AngleKindling: Supporting Journalistic Angle Ideation with Large Language Models

Document Title

AngleKindling: Supporting Journalistic Angle Ideation with Large Language Models

Document Information

• Domain: Human-Computer Interaction, Journalism, and Large Language Models
• Keywords: Journalism, Creative Generation, Idea Divergence, Large Language Models, Generative AI, Natural Language Processing, Tool Design, Historical Context Integration

Research Background and Problem

• Problems and Challenges:

  1. Journalists often use documents (e.g., press releases) to find story inspiration, but these source materials tend to contain biases, requiring in-depth analysis and interpretation.
  2. Current journalist tools typically support news discovery (CND) but lack support for divergent angle generation from documents.
  3. Time and resource constraints make it difficult for journalists to comprehensively explore multiple angles within a document.
  4. Traditional methods for generating story angles are cognitively demanding, and inspiration is often limited by the journalist's biases or knowledge constraints.

• Importance: Analyzing press releases from multiple angles is a critical foundation for high-quality journalism. This approach not only reveals hidden social controversies and potential issues but also clarifies directions for interviews and information gathering.

• Research Motivation and Related Work:

  1. Large Language Models (LLMs) have shown tremendous potential in text generation and creative support tasks in recent years, but their application in divergent reporting within journalism remains underexplored.
  2. Existing tools (e.g., INJECT) focus more on searching and categorizing news leads, with less emphasis on generating specific angles deeply connected to the text.

Solution

• Method or Solution:
  The authors developed an interactive tool, AngleKindling, leveraging the commonsense reasoning capabilities of large language models (GPT-3) to support journalists in generating divergent reporting angles from press releases.

• Innovations:

  1. Directly using large language models to generate angles associated with press releases, offering in-depth and specific controversial viewpoints or questions.
  2. Linking generated content to source documents through technical methods, enhancing the credibility of angle generation.
  3. Providing historical context for generated angles to enrich background information.
  4. Categorizing journalistic angles into major points, controversies, negative outcomes, and unexplored questions, supporting creative generation from multiple dimensions.

• Implementation Steps and Techniques:

  1. Using GPT-3 to segment and parse press releases, generating major points and different angles (e.g., controversies, investigative points, and negative outcomes).
  2. Utilizing Sentence-BERT to semantically match generated angles with sentences in the press release, enabling journalists to quickly verify and reference them.
  3. Integrating the New York Times database to provide related historical articles for different angles, while extracting keywords and critical information from the articles.
  4. Designing a Flask-based front-end interface that displays generated angles and background information in the left-hand toolbar, visually linked to the press release content on the right-hand side.

Research Findings

• Specific Findings:

  1. AngleKindling was proven to be a more helpful creative support system compared to existing tools (e.g., INJECT).
  2. Experiments showed that participants felt AngleKindling significantly reduced cognitive load during creative generation and helped them quickly grasp the content and potential reporting angles of press releases.
  3. The diverse angles provided by the system were suitable for both short-term content creation (e.g., same-day reporting) and long-term investigative reporting requiring deeper exploration.

• Comparative Advantages over Existing Solutions:

  1. AngleKindling provides more specific and direct angles closely tied to press release content, whereas INJECT primarily relies on existing news articles, focusing on associative reporting.
  2. The system's built-in functionality for generating controversial and negative outcome angles enables journalists to more effectively uncover deeper meanings hidden in press releases.

• Experimental or Evaluation Results:

  1. The experiment involved 12 professional journalists, and after comparing AngleKindling with INJECT, all participants preferred AngleKindling.
  2. AngleKindling demonstrated significant effectiveness in reducing cognitive load and inspiring various types of story ideas, with the "controversy points" and "negative outcomes" features being the most popular.
  3. While historical context information was helpful, the relevance of the articles provided was occasionally low, requiring further optimization.

• Limitations and Future Directions:

  1. The current GPT-3 model may generate overly generic or irrelevant angles; future improvements could focus on better prompts and semantic segmentation strategies.
  2. Press release summaries lack emphasis on specific statistical data and implementation details; future enhancements could involve models with longer input lengths or customized algorithms.
  3. Exploring ways to help journalists filter and evaluate the "most promising" reporting angles, incorporating data or personalized user needs to rank generated results.
  4. While the tool currently focuses on journalism, expanding its application to other high-text-density domains (e.g., court rulings or academic papers) is a potential research direction, particularly in using LLMs to discuss ethical implications and potential societal consequences of decisions in these fields.