The head-preference effect in compound word processing suggests that readers tend to recognise head constituents more easily than modifier constituents (Gagné et al., 2009; Libben et al., 2003; Marelli et al., 2009). However, it is less clear if this effect generalizes to novel compounds, which we assessed in the present novel word training study. Hundred-sixty-five native speakers of English (78 in Experiment 1 and 87 in Experiment 2) participated in a training session in which they were exposed to the spoken form (Experiment 1) or spoken and written form (Experiment 2) and the meaning of the novel compound constituents. Their learning performance was tested using two learning check tasks (e.g., picture naming and semantic categorization for the constituents). In the post-training session, participants completed a semantic categorization task, including novel compound targets with two trained constituents (trained + trained) and novel compound targets with one trained and one untrained constituent (trained + untrained; untrained + trained). Participants had to decide whether the novel compound was living or non-living. In Experiment 1, where participants were trained on the spoken form of the novel constituents only, the head-preference effect on post-training was absent. In Experiment 2 however, where the training opportunity was increased by exposing participants to the spoken and written form of the novel constituents, a difference between head and modifier responses in post-training was found. In the trained + trained condition, participants responded more slowly to heads (2543 ms) than modifiers (2212 ms), pointing to a larger semantic interference effect for head compared to modifier constituents. In the trained + untrained and untrained + trained conditions, there was no significant RT difference between the conditions where the head was trained and the conditions where the modifier was trained, suggesting that the head-preference effect did not generalise to the untrained context. However, individual differences in learning outcomes revealed that participants who reached better outcomes in the training task found it easier to classify novel compound words including trained constituents in the post-training task, suggesting that better learners were more expert at identifying embedded constituents. The current findings point to differences in the role of head and modifier constituent in the acquisition of compound words.