In some sense, news is probably never free from the agencies' subjective valuation and external forces such as owners and advertisers. As a result, the perspective of news content may be biased. To clarify such a bias, we propose a novel method to extract characteristic descriptions on a certain entity (person, location, organization, etc.) in articles of a news agency For a given entity, a description is one tuple (called SVO tuple) that consists of that entity and the other words or phrases appearing in the same sentence on the basis of their SVO (Subject(S), Verb(V) and Object(O)) roles. By computing the frequency and inverse agency frequency of each description, we extract the characteristic description on a certain entity. Intuitively a SVO tuple, which is often used by the news agency but not commonly used by the others, has high probability of being of a characteristic description. To validate our method, we carried out an experiment to extract characteristic descriptions on persons by using articles from three well-known Japanese newspaper agencies. The experimental results show that our method can elucidate the different features of each agency's writing style. We discuss the useful application using our method and further work.