Abstract
The word frequency effect refers to the observation that high-frequency words are processed more efficiently than low-frequency words. Although the effect was first described over 80 years ago, in recent years it has been investigated in more detail. It has become clear that considerable quality differences exist between frequency estimates and that we need a new standardized frequency measure that does not mislead users. Research also points to consistent individual differences in the word frequency effect, meaning that the effect will be present at different word frequency ranges for people with different degrees of language exposure. Finally, a few ongoing developments point to the importance of semantic diversity rather than mere differences in the number of times words have been encountered and to the importance of taking into account word prevalence in addition to word frequency.
When word recognition is analyzed, frequency of occurrence is one of the strongest predictors of processing efficiency. High-frequency words are known to more people and are processed faster than low-frequency words (the word frequency effect; Monsell, Doyle, & Haggard, 1989). This is true for tasks such as word naming, lexical decision (does the letter string refer to an existing word or not?), and semantic decision (e.g., does the word refer to an animal?). Word frequency is also of importance for memory performance. In memory research, participants first study a list of words and are later required to recall the stimuli or to discriminate them from lures (new items). Interestingly, the pattern of results depends on the task: Low-frequency words in general are more difficult to recall but lead to better performance in a recognition task (Yonelinas, 2002).
Word frequency typically explains some 30% to 40% of the variance in word recognition tasks (Brysbaert, Stevens, Mandera, & Keuleers, 2016). This effect was first reported by Preston (1935) and has received renewed attention in recent years because new, improved word frequency norms and word processing data for thousands of words have been collected. The present article is an update of a review written a few years ago (Brysbaert et al., 2011).
Not All Word Frequency Measures Are Equal
For a long time, researchers did not have much choice about which word frequency measure to use. Because counting words in printed books and newspapers was time intensive, only one or two lists (if any) existed per language. The situation changed dramatically when texts became available digitally. Then it became much easier to gather a sample of texts (called a corpus) and count the words in them.
Surprisingly, psychologists were not eager to turn to the new word frequency lists. They preferred to continue working with the established and familiar lists (such as Kucera & Francis, 1967, for English), arguably because they did not trust the validity of the new word counts. 1 This situation did not change until frequency lists could be validated against word processing times for thousands of words (collected in megastudies). The validation studies showed that the best word frequency norms are based on language the participants are likely to have been exposed to. This may sound like a truism, but before the validation studies, researchers typically used word frequencies based mainly on nonfiction texts, such as newspapers, magazines, and scientific books. When fiction materials were included, they consisted of a limited number of novels and stories. The new frequency measures explain more than 10% extra variance in word recognition performance than the Kucera and Francis (1967) measure.
For undergraduate students (the most commonly used participants in psychology studies), the best word frequency measures turned out to be based on corpora of television subtitles (Brysbaert & New, 2009), social media (Gimenes & New, 2016; Herdağdelen & Marelli, 2017), and blogs (Gimenes & New, 2016). In general, a combination of these sources gives better results than each source alone. For older participants, traditional word frequency measures based on books are sometimes better (Brysbaert & Ellis, 2016). Johns, Jones, and Mewhort (2016) suggest that further gains may be possible by compiling frequency lists tailored to the participants of a study, depending on their learning histories (e.g., how much television they watched, which book authors they read, how active they have been on social media, which schoolbooks they used). Good frequency lists are based on a large corpus (not smaller than 20 million words). Such lists also include information about the syntactic roles played by the words (e.g., nouns, verbs) so that this information can be used in research as well. In addition, good frequency lists should have proven their mettle in validation tests based on megastudy data.
A Good Standardized Measure of Word Frequency
Because frequency counts depend on the size of the corpus, researchers typically work with a standardized measure so that the various counts can be compared. Thus far, the main standardized measure has been frequency per million words (fpm). Low-frequency words are typically defined as having less than 5 fpm (e.g., “gloom,” “frenzy,” “objection”); high-frequency words have more than 100 fpm (e.g., “energy,” “market,” “area”).
There are two problems with the fpm measure. First, in corpora with tens of millions of words, most words have a frequency lower than 1 fpm. For instance, in the SUBTLEX-US corpus (Brysbaert & New, 2009), which contains 50 million words, three quarters of the words (i.e., 56,000 of the 74,000 word types) occur with a frequency below 1 fpm. The percentage becomes even higher for word frequency lists based on larger corpora. Since many of these words are well-known, more than half of the word frequency effect is situated below the intuitive start of the scale (1 fpm).
