We aim to apply cognitive neuroscience insights to vocabulary learning practice. Towards this end, we review currenteducational methods in relation to important characteristics of the mental lexicon, such as similarity-coding. This showsthat methods relate poorly to the mental lexicon, and that especially contrasting - explicitly distinguishing similarities -receives little attention. To remedy this, we run experiments to put these findings into practice. First, we ask participants tolearn artificial vocabulary using retrieval practice multiple-choice, manipulating the orthographic and semantic similarityof distractors. The prediction is that learning will be harder but more effective depending on similarity and translationdirection. Second, we test whether participants show indications of gradient descent learning when guessing in recallretrieval practice. Thirdly, we use cognitive neuroscience and large scale word learning data to model the mental lexicon.Combined, these studies potentially offer relevant scientific and societal insights, applicable to school settings.