This paper presents a method for improving text classification by combining Expectation-Maximization (EM) and active learning, using a large pool of unlabeled documents. The goal is to reduce the need for labeled training examples while maintaining high classification accuracy. The paper introduces a pool-based active learning approach that selects the most informative examples from the entire pool of unlabeled documents, rather than generating them synthetically or selecting from a stream. This approach is combined with EM to estimate the class labels of unlabeled documents, thereby improving the classifier's performance. The paper first describes a probabilistic framework for text classification based on naive Bayes, which assumes that words in a document are independent given the class. It then explains how EM can be used to improve the classifier by estimating the class labels of unlabeled documents. Active learning is then introduced as a method to select the most informative examples for labeling, with the Query-by-Committee (QBC) algorithm being used to measure classification variance indirectly by examining the disagreement among multiple classifier variants. The paper shows that combining EM with active learning significantly reduces the number of labeled examples needed to achieve high accuracy. Specifically, the combination of QBC and EM requires only 58% as many labeled examples as EM alone and 26% as many as QBC alone. The paper also introduces a new approach called pool-leveraged sampling, which interleaves EM and active learning to further improve performance. The paper presents experimental results on real-world text datasets, such as the Newsgroups and Reuters datasets, showing that the combination of EM and active learning outperforms both methods individually. The results indicate that using a pool of unlabeled documents can significantly improve active learning by reducing the need for labeled examples and increasing classification accuracy. The paper concludes that combining EM and active learning provides a substantial benefit in text classification tasks, especially when dealing with sparse labeled data.This paper presents a method for improving text classification by combining Expectation-Maximization (EM) and active learning, using a large pool of unlabeled documents. The goal is to reduce the need for labeled training examples while maintaining high classification accuracy. The paper introduces a pool-based active learning approach that selects the most informative examples from the entire pool of unlabeled documents, rather than generating them synthetically or selecting from a stream. This approach is combined with EM to estimate the class labels of unlabeled documents, thereby improving the classifier's performance. The paper first describes a probabilistic framework for text classification based on naive Bayes, which assumes that words in a document are independent given the class. It then explains how EM can be used to improve the classifier by estimating the class labels of unlabeled documents. Active learning is then introduced as a method to select the most informative examples for labeling, with the Query-by-Committee (QBC) algorithm being used to measure classification variance indirectly by examining the disagreement among multiple classifier variants. The paper shows that combining EM with active learning significantly reduces the number of labeled examples needed to achieve high accuracy. Specifically, the combination of QBC and EM requires only 58% as many labeled examples as EM alone and 26% as many as QBC alone. The paper also introduces a new approach called pool-leveraged sampling, which interleaves EM and active learning to further improve performance. The paper presents experimental results on real-world text datasets, such as the Newsgroups and Reuters datasets, showing that the combination of EM and active learning outperforms both methods individually. The results indicate that using a pool of unlabeled documents can significantly improve active learning by reducing the need for labeled examples and increasing classification accuracy. The paper concludes that combining EM and active learning provides a substantial benefit in text classification tasks, especially when dealing with sparse labeled data.