A method for training a machine learning model for open domain question answering includes receiving trained classifiers for question answering. The received trained classifiers are used to generate a set of candidate answers to a question. Second trained classifiers are used for scoring each of the candidate answers. The scoring indicates a measure of how well each candidate answer answers the question. Using the second trained classifiers for scoring each of the candidate answers includes comparing each candidate answer to a first ground truth corresponding to the question. A set of top-scoring candidate answers is presented to a human operator who marks each as correct or incorrect. The correct candidate answers are treated as additional ground truths for further training the first trained classifiers.