Abstract
Researchers Prafulla Kumar Choubey and Ruihong Huang (EMNLP 2021) present a novel actor-critic framework for news discourse profiling. By explicitly modeling articles into subtopics using journalistic structures like the inverted pyramid, the model achieves state-of-the-art performance in identifying the functional roles of sentences within news stories.
Introduction: The Challenge of News Discourse Profiling
In Natural Language Processing (NLP), news discourse profiling is the task of categorizing sentences based on their function—distinguishing between a Main Event, Historical Context, or an Expectation of future consequences. Understanding these roles is vital for automated summarization and event extraction.
However, current models often struggle by treating documents as flat sequences or simple hierarchies. This approach loses the nuance of journalistic organization. News articles follow specific patterns, such as the "inverted pyramid" (prioritizing the most important information) or topical shifts, which provide essential context for determining a sentence's role.
Methodology: The Actor-Critic Framework
The authors propose a sophisticated Actor-Critic framework designed to induce latent subtopic structures:
- The Actor ($f_T$): A pointer decoder network that identifies sentences representing subtopic boundaries.
- The Critics: Multiple "critics" guide the actor based on known journalistic structures:
- TextTiling: Segments text based on lexical shifts.
- Locally Inverted Pyramid: Identifies boundaries when content relevance suddenly increases (e.g., shifting from background back to core events).
- Training Strategy: A hybrid of Reinforcement Learning (RL) and Imitation Learning. If the Actor’s predicted segmentation outperforms the Critics on the final task, RL (REINFORCE) rewards the Actor. Otherwise, the Actor is forced to imitate the Critic via cross-entropy loss.
- The Profiler ($f_C$): A hierarchical neural network that uses these subtopics to create local representations for sentence classification.
Datasets and Evaluation
The model was evaluated on the NewsDiscourse Corpus, containing 802 English news articles across domains like business, crime, and politics. Each sentence is annotated with one of eight roles: Main Event, Consequence, Previous Events, Current Context, Historical Event, Anecdotal Event, Evaluation, and Expectation.
Key Results
The proposed model, specifically the version utilizing both TextTiling and Inverted Pyramid critics (RL-IP/TT), significantly outperformed standard hierarchical baselines:
- Macro F1 Score: Improved by 2.6 points (from 54.4 to 57.0).
- Micro F1 Score: Improved by 1.3 points (from 60.9 to 62.2).
- Specific Roles: The model showed particularly strong gains in identifying Historical Events (D1) and Anecdotal Events (D2), which are traditionally difficult to distinguish from the main narrative without clear subtopic boundaries.
Discussion and Insights
The study highlights that a sentence's role is often determined by its immediate subtopical neighborhood rather than the overall document theme. By successfully converting classical journalism theories into effective "critics," the authors demonstrate that modern deep learning models can be significantly enhanced by domain-specific heuristics. The flexibility of using multiple critics allows the model to adapt its organizational strategy to the specific document at hand.
Future work may explore incorporating other discourse structures, such as rhetorical relations (e.g., Contrast, Elaboration), or expanding the framework to handle non-news genres that do not follow the inverted pyramid structure.
Summary updated