Abstract
Zeyu Dai, Himanshu Taneja, Ruihong Huang. Proceedings of the Workshop Events and Stories in the News 2018. 2018.
This paper introduces a novel NLP task: categorizing news articles based on their underlying narrative structures (such as the "Inverted Pyramid" or "Kabob") rather than just their topics or authors. The authors define a set of foundational "news elements," map them to four common journalistic structures, and introduce a newly annotated dataset of 853 articles. Using an SVM classifier with custom structural and stylistic features, they demonstrate that algorithms can successfully identify these macro-level news genres.
Introduction
Historically, automated text categorization has focused on classifying documents by topic (e.g., sports vs. politics), author, or general sentiment. However, journalists are trained to organize information using specific structural templates. For example, breaking news often uses an "Inverted Pyramid" (most critical facts first, followed by minor details), while feature stories might use a "Kabob" (starting with a catchy anecdote, transitioning to the main point, and ending with broad details).
Recognizing this macro-structure is highly beneficial for downstream NLP tasks like text summarization, discourse parsing, and information extraction. If an AI knows an article is written as a chronological narrative, it can better extract timelines. Despite its utility, there were previously no datasets or computational models designed to classify news based on these fine-grained content organization structures.
Methodology
The authors tackle this by formalizing journalistic theory into a machine learning framework. They break articles down into five functional building blocks based on their role and writing style: Standard Lede (direct introduction), Image Lede (anecdotal hook), Synopsis (summary bridge), Narration (chronological details), and Body Section (supplementary facts). These elements are sequenced to form four primary news structures.
Classification Model and Features
The authors created a first-of-its-kind annotated dataset of 853 articles from the New York Times section of the English Gigaword corpus, spanning politics, crime, business, and disasters. They trained a multi-class Support Vector Machine (SVM) to predict the structure of an article using several feature sets:
- N-grams: Standard unigrams and bigrams.
- Writing Style Features: Syntactic production rules and semantic categories (using the LIWC dictionary) to differentiate narrative from expository writing.
- Key Event Placement (KEP): A feature measuring word overlap between the article's title and its first paragraph to detect structures that introduce key events immediately (like the Inverted Pyramid).
Results and Discussion
The experimental results highlight the importance of structural and stylistic features over simple bag-of-words approaches.
As shown in the table above, the best SVM model (using all features) achieved a Micro F1 score of 71% and a Macro F1 score of 60.2%. Notably, using only N-grams completely failed to identify the "Martini Glass" structure (F1 of 0). Adding Writing Style features significantly improved the detection of Narrative and Martini Glass structures, boosting Macro F1 from 46.7% to 57.3%.
The dataset revealed a heavy imbalance in real-world news: the "Inverted Pyramid" is overwhelmingly the most common (482 articles), while the "Martini Glass" is quite rare (37 articles), making it the most difficult to classify even with engineered features (best F1 of 19.1%).
Conclusion
This research demonstrates that news can be systematically categorized by its macro-structure, independent of topic. While standard NLP features are insufficient for nuanced structures, incorporating syntax and information placement allows for effective classification. Future work will investigate how these macro-level structures interact with internal event ordering to further bridge the gap between document classification and information extraction.
Summary updated