Toddlers can learn about the meanings of individual words from the structure and semantics of the sentences in which they are embedded. However, it remains unknown whether toddlers encode similarities among novel words based on their positions within sentences. In three experiments, two-year-olds listened to novel words embedded in familiar sentence frames. Some novel words consistently occurred in the subject position across sentences, and others in the object position across sentences. An auditory semantic task was used to test whether toddlers encoded similarities based on sentential position, for (a) pairs of novel words that occurred within the same sentence, and (b) pairs of novel words that occurred in the same position across sentences. The results suggest that while toddlers readily encoded similarity based on within-sentence occurrences, only toddlers with more advanced grammatical knowledge encoded the positional similarities of novel words across sentences. Moreover, the encoding of these cross-sentential relationships only occurred if the exposure sentences included a familiar verb. These studies suggest that the types of lexical relationships that toddlers learn depend on the child’s current level of language development, as well as the structure and meaning of the sentences surrounding the novel words.