A second problem with the fpm measure is that the frequency effect is compressed, typically represented by a logarithmic curve. That is, the difference between 1 fpm and 2 fpm has more or less the same effect on processing times as the difference between 10 fpm and 20 fpm, between 100 fpm and 200 fpm, and between 1,000 fpm and 2,000 fpm. This compressed nature is even more of a problem for words below 1 fpm, because it means that the difference between 1 fpm and 2 fpm has the same effect as the difference between .1 fpm (1 per 10 million words) and .2 fpm (2 per 10 million words) and even between .01 fpm (1 per 100 million words) and .02 fpm (2 per 100 million words) if these words are known.
Because the fpm scale provides users with the wrong intuitions (1 fpm is the start of the scale, differences lower than 5 fpm are negligible), van Heuven, Mandera, Keuleers, and Brysbaert (2014) proposed the Zipf scale as an alternative. This scale is logarithmic (like the decibel scale for loudness or the Richter scale for earthquakes) and is calculated as follows 2 : Zipf = log10 (frequency per billion words). In practice, the scale runs from 1 (1 per 100 million words) to 6 (1,000 per million words). The lower half of the scale (1–3) represents the low-frequency words, the upper half (4–6) the high-frequency words. There are few words with frequencies higher than 6 Zipf, and they are nearly all function words (“the,” “you,” “but,” “with,” etc.). Similarly, in a corpus larger than 100 million words, there are words with frequencies below 1 Zipf, but few of these are known.
Figure 1 shows the distribution of Zipf values for the 29,902 words known to more than 75% of the participants in the lexical decision task of the English Lexicon Project, a megastudy with naming and lexical decision times for 40,000 English words (Balota et al., 2007).
Distribution of English words on the Zipf scale of word frequency. Words were taken from the English Lexicon Project (ELP; Balota et al., 2007). Data are limited to the words that more than 75% of the participants who took part in a lexical decision task identified as being real. Frequencies are based on the average of the measures collected by Brysbaert and New (2009) and Gimenes and New (2016), which accounted for more variance in the lexical decision times than the individual measures.
Individual Differences in the Word Frequency Effect
Preston (1935) already reported that the frequency effect is larger for university students with a small vocabulary than for university students with a large vocabulary. This observation has largely been lost since that time but has regained momentum in recent years (Davies, Arnell, Birchenough, Grimmond, & Houlson, 2017; Kuperman & Van Dyke, 2013; Mandera, 2016), partly after it was discovered that the larger frequency effect is also observed in second-language speakers (Cop, Keuleers, Drieghe, & Duyck, 2015). This suggests that the difference in frequency effect is not related to differences in intelligence (second-language speakers, on average, are not less intelligent than native speakers) but to differences in language exposure, which can be measured with a vocabulary test (Brysbaert, Lagrou, & Stevens, 2017).
Monaghan, Chang, Welbourne, and Brysbaert (2017) reported that the decrease of the word frequency effect as a function of word exposure can be simulated with a connectionist network. As the network gets more practice with input words, the word frequency effect diminishes. At the same time, simulations showed that a network needs some exposure to words before a word frequency effect emerges (which is understandable, as words need to be encountered a few times before they can show an effect). The net result is that a learning network (and person) at first will show a small frequency effect, which initially grows and then again decreases.
The findings of Monaghan et al. (2017) help us to more deeply understand the individual differences observed in the word frequency effect. At each point in time, there are a range of word frequencies to which individuals show a strong frequency effect, lower-frequency words for which they show a small frequency effect (because they hardly know these words), and high-frequency words for which they show a smaller frequency effect as well (because these words are overlearned). Figure 2 presents the word frequency effects that can be observed for people with different word exposure levels (estimated via vocabulary size), both for accuracy and response times.
