Modeling semantic and orthographic similarity effects on memory for individual words

Part III:

Several different explanations for the word-frequency effect have been proposed; probably because word frequency is correlated with many variables. The advantage for low frequency words has been attributed to elevated attention (Brown, 1976; Glanzer & Adams, 1990; Lockhart, Craik, & Jacoby, 1976; Maddox & Estes, 1997; Shepard, 1967), extra rehearsal time (Mandler, 1980), differences in pre-experimental recency (Scarborough, Cortese, & Scarborough, 1977; Underwood & Schultz, 1960), noise from extra-list memory (Estes, 1994; Maddox & Estes, 1997; Shiffrin & Steyvers, 1997), number of different contexts (Dennis & Humphreys, in review) and differences in the variability with which words are encoded (McClelland & Chappell, 1998). The Retrieving Effectively from Memory theory (REM, Shiffrin & Steyvers, 1997, 1998) accounts for the word-frequency effect on the assumption that the memory representations of low-frequency words tend to be made up of less common features than the memory representations of high-frequency words. It is of course possible that several or all of the mechanisms proposed are operating simultaneously. It should be pointed out that while Shiffrin and Steyvers (1997) employed the feature frequency assumption as the sole mechanism to predict word frequency effects, they were careful to point out that many other plausible factors could also contribute to word frequency effects. In this paper, however, we empirically test the feature-frequency assumption.

Landauer and Streeter (1973) pointed out that the frequency distributions of orthographic and phonetic features are dependent on normative word-frequency. For example, the letter “X” is twice as likely to occur in rare words than in common words. Almost all implementations of the REM model assume that features vary in their environmental frequency, or ‘base rate’. This feature frequency assumption can be used to explain word frequency effects: because high frequency words are encountered more often, the features that make up high frequency words are also encountered more often. This means that feature frequency is correlated with normative word frequency. In REM (Shiffrin & Steyvers, 1997), high frequency words were represented with vectors having more common feature values and low frequency words were represented with vectors having more rare feature values. Because the REM model is sensitive to the diagnosticity of the features that make up words (memory traces with rare features that match the test features provide better evidence), it predicted an advantage for low frequency words over high frequency words as well as mirror effects for hit and false alarm rates.

Convergent evidence for the feature-frequency assumption comes from a set of experiments by Zechmeister (1969, 1972) that showed that words that were rated as orthographically distinct (e.g. sylph) were better recognized than words rated less orthographically distinct (e.g. parse). He also showed that the distinctiveness ratings were related to both the frequency of letter combinations and orthographic distinctiveness.

In this study, instead of using ratings, we assess feature frequency by measures that are directly based on the frequencies of the individual letters that make up words. The results of this study will be modeled by two versions of the REM model. The first model is based on the REM model as described by Shiffrin and Steyvers (1997) in which words are represented by arbitrary feature values. In the second model, the representation of the words is directly based on the orthography of the words used in the experiment and on the environmental base rates of letters occurring in words.