Extraction of Characteristic Description for Analyzing News Agencies
Ishida, Shin, Ma, Qiang, Yoshikawa, Masatoshi, Journal of Digital Information Management
An increasing number of people are reading news articles on the Web these days. News articles are widely available on the Web from diverse sources of information, and the main ones being news agencies. However, the bias of news agencies is sometimes reflected in the way news stories are reported, raising the possibility that readers may be mislead by what they read.
There are many services and a large amount of published research that deal with the news bias issue. As an example of the former, FairSpin  is a service that shows readers whether an article is biased or not. Readers using this services can vote on whether they perceive an article to be slanted left or right and see the aggregate results of the poll. Services like this are useful as a means of providing an understanding of bias appearing in news articles, but they do not to any significant degree consider the salient characteristics of news agencies, such as that agencies usually describe certain persons in a rather negative way. For instance, in the issues related with America and Iraq, such as Iraq war, P news agency's descriptions are often affirmative to America, while those of Q news agency are frequently affirmative to Iraq We attempted to overcome this deficiency by ascertaining the characteristics of news agencies to enable readers to obtain a more comprehensive understanding of news articles.
One of the features or characteristics of news agencies may be seen in the way they use words in their articles, i.e, their writing style. In particular, they tend to show this in their descriptions of certain entities (persons, locations, organizations, etc). We define such descriptions as "characteristic descriptions" and propose a method of extracting them.
To extract characteristic descriptions, we analyze words or phrases that appear in the same sentence on the basis of whether they are a subject (S), verb (V), or object (O). When a given entity appears in a sentence as S or O, the words playing the other roles (i.e., O and V if the entity is S, or S and V if it is O) are defined as "descriptions" of the given entity. We define such descriptions as "SVO tuples".
We extract characteristic descriptions from two aspects: local features and global features.
* Local Feature
When a certain description often appears in a news agency report, it tends to represent the way that the agency typically reports news. We define this tendency as a news agency's "local feature".
* Global Feature
When a certain description rarely appears in other news agency reports, this also tends to represent the way that a certain agency typically reports news. We define this tendency as a news agency's "global feature".
We also present characteristic descriptions of each news agency chronologically. Using our presentation method, users can understand that there are two major differences in the way that news agencies report stories, i.e., "inter differences" and "intra differences".
* Inter differences
Differences in opinion between competing news agencies. These can be ascertained by comparing characteristic descriptions among news agencies.
* Intra differences
Variations in opinion that vary over time between competing news agencies. These can be ascertained by comparing characteristic descriptions chronologically.
To estimate the effectiveness of our method, we carried out experiments using articles published by three major Japanese newspaper agencies, i.e. the Asahi Shinbun, the Yomiuri Shinbun, and the Mainichi Shinbun (Readers' note: "Shinbun" is the Japanese word for "newspaper" or "newspaper company"). From these articles we extracted characteristic descriptions of certain persons and organizations and evaluated them. The experimental results we obtained showed that our method can be used to elucidate the features of each agency. …