Wafer map defect patterns provide valuable information for root cause analysis and yield learning. Previous studies show that supervised machine learning (ML) methods can achieve good defect pattern recognition rate. However, the effectiveness of these methods rely on accurately labeled samples, which leads to two problems. First, oftentimes there are mislabeled data, which may prevent an accurate model from being established. Secondly, defect patterns that are not defined a priori will not be recognized. In this paper, we propose a semi-supervised learning method to deal with these problems. Labeled wafer maps are first used to train a prediction model, and questionable samples are excluded in the process. The prediction model is then used to classify unlabeled data. The remaining data that cannot be properly classified are then sent to an unsupervised learning algorithm to extract more defect patterns for enhanced labeling. The proposed approach is validated using the WM-811K database. With the proposed method, we are able to define five new defect pattern types, and the 14 defect types can be recognized with high accuracy.