Praising behavior is considered an important method of communication. It is considered effective in building good relationships with others and bringing out the best in them. However, there are quite a few people who have difficulty in praising successfully.These people have difficulty judging and improving their praising skills. The reason is that they have not clarified which behaviors are important for successfully praising. In this paper, we analyze praising behavior by focusing on the content of utterances in Japanese. We construct machine learning models for estimating praising skills to analyze behaviors for successfully praising. Then, we extract features from the utterances of persons who give praise (praiser) and those who receive praise (receiver). The results showed that the best performance of our models was F1 = 0.632. We analyzed praising skills from the results of that model. An utterance of praising scene, one receiver utterance immediately preceding the utterance of praising scene, and four receiver utterances immediately following it are crucial for estimating praising skills. The cosine similarity between an utterance by a praiser and one by a receiver is also important as a feature. Thus, we confirmed that the utterances of receiver are similar to the utterances of praiser.