Paper Accepted

CtrlNews: LLM-based Multi-Agent Controllable News Writing via Knowledge Gravitational Field has been accepted to EMNLP 2025

Overview

This is an automated news generation system powered by large language models (LLMs) and a "gravitational field" approach. The system automatically generates news articles with controllable sentiment through multi-agent discussions, knowledge graph expansion, and opinion generation.

Key Features

• Multi-agent discussion to generate initial question sets
• Directed gravitational field model construction
• Automated web search for relevant information
• Multi-round knowledge graph expansion
• Sentiment-controlled opinion generation (adjustable positive/neutral/negative ratios)
• Automated article writing and optimization
• Batch generation of multi-topic news articles

Technical Architecture

• Based on DeepSeek series large language models
• DuckDuckGo search tool for information retrieval
• NewsGravityField for knowledge structure representation
• ArticleNewsWriter for article generation
• SimpleArticleReviewer for article evaluation and optimization

Installation

Prerequisites

• Python 3.8+
• Anaconda environment (recommended)

Installation Steps

1. Clone the repository
2. Create and activate a new conda environment

   conda create -n ctrlnews python=3.8
   conda activate ctrlnews

3. Install the dependencies

Usage

Configuration

In the main function, you can modify the following configuration parameters:

• topics_file: Path to the topics JSON file
• output_dir: Output directory for generated articles
• generations_per_topic: Number of articles to generate per topic
• sentiment_ratios: Sentiment ratio configuration
• max_expansion_rounds: Maximum knowledge expansion rounds

Running

python generate_controlled_wf_gra_full.py

Project Structure

./
├── data/                     # Data directory
│   ├── news_topics.json      # News topics configuration
│   └── workflow/             # Generated news articles
├── tools/                    # Tool classes
│   ├── duckduckgo_searchtool.py  # Search tool
│   └── lm.py                 # Language model tools
├── GravitionalField.py       # Gravitational field model
├── ArticleNewsWriter.py      # Article generator
├── simple_reviewer.py        # Article review component
└── generate_controlled_wf_gra_full.py  # Main program

Workflow

1. Load topics from configuration file
2. For each topic:
   • Conduct multi-agent discussion to generate initial questions
   • Build directed gravitational field and add questions
   • Search and answer questions using DuckDuckGo
   • Build gravitational connections
   • Expand knowledge graph through multiple rounds
   • Apply sentiment control to peripheral nodes
   • Generate opinions for peripheral nodes
   • Generate news article using the opinions
   • Review and optimize the article
   • Save the article and used opinions

Notes

• Ensure sufficient API access rights and quotas
• Generating numerous articles may require significant time
• Stable network connection is required for search functionality
• The sentiment ratio can be adjusted to control the opinion bias

Future Improvements

• Add more language models support
• Improve search capabilities
• Enhance article generation quality
• Add more topics and domains

License

This project is licensed under the MIT License - see the LICENSE file for details.