Causal graph extraction from news: a comparative study of time-series causality learning techniques.

Area of Science:

• Computational Social Science
• Natural Language Processing
• Causal Inference

Background:

• Understanding complex scenarios and predicting events often requires identifying cause-effect relationships.
• Extracting causal graphs from news articles is challenging, especially at scale.
• Existing methods lack robust approaches for large-scale causal graph extraction from digital text media.

Purpose of the Study:

• To present a novel framework for large-scale causal graph extraction from digital text media.
• To identify topic-relevant variables and model event relationships using natural language processing and information retrieval.
• To apply causal structure learning techniques to time-series data derived from news articles.

Main Methods:

• Developed information retrieval and NLP methods to select topic-relevant variables and ongoing events.
• Utilized event-phrase embeddings to cluster similar events semantically.
• Applied causal structure learning techniques to time-series data of selected variables.
• Evaluated nine state-of-the-art and two novel ensemble causality learning methods on synthetic data.
• Validated the framework with domain experts using a New York Times dataset spanning 246 months.

Main Results:

• The framework successfully extracts causal graphs from large news datasets.
• Event-phrase embeddings effectively group semantically similar events.
• An evaluation identified promising time-series causality learning techniques for this task.
• Expert evaluation in a real-world scenario confirmed the framework's utility.

Conclusions:

• The proposed framework offers a scalable solution for causal graph extraction from news.
• It provides valuable insights into identifying key variables and learning causal relationships from time-series news data.
• This approach enhances the ability to explain and predict complex events through data-driven causal discovery.