It has been hypothesized that known words in the lexicon strengthen newly formed representations of novel words, resulting in words with dense neighborhoods being learned more quickly than words with sparse neighborhoods. Tests of this hypothesis in a connectionist network showed that words with dense neighborhoods were learned better than words with sparse neighborhoods when the network was exposed to the words all at once (Experiment 1), or gradually over time, like human word-learners (Experiment 2). This pattern was also observed despite variation in the availability of processing resources in the networks (Experiment 3). A learning advantage for words with sparse neighborhoods was observed only when the network was initially exposed to words with sparse neighborhoods and exposed to dense neighborhoods later in training (Experiment 4). The benefits of computational experiments for increasing our understanding of language processes and for the treatment of language processing disorders are discussed.