Porkonsale/International_Trade_News_Tension_Analysis_Dataset

China Trade Sentiment Analysis Dataset

Dataset Description

This dataset provides labeled sentences from English news articles, specifically annotated for the rhetoric intensity of trade-related discourse between two key bilateral pairs:

• China-Japan
• China-US

The core label (intensity_score) quantifies the tone of trade rhetoric, enabling downstream tasks like trade tension trend analysis, sentiment classification, and cross-country discourse comparison.

Scoring Methodology (DeepSeek Prompt)

Intensity scores were generated via standardized prompting of the DeepSeek model, with strict adherence to the following instruction template to ensure annotation consistency:

Rate the intensity of rhetoric for sentences in the following English text:

1. Scoring rule: Use a continuous scale from -1 (completely mild) to 1 (extremely intense), accurate to 1 decimal place. Score based only on tone, word choice, and sentence structure (neutral statement → near -1, slight emotion → near 0, aggressive/extreme expression → near 1), independent of content truthfulness or stance.
2. The input contains complete sentences, no need to split. Each sentence corresponds to one score.

Output format requirement (strictly follow for parsing, no extra text): Sentence 1: [Input sentence content] Score 1: [Specific score, e.g., -0.2] Sentence 2: [Input sentence content] Score 2: [Specific score, e.g., 0.5] ...

Text content: {para}

Dataset Structure

Data Fields

All fields are structured with clear semantic definitions, as detailed below:

Field	Type	Description
uuid	string	Unique identifier for the parent news article
title	string	Title of the source news article
paragraph_count	int64	Total number of paragraphs in the original article
sentence	string	Individual sentence extracted from the article (annotation target)
intensity_score	float64	Rhetoric intensity score (-1 = mild, 0 = neutral, 1 = extremely intense)
source_type	string	Source of the sentence (either "title" or "paragraph")
source_index	int64	0-based index of the paragraph in the original article (for paragraph sources)
country_pair	string	Bilateral trade pair (fixed values: China-Japan / China-US)

Dataset Splits

The dataset is partitioned by country pair, with the following size metrics:

Split	Number of Examples	Total Bytes
china_japan	10,974	3,220,612
china_us	7,832	2,207,560
china-eu	28,059	10,390,939
Total	46,865	15,819,111

Usage Examples

Load Dataset via Hugging Face datasets Library

from datasets import load_dataset

# Load full dataset (default config)
dataset = load_dataset("Porkonsale/International_Trade_News_Tension_Analysis_Dataset")

# Access China-Japan trade data split
china_japan_split = dataset["china_japan"]
print(f"China-Japan samples: {len(china_japan_split)}")
print(f"Sample sentence: {china_japan_split[0]['sentence']}")
print(f"Intensity score: {china_japan_split[0]['intensity_score']}")

# Access China-US trade data split
china_us_split = dataset["china_us"]
print(f"China-US samples: {len(china_us_split)}")

Citation

If you use this dataset in your research, please cite:

@dataset{china_trade_sentiment_2025,
  title={China Trade Sentiment Analysis Dataset},
  author={Porkonsale},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/Porkonsale/International_Trade_News_Tension_Analysis_Dataset}}
}

License

This dataset is released under the Apache 2.0 License.