TradeNewsEventDedup / README.md
lyutovad's picture
Update README.md
1b37ee6 verified
|
Raw History Blame Contribute Delete
4.65 kB
metadata
pretty_name: Trade News Event-Level Deduplication Dataset
language:
  - en
license: cc-by-4.0
task_categories:
  - text-classification
task_ids:
  - semantic-similarity-classification
tags:
  - duplicate-detection
  - deduplication
  - information-retrieval
  - news
  - trade
  - weak-supervision
  - hard-negatives
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.parquet
      - split: test
        path: test.parquet

Trade News Event-Level Deduplication Dataset

Dataset Summary

This dataset is designed for event-level duplicate detection in trade and foreign economic news.

Each instance consists of a pair of English summaries labeled according to whether they describe the same underlying trade event (case).

The dataset follows a Gold/Silver protocol:

  • train (SilverSet) — weakly labeled data used for training
  • test (GoldSet) — manually validated data used exclusively for evaluation

The task goes beyond surface similarity and focuses on event identity, requiring consistency across attributes such as:

  • countries and actors
  • trade objects (commodities)
  • numerical indicators
  • time references

Dataset Structure

Data Instances

Each row contains:

  • pair_id: unique identifier
  • document_id_1, document_id_2: document identifiers
  • summary_1, summary_2: English summaries
  • label_5way: one of D, U, R, N, O
  • binary_label:
    • 1 — duplicate
    • 0 — non-duplicate
  • difficulty_tag: Easy, Hard_negative, or NA
  • confidence_score: (SilverSet only)
  • justification: optional explanation
  • split: train or test

Label Semantics

The final task is binary classification:

  • Duplicate (D) → 1
  • Non-duplicate (U, R, N, O) → 0

Additionally, a diagnostic labeling scheme is provided:

  • D — duplicate
  • U — update of the same topic but not the same event
  • R — related event
  • N — not related
  • O — other

These labels are used for error analysis, not for multi-class training.


Dataset Creation

The dataset is constructed from a large collection of trade and foreign economic news articles.

Pipeline:

  1. Articles are summarized using a domain-adapted abstractive summarization model based on mBART50 (lyutovad/mbart50-tradenewssum), designed to preserve key trade-related attributes.
  2. All summaries are translated into English using an automatic machine translation model to ensure a unified semantic space.
  3. Candidate pairs are generated using hierarchical clustering of summary embeddings under a temporal constraint: only documents within a ±7 day window are considered.
  4. Clustering is used solely for candidate generation and does not affect evaluation.
  5. Initial labels are assigned using an LLM-based annotation pipeline, comparing key trade attributes such as countries, commodities, numerical indicators, and time references.
  6. All candidate pairs are subsequently manually validated to construct the GoldSet.
  7. The SilverSet is used exclusively for training, while the GoldSet is reserved for evaluation.

Dataset Statistics

GoldSet (test)

  • 1,469 pairs
  • 41.7% duplicates
  • manually validated

SilverSet (train)

  • 16,097 pairs
  • 33.9% duplicates
  • weakly labeled with confidence scores

Uses

Direct Use

  • duplicate detection
  • reranking and pairwise classification
  • evaluation of semantic similarity models
  • training with hard negatives

Out-of-Scope Use

  • legal or financial decision-making
  • applications requiring verified factual accuracy of news content

Bias, Risks, and Limitations

  • The SilverSet contains weak labels and may include noise
  • Annotation was not fully multi-annotator for all instances
  • The dataset focuses on trade news, limiting generalization
  • Distinguishing between duplicates and updates remains inherently ambiguous

Licensing Information

This dataset is released under the Creative Commons Attribution 4.0 (CC-BY-4.0) license.

The dataset consists of derived data, including:

  • generated summaries
  • translated text
  • annotations and labels

These transformations are intended to reduce direct overlap with original news articles.

Important notice:

  • The original news articles may be subject to their own copyright and licensing restrictions
  • This dataset does not include full original articles
  • Users are responsible for ensuring compliance with applicable laws and source licenses when using the data