Fig. 1. An abstract of the co-training algorithm.

Fig. 2. An example news article from Reuters-21578 corpus.

Fig. 3. The idea of support vector machines.

Fig. 4. The performance improvement by using both lexical and syntactic information. Two different measures are used: LF1 and LF2.

Fig. 5. The analysis of the effects of unlabeled documents for the topics with a great number of relevant documents in TREC-7. The performance measure is LF1.

Fig. 6. The analysis of the effects of unlabeled documents for the topics with a small number of relevant documents in TREC-7. The performance measure is LF1.

Fig. 7. The improvement in accuracy by using additional unlabeled documents.

Fig. 8. A perceptron designed to see the effect of the first assumption in the co-training algorithm.

Table 1. Syntactic features for document classification

Table 2. The accuracy improvement by using syntactic information on Reuters-21578 corpus

Lexical implies when we use only tf·idf, Syntactic when we use only text chunks, and Both when we use both of them. The performance measure is LF1.

Table 3. The result on TREC-7 dataset

The performance measure is the averaged scaled LF1 and LF2 when U(s) is set to 0.

Table 4. The effect of the first assumption in the co-training algorithm

The experiments are performed on TREC-7 dataset.

Table 5. The conditional mutual information and average LF1 for various feature splits

The experiments are performed on Reuters-21578 dataset.

Table 6. The effect of the second assumption in the co-training algorithm

The experiments are performed on Reuters-21578 dataset.