Computational Framing Analysis: Proposing And Applying an Unsupervised Entity-Centric Semantic Relations Approach

Abstract

This dissertation presents a novel computational approach to analyze how the news media construct frames around certain people or groups in their coverage. The approach centers on entity-centric emphasis frames, focusing on the language used to attribute key entities (e.g., shooters and victims in mass shooting incidents). The unsupervised method, named Semantic Relations-based Unsupervised Framing Analysis (SUFA), uses computational techniques to detect and analyze framing patterns based on semantic relations, moving beyond existing bag-of-words, co-occurrence, and frequency-based approaches. The dissertation includes three main projects, each building on the previous to develop, apply, and improve this new approach. Project 1 (Chapter 2) provides a critical review of existing computational methods used for supervised and unsupervised framing analysis. The review highlights limitations in traditional unsupervised approaches, which largely rely on bag-of-words, frequency, and word co-occurrence methods, often failing to capture contextual meaning and relationships between words. The survey article recommends integrating semantic relations into unsupervised framing analysis, proposing a more nuanced method for detecting frames. Project 2 (Chapter 3) builds on the recommendations of Project 1 and introduces Semantic Relations-based Unsupervised Framing Analysis (SUFA) as a new computational framing approach to explore entity-centric emphasis frames. This chapter presents a mixed-method study consisting of qualitative textual analysis and computational analysis applied to 100 news reports (600 paragraphs) on the Uvalde school mass shooting from four major U.S. media outlets, The New York Times, Cable News Network (CNN), Wall Street Journal, and Fox News. The qualitative analysis identifies how semantic relations contribute to frame construction, while the computational analysis employs natural language processing techniques, including dependency parsing, to extract and analyze entity-centric frames (e.g., shooter, victims, incident). The study outlines the strengths, limitations, and practical applications of SUFA, demonstrating its potential as a scalable and context-aware framing analysis approach. Project 3 (Chapter 4) applies SUFA to a large-scale dataset of gun violence coverage in the United States and advances the methodological approach by automating the formation of frames from framing components. This study analyzes one month of news reports (N = 1334) from nine major U.S. news outlets covering the 2022 Uvalde elementary school mass shooting incident in Texas. Three of the outlets were selected from the left-centered bias category: The New York Times (n=227), The Washington Post (n=228), and The USA Today (n=128). Three were selected from the right-centered bias category: The Wall Street Journal (n=47), The New York Post (n=255), and The Dallas Morning News (n=155). And three others were included from the least-biased category: The Hill (n=227), The Indianapolis Star (n=39), and The Des Moines Register (n=28). The media outlets’ biases were determined by scores provided by Media Bias/Fact Check (MBFC), a non-partisan and independent site that provides bias scores for media outlets. The research under Project 3 goes beyond existing topic modeling approaches, which primarily rely on bag-of-words, co-occurrence, and frequency-based methods that often fail to capture semantic relationships between words. Instead, SUFA leverages advanced NLP techniques such as dependency parsing and coreference resolution to identify how words modify or relate to key entities (e.g., shooter, victims). Furthermore, large language models (LLMs) such as OpenAI’s GPT-4o are incorporated to automate the clustering of framing components. This advances SUFA and improves its scalability for unsupervised frame detection. At the same time, it demonstrates how LLMs can be utilized in the coding of textual data and the clustering of framing components. This chapter provides how SUFA’s entity-centric emphasis framing focus offers deeper insights into media narratives and bias, particularly in the framing of shooters and victims in a mass shooting incident. The results revealed that at least six frames were attributed to the shooter, and nine frames were attributed to the victims in different ways. The right-centered news media group deployed some frames, including action attribution, younger age, and allegation certainty, significantly higher compared to left-centered ones, to frame the shooter. As the regression analysis provides, left-centered media are significantly more likely to deploy personalized victim framing (our victims, your victims) and emphasize older victims. In contrast, right-centered media more frequently use dehumanization framing. At the same time, there is also a significant difference in how right-leaning and left-leaning news outlets use individual framing components. These results highlight news media’s ideological differences in how they deploy framing components and frames to attribute to victims and the shooter. Guided by framing and attribution theory, the exploration of frames provides theoretical insights into how media frames assign responsibility for mass shootings through international or external attributions. For instance, one prominent frame, action attribution, highlights the shooter’s agency and responsibility, aligning with internal attribution and frequently used by right-leaning media. This dissertation makes several important contributions to computational framing research. It advances the SUFA approach by applying it to a large-scale dataset and enhancing its methodological rigor through the integration of semantic relations, dependency parsing, coreference resolution, and large language models (LLMs) for automated clustering. The study applies SUFA to mass shooting coverage, offering one of the first large-scale, unsupervised analyses of shooter and victim frames across media bias groups. It bridges attribution theory with computational methods, demonstrating how internal and external attributions of responsibility are reflected in media frames. Additionally, two annotated mass shooting datasets, validated through human and GPT comparison, provide a valuable resource for future research. This dissertation also shows SUFA’s interdisciplinary potential, showing how it can support framing analysis across fields such as communication, political science, and computer science. In terms of implications, SUFA can be potentially used as a powerful tool for real-time media monitoring and enhanced social media analytics by adding framing-based insights that move beyond surface-level metrics like sentiment analysis and mention frequency. It can support crisis communication by helping crisis practitioners understand how key entities are framed during crises, informing more effective response strategies. Computational- or data-driven journalists can use SUFA to uncover and visualize media bias and framing trends. The general public, educators, and activists can also apply the approach to strengthen their media literacy and, expectedly, hold media outlets accountable for biased or misleading representations. Overall, this dissertation develops, applies, and advances a computational framing analysis approach grounded in semantic relations to explore entity-centric emphasis frames. Although SUFA has some methodological limitations, including its primary reliance on textual framing components and its focus on entity-centric frames, its successful application across multiple datasets underscores its potential as a powerful tool for computational media analysis, bridging the gap between social science theories and advanced computational methods. Keywords: Computational framing analysis, natural language processing, machine learning, dependency parsing, semantic relations, method, attribution theory, framing, gun violence, mass shootings, public health crisis, computational strategic communication, social media analytics, media monitoring