Four deaf children’s (mean age = 10 years 10 months) semantic representations of particular vocabulary items were explored in this study. It was intended to investigate how the familiarity of a word in the child’s preferred modality (BSL or English) would determine how much could be learned about the corresponding item in the child’s less preferred modality. On the basis of the pre-test, items used were classified as familiar or less familiar to the child. It was hypothesised that a deeper semantic representation of the item would allow for more knowledge to be retained after training. Four tasks were used to determine the scope of initial semantic representations for each of the items. The children were then trained on these items using one of two methods; either using a picture representation of the item together with the item in the less preferred modality (method 1), or the item in both the child’s preferred and less preferred modality (method 2). The different training methods were hypothesised to support learning of the forms of the items in the less preferred modality, as well as understanding in the preferred modality. After training a follow-up assessment of four tasks investigated the extent of how familiarity and training method affected knowledge retained. It was found semantic knowledge in the preferred modality increased after training, particularly in the less familiar items. Learning of the items in the less preferred form was more variable and less certain; effect varied from child to child. The two training methods used also had strengths in different areas; method 1 generally helped to a greater extent. It can be concluded that depth of semantic representation can affect lexical learning, and further research is required for supporting learning of new forms.