Uncovering the essence of diverse media biases from the semantic embedding space

The proliferation of news media in the digital age underscores its crucial role in shaping public opinion. However, the pervasive presence of media bias, or slanted news coverage, poses a significant challenge. This bias can heavily skew public perception, potentially leading to severe social consequences. Examples, such as the disproportionate representation of management over workers in news coverage of strikes, highlight the detrimental effects of biased reporting. Media bias encompasses various forms, including event selection, tone, framing, and word choice. News organizations inherently select which events to cover, leading to a perception of bias, whether intentional or unintentional. News values, though offering a framework for understanding news selection, vary across organizations, resulting in skewed coverage, especially regarding underrepresented topics, like women's issues. Once an event is selected, the choice of tone, framing, and words employed introduces further bias. The same event can be presented dramatically differently depending on the media outlet's perspective. Furthermore, media bias is multifaceted, influenced by factors such as geographic location, media position, editorial guidelines, political ideology, business reasons, and even personal factors. These complex interactions make the emergence of bias inevitable, threatening objective judgment and potentially exacerbating social prejudices.

Existing research on media bias spans various disciplines. Social science studies, often qualitative, analyze opinions expressed in editorials or identify biased instances through human annotation, a process that is labor-intensive and subjective. Quantitative approaches often count keyword frequencies, while some automated methods rely on text similarity and sentiment analysis, but these are typically limited to specific bias types. Computer science research on social media is extensive, yet few methods specifically target media bias, and those available often focus on single bias types. Natural Language Processing (NLP) research on bias in pre-trained models is relevant, revealing that these models often reflect and amplify existing human biases. However, these studies primarily focus on the models themselves, not directly on media bias analysis. A key challenge is the subjective nature of bias evaluation; what one person sees as neutral may appear biased to another. Addressing this necessitates an objective and comprehensive framework.

This study addresses the limitations of previous approaches by proposing a novel framework that leverages embedding techniques from NLP. The framework analyzes media bias from two perspectives: macro (event selection bias) and micro (bias in word choice and sentence structure). The macro-level analysis employs two datasets: the GDELT Mention Table (containing over 8 million event records) and MediaCloud (yielding over 1.2 million news articles from 12 mainstream US news outlets). From GDELT, a "media-event" matrix is constructed, quantifying how often each media outlet reports on each event. Latent Semantic Analysis (LSA), using Truncated Singular Value Decomposition (Truncated SVD), generates media embeddings. These embeddings represent each media outlet as a vector in a high-dimensional space; outlets with similar event selection biases cluster together. Word Mover's Distance (WMD) measures the similarity between these embeddings. The micro-level analysis uses the MediaCloud dataset. A Word2Vec model is pre-trained on the combined corpora of all news outlets and then fine-tuned separately for each outlet. This allows for the quantification of each outlet's bias using a method inspired by Semantic Differential, comparing embedding distances between a target word (e.g., "scientist") and sets of words with opposing semantics (e.g., male-related vs. female-related words). The cosine similarity function measures the similarity between word embeddings. Up-sampling ensures equal corpus size across outlets for the Word2Vec model training. The framework integrates both macro and micro-level analyses, providing a comprehensive assessment of media bias.

The study's results reveal several key findings. First, media outlets exhibit significant clustering based on geographical location and organizational affiliation. Media from the same country tend to group together, reflecting a regional bias in event selection. International news agencies like AP and Reuters, however, cluster together due to their global coverage. Secondly, international events such as the Russia-Ukraine conflict significantly impact event selection bias. During the conflict's peak, media from various countries showed increased similarity in their event coverage, converging around the conflict's narrative. This effect lessened as the conflict became more normalized in media cycles. Thirdly, the analysis of US news outlets demonstrated diverse biases across different topics (gender, income, political affiliation). For gender bias related to occupation, the study found that certain occupations were consistently associated with specific genders, often reflecting real-world gender ratios but potentially reinforcing stereotypes. Analyzing income bias across races and ethnicities largely reflected income disparities in the US, though anomalies such as ESPN's coverage (attributable to its focus on sports popular with specific demographics) exist. Examining political bias concerning US states showed a correlation between state political leanings (red vs. blue states) and media coverage, but also exhibited some anomalies possibly due to the timeframe of the data (Trump’s presidency influencing coverage). The use of up-sampling introduced randomness, but repeated experiments with different random seeds did not yield significantly different results across the topics analyzed, suggesting robustness of the methodology.

The study's findings contribute significantly to the understanding and quantification of media bias. The integrated framework objectively assesses bias from both macro and micro perspectives, overcoming limitations of previous methods. The results demonstrate that media bias is geographically influenced and responsive to global events, with international events leading to convergence in coverage. The analysis of US news outlets highlights the diversity of bias across topics but also reveals the perpetuation of stereotypes. The consistency between observed biases and real-world statistics underscores the potential of media to reinforce existing biases. The framework's ability to identify both event selection bias and wording bias offers a more holistic view.

This study presents a novel, data-driven framework for analyzing media bias using semantic embedding techniques. The framework's effectiveness is demonstrated through the analysis of large datasets, revealing regional variations in bias, the impact of significant global events on coverage, and the presence of diverse and persistent biases in US news outlets. Future research could focus on enhancing the interpretability of media embeddings, developing more sophisticated methods for handling complex semantic relationships, and expanding the application of this framework to other forms of media bias.

The study acknowledges several limitations. Interpreting the continuous numerical vectors of media embeddings could be enhanced. The event selection bias analysis focuses on relative topic coverage. For complex topics, bias estimation using antonym pairs might not always be sufficient. The datasets used might not represent the entirety of global media, impacting generalizability. Refining the selection and interpretation of word embeddings could also enhance accuracy. Future work should aim to mitigate these limitations for a more robust assessment of media bias.