A domestic robot must recognize its current place accurately and interact with human beings effectively, thus we desire efficient and semantically meaningful scene representation. In this article, we introduce weighted component pooling to analyze indoor scenes, and probabilistic semantic mapping to represent them based on interactive robot learning. We test this algorithm with 10 scene types from an indoor scene recognition image set and 5 scene types with a humanoid robot in domestic settings. Our result shows that the robot can learn and find desired place according to our verbal commands accurately.