The bias in the news media is an inherent flaw of the news production process. The resulting bias often causes a sharp increase in political polarization and in the cost of conflict on social issues such as the Iraq war. It is very difficult, if not impossible, for readers to have penetrating views on realities against such bias. The problem has been extensively studied considering its serious and chronic impact. News producers put much effort to avoid the creation of bias at the production stage, e.g., journalism ethics and standards and adversarial formats of reporting such as point-counterpoint roundtable discussions. Media experts and researchers make efforts such as frame analysis to deal with the created biases of the produced contents. However, media bias is still widespread.
This thesis investigates the media bias problem from a computational perspective and develops a practical solution. We present the approach media bias mitigation, reducing the effects of bias by making readers themselves to actively overcome biased views. The thesis aims to establish a framework providing readers with tools for active interaction with which they easily discover and compare diversity of existing biased views. For this, we develop the bias mitigation approach for three important news article domains: straight news articles, news articles on contentious issues, and political news articles.
The first part of the thesis presents aspect-level news browsing as a solution to mitigate bias in straight news articles. Aspect-level news browsing provides readers with a classified view of news articles on the same event with different viewpoints. It effectively helps readers understand the event from a plural of viewpoints and formulate their own, more balanced viewpoints free from specific biased views. Realizing aspect-level browsing raises important challenges, mainly due to the lack of semantic knowledge with which to abstract and classify the intended salient aspects of articles. We first demonstrate the feasibility of aspect-level news browsing through user studies. We then deeply look into the news article production process and develop news structure-based extraction and framing cycle-aware clustering. The evaluation results show that the developed method performs classification more accurately than other methods.
The second part develops and evaluates NewsCube, a fully automatic news system for aspect-level browsing. NewsCube automatically creates and promptly provides readers with multiple classified viewpoints on a news event of interest. NewsCube effectively achieves its goal by classifying aspects, presenting them fairly, and recommending articles with contrasting aspects. NewsCube quickly processes a large amount of news articles and keeps up the fast news production cycle. We discuss the effect of the service through various user studies. The studies demonstrate the potential of NewsCube for bias mitigation. The system influences users’ news reading behavior and opinions.
The coverage of contentious issues is frequently biased and fails to fairly deliver conflicting arguments of the issue. Contentious issues such as the ‘4 river project’ and ‘Cheonan sinking incident’ continuously arise in various domains, such as politics, economy, environment. Each issue involves various dispute topics and complex arguments. The thesis develops a method to provide readers with opposing views of contentious issues. We observe that the disputants of a contention are an important feature for understanding the discourse. It performs unsupervised classification on news articles based on disputant relations, and helps readers intuitively view the articles through the opponent-based frame.
The last part of the thesis presents a novel social annotation analysis approach for identification of news articles’ political orientation. Political views frequently conflict in the coverage of contentious political issues. The approach focuses on the behavior of individual commenters. It uncovers commenters’ sentiment patterns towards political news articles, and predicts the political orientation from the sentiments expressed in the comments. It takes advantage of commenters’ participation as well as their knowledge and intelligence condensed in the sentiment of comments, thereby greatly reduces the high complexity of political view identification. The thesis conducts extensive study on commenters’ behaviors and discovers predictive commenters showing a high degree of regularity in their sentiment patterns. We develop and evaluate sentiment pattern-based methods for political view identification.
매체 편향은 뉴스 생산과정의 본질적 한계로서, 현실을 전달하는 과정에서 주관이 개입되는 현상이다. 매체의 편향 현상은 뉴스 수집, 기사 작성, 편집을 포함한 뉴스 생산의 모든 단계에서 가치 판단이 개입됨으로 인해 피할 수 없는 현상이다. 뉴스 생산자들은 각기 다른 사건을 기사화하며, 다른 논조로 기술하며, 또 다른 방식의 편집을 수행한다. 종전의 뉴스 소비 환경에서 독자들이 손 쉽게 다른 뉴스 생산자들의 관점을 파악하고 종합적인 이해를 갖는 것은 매우 어렵다. 따라서 편향된 생산의 결과로 나온 매체의 편향된 보도는 자주 극단적 정치적 대립이나 중요 사회 이슈에 대한 갈등 비용의 상승으로 이어진다.
본 논문은 매체 편향의 문제를 전산학적인 관점에서 접근하여 실용적인 방법을 제시한다. 본 논문은 ‘매체 편향 완화’ 방법을 제안하고, 이를 구현하는 전산학적 프레임워크를 기술한다. 매체 편향 완화 기법은 뉴스 소비자가 여러 생산자의 관점을 비교-대조할 수 있게 하여 편향된 관점의 영향을 완화 하는 것이다. 전산학적 프레임워크는 세가지 주요 기사 영역에 대해 자동으로 관점의 다양성을 포착, 분류 및 제시하는 기법을 구현한다. 제안된 프레임워크는 사용자가 사건을 보다 다양한 관점에서 바라보고 보다 종합적인 이해를 가질 수 있게 도우며 편향된 특정 관점을 극복하고 보다 균형적인 관점을 가질 수 있도록 유도한다.