One of the most studied and least well understood phenomena in episodic memory is the word frequency effect (WFE). The WFE is expressed as a mirror pattern where uncommon low frequency words (LF) are better recognized than common high frequency words (HF) by way of a higher HR and lower FAR. One explanation for the HR difference is the early-phase elevated-attention hypothesis which proposes two stages of encoding. In the first, called the early-phase, words are identified based on orthographic and/or phonological characteristics. LF words are composed of atypical features making their identification more difficult than HF words. This relative difficulty during the early-phase results in the LF HR advantage. The first two experiments test the proposal that LF words are better recognized due to their distinct lexical features. The second stage of encoding, called the late phase, consists of controlled processing where the semantic features of the item are paramount. According to the early-phase elevated-attention hypothesis, semantic features of HF and LF words do not differ in diagnosticity and do not contribute to the word frequency effect. We find evidence for this assumption in the final experiment by comparing memory for words and objects. (Contains 4 figures and 3 tables.)