notnews

A Python library for classifying news articles using URL patterns, trained US and UK models, or optional large language models.

Features

The package supports three classification methods:

• URL pattern analysis
• trained models for US and UK news
• optional Claude or OpenAI classification

The Python API and Click CLI expose the same US and UK classifiers.

Streamlit Demo: https://notnews-notnews-streamlitstreamlit-app-u8j3a6.streamlit.app/

Quick Start

Python API

import pandas as pd
import notnews

# Load your data
df = pd.read_csv("news_articles.csv")

# Method 1: URL Pattern Classification (fastest)
df_url = notnews.classify_by_url(df, url_col="url", region="us")
print(df_url[["url", "hard_news", "soft_news"]].head())

# Method 2: ML Model Prediction (most accurate)
df_ml = notnews.predict_soft_news(df, text_col="text", region="us")
print(df_ml[["text", "prob_soft_news_us"]].head())

# Method 3: LLM Classification (most flexible)
# Requires ANTHROPIC_API_KEY or OPENAI_API_KEY environment variable
df_llm = notnews.classify_with_llm(df, text_col="text", provider="claude")
print(df_llm[["text", "llm_category", "llm_confidence"]].head())

# Detailed Categories (US only)
df_categories = notnews.predict_news_category(df, text_col="text")
print(df_categories[["text", "pred_category", "prob_soft_news"]].head())

Command Line Interface

# Install the package
pip install notnews
# or with uv
uv add notnews

# URL pattern classification
notnews classify-urls articles.csv --region us --output results.csv

# ML model prediction  
notnews predict-ml articles.csv --region uk --text-col content

# LLM classification
notnews classify-llm articles.csv --provider claude --api-key your_key

# Run all methods together
notnews classify-all articles.csv --region us

# Get help
notnews --help
notnews classify-urls --help

Installation

Standard Installation

pip install notnews

Fast Installation with UV

uv add notnews

Requirements

• Python: 3.11, 3.12, or 3.13
• Core: pandas, numpy, pyarrow
• Web: requests, beautifulsoup4
• CLI: click 8.0+
• Optional: anthropic, openai (for LLM classification)

LLM Setup

For LLM classification, set your API key:

# For Claude
export ANTHROPIC_API_KEY="your_key_here"

# For OpenAI  
export OPENAI_API_KEY="your_key_here"

API Reference

Core Functions

classify_by_url(df, url_col="url", region="us")

Classify articles using URL pattern matching.

Args:

• df: DataFrame with articles
• url_col: Column containing URLs
• region: "us" or "uk" for region-specific patterns

Returns: DataFrame with hard_news and soft_news columns

predict_soft_news(df, text_col="text", region="us")

Predict soft news probability using trained ML models.

Args:

• df: DataFrame with articles
• text_col: Column containing article text
• region: "us" or "uk" for model selection

Returns: DataFrame with prob_soft_news_{region} column

classify_with_llm(df, text_col="text", provider="claude", **kwargs)

Classify articles using Large Language Models.

Args:

• df: DataFrame with articles
• text_col: Column containing article text
• provider: "claude" or "openai"
• categories: Optional custom categories dict
• api_key: Optional API key (uses env var if not provided)

Returns: DataFrame with llm_category, llm_confidence, llm_reasoning columns

Advanced Usage

# Custom LLM categories
custom_categories = {
    "breaking": {"description": "Breaking news and urgent updates"},
    "analysis": {"description": "In-depth analysis and commentary"},
    "lifestyle": {"description": "Lifestyle and entertainment content"}
}

df_custom = notnews.classify_with_llm(
    df, 
    provider="claude",
    categories=custom_categories
)

# Fetch content from URLs
content = notnews.fetch_web_content("https://example.com")

Model Information

URL Patterns

• US: Politics, economics, international affairs vs. sports, entertainment, lifestyle
• UK: Includes UK-specific patterns like "uk-news", "scottish-news"

ML Models

• US: NYT-based models trained on headline and content text
• UK: URL-based model trained on UK news outlets
• Runtime inference uses version-independent, typed Parquet exports of the original calibrated linear models.

Data Sources

• US Model: Based on NYT data
• UK Model: Based on UK news analysis

The fitted classifiers and vectorizers are published at gojiberries/notnews. Notnews downloads them from an immutable Hugging Face commit on first use. Set NOTNEWS_MODEL_DIR to use an explicitly managed local copy.

Applications

Research using notnews:

• US Soft News Analysis
• UK Soft News Analysis

Documentation

Full documentation: notnews.github.io/notnews

Contributing

We welcome contributions! Please see our Contributor Code of Conduct.

Development Setup

git clone https://github.com/notnews/notnews.git
cd notnews
uv sync --all-groups --all-extras
uv run pytest

Authors

• Suriyan Laohaprapanon
• Gaurav Sood

License

MIT License