dang, Thai-Son and Le, Thanh-Huong},
  journal = {Applied Intelligence},
  year    = {2026},
  note    = {Under review}
}

Dataset Card Contact

For questions about the dataset, please open a discussion on this repository or contact the corresponding author (huonglt@soict.hust.edu.vn).