There are two important problems in improving text classification systems based on machine learning approach. The first one, called "selection problem", is how to select a minimum number of informative documents from a given document collection. The second one, called "composition problem", is how to reorganize selected training documents so that they can fit an adopted learning method. The former problem is addressed in "active learning" algorithms, and the latter is discussed in "boosting" algorithms. This paper proposes a new learning method, called AdaBUS, which proactively solves the above problems in the context of Naive Bayes classification systems. The proposed method constructs more accurate classification hypothesis by increasing the valiance in "weak" hypotheses that determine the final classification hypothesis. Consequently, the proposed algorithm yields perturbation effect makes the boosting algorithm work properly. Through the empirical experiment using the Routers-21578 document collection, we show that the AdaBUS algorithm more significantly improves the Naive Bayes-based classification system than other conventional learning methodson system than other conventional learning methods