Loading...
Profiling News Media with Graphs and Advancing Empathetic Language Models
Manzoor, Muhammad Arslan
Manzoor, Muhammad Arslan
Files
Author
Supervisor
Department
Natural Language Processing
Embargo End Date
30/05/2025
Type
Dissertation
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
The news media exerts a lot of influence via mass reporting and an extensive social media presence, reaching millions of readers and followers. Analyzing these sources is essential for promoting social wellbeing, particularly by raising awareness among readers about the existence of political bias and varying levels of factuality in the reporting. The rapid expansion of media sources and the constraints of manual evaluation have led to research on automated analysis of news media sources with the aim of detecting their political bias and factuality of reporting. Most prior work has focused on analyzing the textual content of the articles published by the news outlet. However, various challenges, such as text noise, restrictions on scraping content from media websites, and the absence of explicit relationships between news outlets, their published content, and their audiences have prompted researchers to explore other sources of information that go beyond the content of the articles the news outlet has published. In particular, some work has explored the inherent relationships between media sources (e.g., citation, audience overlap), modeled as graphs. These graphs allow graph neural networks (GNNs) to reason over connections, leveraging structured patterns to enhance media profiling capabilities. Media graphs offer several advantages, such as efficiency, reduced dependence on cumbersome feature collection processes, and realtime profiling. While effective, these techniques also present challenges, including label sparsity and disconnected components, which hinder GNNs from capturing longrange dependencies and limit their ability to learn expressive node representations. To address these limitations, Media Graph Mind (MGM) is proposed as a robust solution to utilize features, structural patterns, and label information from structurally similar nodes throughout the media graph. This not only enables GNNs to capture long-range dependencies to learn more expressive node representations, but also enhances language models (LMs) by integrating structural information, ultimately improving the performance of both models. Furthermore, I introduce an expanded set of news media labels to address the challenge of label scarcity and construct multi-dimensional representations that encompass media articles, Wikipedia descriptions, and novel approaches for generating media graphs to enhance diversity and informativeness. Graphs include Alexa graphs, hyperlink graphs, and large language model (LLM)-generated graphs. I present a comprehensive framework for systematically analyzing and evaluating these representations in various combinations, assessing their significance and relevance for factuality and political bias detection tasks. Moreover, as users are increasingly relying on LLMs as a source for consuming news, it is important to ensure that the generated content is context-aware, reflects user preferences while being mindful of existing media biases. This would enable an LLM to have a balanced and informed understanding of media, allowing it to produce content that resonates empathetically with users. Toward this goal, I propose and evaluate strategies such as contrastive learning, gold label-guided explanation, supervised fine-tuning, and Chain-of-Thought to enhance LLMs’ ability to recognize nuanced emotional expressions. This research lays the foundation for developing LLMs that are capable of understanding the news media and of generating content that resonates both ideologically and emotionally with users.
Citation
Muhammad Arslan Manzoor, “Profiling News Media with Graphs and Advancing Empathetic Language Models,” Doctor of Philosophy thesis, Natural Language Processing, MBZUAI, 2025.
Source
Conference
Keywords
News Media Profiling, Media Graphs, Political Bias and Factuality of News Media, Empathic Understanding, Enhancing Empathy in Language Models