Alternatively, it could indicate quality problems in the training data. I wonder how often the extracted data was repeated in the training dataset?