From Text to Context: An Entailment Approach for News Stakeholder Classification

TL;DR

This work tackles automatic stakeholder classification in news articles by reframing the problem as natural language inference (NLI) to enable zero-shot reasoning over unseen stakeholder classes. It constructs enriched entity representations using within-document and cross-document context, augmented with Wikipedia background knowledge, and trains a domain-agnostic classifier on a weakly supervised entailment dataset. The approach leverages two architectures (RoBERTa and BART) and demonstrates competitive performance on seen labels while achieving gains in zero-shot scenarios, with robustness improved via prompt design and P-tuning. The methodology supports broader analysis of news narratives and could inform bias detection and context-aware journalism research across diverse policy domains.

Abstract

Navigating the complex landscape of news articles involves understanding the various actors or entities involved, referred to as news stakeholders. These stakeholders, ranging from policymakers to opposition figures, citizens, and more, play pivotal roles in shaping news narratives. Recognizing their stakeholder types, reflecting their roles, political alignments, social standing, and more, is paramount for a nuanced comprehension of news content. Despite existing works focusing on salient entity extraction, coverage variations, and political affiliations through social media data, the automated detection of stakeholder roles within news content remains an underexplored domain. In this paper, we bridge this gap by introducing an effective approach to classify stakeholder types in news articles. Our method involves transforming the stakeholder classification problem into a natural language inference task, utilizing contextual information from news articles and external knowledge to enhance the accuracy of stakeholder type detection. Moreover, our proposed model showcases efficacy in zero-shot settings, further extending its applicability to diverse news contexts.

Figures (2)

• Figure 1: Illustration of the entailment approach for zero-shot stakeholder classification. The left-hand side depicts how the model is trained on entailment task, and on the right-hand side, we demonstrate how the fine-tuned model predicts new stakeholder classes for the query entity.
• Figure 2: Zero-shot classification Performance of RoBERTa, RoBERTa+P-tuning and bart-large-mnli model (from Facebook) on different hypothesis templates. Here, Template1: The entity {placeholder phrase} is {placeholder label}; Template2: The entity {placeholder phrase} is of stakeholder type {placeholder label}; and Original Template refers to the template mentioned in the Table 1.