Data patterns that can be expected from models investigating the word frequency effect: mean probability of words being known (left) and mean response time in a lexical decision task (right) as a function of word frequency and vocabulary size. For vocabulary size, low was defined as knowing a few thousand words from the English Lexicon Project (ELP; Balota et al., 2007), medium as knowing slightly more than half of the words from the ELP, and high as knowing most of the words from the ELP. According to the models, individuals with a low vocabulary size would show a word frequency effect only for high-frequency words (Zipf > 4); the other words would not be known to them. People with a medium vocabulary size would show the largest frequency effect (between 3 and 5 on the Zipf scale). Finally, persons with a high vocabulary size would show the clearest frequency effect for low-frequency words (Zipf = 2–4). All participants respond equally quickly to very-high-frequency words. Participants with less exposure respond more slowly (and less accurately) to low-frequency words. For the empirical data on which the curves are based, see Keuleers, Diependaele, and Brysbaert (2010) and Brysbaert, Lagrou, and Stevens (2017).
Interpretations of the Word Frequency Effect
The standard interpretation of the word frequency effect is that it is a learning effect. Indeed, the compressed frequency effect has much overlap with the decelerating learning curve observed in repeated tasks. 3 In computational models of word processing, learning is captured by adapting the activation levels of word representations as a function of their frequency (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001) or by having stronger connections between frequently coactivated representations. The latter follows naturally from the way in which connectionist models learn associations between input patterns and output patterns (Harm & Seidenberg, 2004).
Not everyone agrees that the word frequency effect is a simple learning effect, however. For a start, word frequency is highly correlated with a number of other word features: word length, age at which the word was acquired, and similarity to other words. So, in principle, the word frequency effect could be confounded with any of these variables (or alternatively, the effect of any of these variables could be a word frequency effect in disguise). Analyses of the reaction times obtained in megastudies suggest that all of these potential confounds have an independent effect on word processing (e.g., Brysbaert et al., 2016). For instance, even when the effects of all other variables are partialed out, there is still a robust word frequency effect (although its impact is diminished to some 5–10% of the variance explained). In addition, word frequency interacts with these variables. Typically, the effect of a variable is stronger for low-frequency words than for high-frequency words.
Another possibility is that it is not so much the number of encounters that matters but the diversity of situations in which the words are encountered. According to this view, words found in many different settings are responded to more efficiently than words present in a small range of settings only. This idea was brought to the forefront by Adelman, Brown, and Quesada (2006), who argued that contextual diversity was a better predictor of word processing efficiency than word frequency. Contextual diversity is defined as the number of texts in which a word appears (rather than the total number of times the word is encountered). A similar idea was defended by Jones, Johns, and Recchia (2012) and Johns, Dye, and Jones (2016). They showed that when novel words are encountered across distinct discourse contexts, people are both faster and more accurate at recognizing them than when the words are seen in redundant contexts. It will be interesting to find out to what extent these alternative measures provide a better account of the frequency effect outlined in Figure 2 (see Joseph & Nation, 2018, for counterevidence).
Finally, a challenge for the word frequency effect is that there are many words in the low-frequency range that are responded to rapidly and accurately. The curves of Figure 2 show the main effects but not the scatter observed at the low-frequency end. Indeed, some low-frequency words are recognized as quickly and accurately as high-frequency words are. As a result, word frequency does not explain all of the systematic variance in megastudies but only some 30% to 40% (Brysbaert et al., 2016). Several factors are responsible for the finding that low-frequency words are not all equally difficult.
First, many low-frequency words are related to high-frequency words through inflection, derivation, and compounding (e.g., “distinctively,” “microbiologist,” “reusable,” “unsweetened,” “screenshot”). Such words can be recognized by decomposing them into their components. Second, some words are rarely spoken, even though people are familiar with them (e.g., “ladle,” “hinge,” “sanitizer”). The frequency of other words may also be misjudged because of the language register tapped into by the corpus (e.g., subtitles, texts, social media). Finally, the word frequency effect is built on the idea that each encounter with a word has the same weight. This need not be the case. Some words are much easier to learn than others. Indeed, some words seem to be remembered for the rest of one’s life after their first encounter (e.g., a film or a book about a unicorn or a gnome), whereas other words tend to be forgotten easily (e.g., “kestrel,” “hangar,” “cinch”). So the number of encounters itself may not be the best measure of word knowledge.
To counter the shortcomings of word frequency norms, Brysbaert et al. (2016) introduced the variable of word prevalence, defined as the percentage of people who know the word. This variable explains some extra 7% of response times in lexical decision megastudies in addition to all the known variables, particularly at the low end of the word frequency range. Thus far, word prevalence has been collected only for the Dutch language. The English language will follow soon.
