Computational and animal models suggest that the unlearning or pruning of incorrect meanings matters for word learning. However, it is currently unclear how such pruning occurs during word learning and to what extent it depends on supervised and unsupervised learning. In two experiments (N[subscript 1] = 40; N[subscript 2] = 42), adult participants first completed a pretraining, in which each word was paired with two objects across trials: its target and another object (termed secondary target [T2]). Subsequently, participants learned the correct word-object-mappings in a supervised training paradigm and were then tested on the word meanings. During training, trials were structured such that some T2s never occurred with the targets, while others did, allowing us to disentangle the contributions of supervised and unsupervised pruning accounts. Eye movements were tracked during training and testing to measure the activation strength of alternative meanings. The experiments were identical but differed in how often the word was paired with the T2 during pretraining. We found that while weak incorrect associations were pruned quickly (Experiment 1), stronger ones remained present even after ceiling performance (Experiment 2), suggesting that the extent to which incorrect associations are unlearned depends on the strength of the initial mappings. Additionally, pruning was observed even for T2s that did not co-occur with their corresponding word during training in line with unsupervised pruning. Overall, these findings imply that subtle incorrect associations may remain in the lexicon and contribute to other language processes (e.g., word recognition) even after word learning is completed.