timerlanmukhtarov/trade-news-datasets

Trade News Semantic Duplicates (URNO)

Dataset for detecting semantic duplicates in international trade/economic news, built as part of a Master's thesis (ВКР, ИТМО).

Pairs of news articles are labeled using the URNO scheme — a four-way diagnostic taxonomy that distinguishes structurally different types of negatives:

Label	Meaning
D	Duplicate — same trade event, same information
U	Update — same event, new data or development (hard negative)
R	Related — different but similar events, e.g. same sector (hard negative)
N	Not related — different topics/events
O	Out of scope — non-trade content

binary_label = 1 for D, 0 for all others.

Configurations

goldset — Gold Standard (2,250 pairs)

100% human-annotated. Used as the evaluation benchmark. Label distribution: D=54%, U=24%, N=19%, R=2%, O=1%. LLM ↔ Human agreement: 98.98%.

silverset — Silver Set (4,591 pairs)

LLM-annotated (DeepSeek) with train/validation splits. Used for fine-tuning duplicate detection models.

Usage

from datasets import load_dataset

# Gold standard (evaluation)
gold = load_dataset("timerlanmukhtarov/trade-news-datasets", "goldset")

# Silver set (fine-tuning)
silver = load_dataset("timerlanmukhtarov/trade-news-datasets", "silverset")
silver["train"][0]

Fields

goldset

Field	Type	Description
id	int	Row index
pair_id	string	Unique pair identifier
document_id_1	string	ID of article A
document_id_2	string	ID of article B
text_1	string	English summary of article A
text_2	string	English summary of article B
model_label	string	LLM annotation (D/U/R/N/O)
human_label	string	Human annotation (D/U/R/N/O) — ground truth
human_difficulty_tag	string	Annotator difficulty tag
binary_label	int	1=Duplicate (D), 0=Non-duplicate

silverset

Field	Type	Description
id	int	Row index
pair_id	string	Unique pair identifier
document_id_1	string	ID of article A
document_id_2	string	ID of article B
text_1	string	English summary of article A
text_2	string	English summary of article B
label	string	LLM annotation (D/U/R/N/O)
difficulty_tag	string	LLM-assigned difficulty
confidence_score	float	LLM confidence (0–1)
binary_label	int	1=Duplicate (D), 0=Non-duplicate

Source corpus

86,913 articles from 250+ multilingual news sources (EN/RU/PT/FR). Articles were embedded with BAAI/bge-m3, clustered via agglomerative clustering, then pairs were sampled per URNO category and annotated.

Citation

@mastersthesis{mukhtarov2026trade,
  author  = {Mukhtarov, Timerlan},
  title   = {Анализ и выявление семантических дублей в торговых новостях},
  school  = {ITMO University},
  year    = {2026}
}