Below I have copy and pasted the dissertation-y language that describes the stimuli we constructed for Jacobs, Dell, Benjamin & Bannard (2016) and Jacobs, Dell, & Bannard (2017). If you use any of these sets of materials, please cite the appropriate paper of the below:
-
Jacobs, C. L., Dell, G. S., & Bannard, C. (2017). Phrase frequency effects in free recall: Evidence for redintegration. Journal of Memory and Language, 97C, 1-16.
-
Jacobs, C. L., Dell, G. S., Benjamin, A. S., & Bannard, C. (2016). Part and whole linguistic experience affect recognition memory for multiword sequences. Journal of Memory and Language, 87, 38-58. doi: 10.1016/j.jml.2015.11.001
These were extracted from the Google 1T n-gram corpus (Brants & Franz, 2006) using word lists for each category extracted from the part-of-speech-tagged British National Corpus (BNC). In these phrases, both the nouns and the adjectives had a restricted log frequency range of 19-23.5. The phrases exhibited a relatively broad phrase frequency range (log frequencies from 5.4 to 19.7). The most common phrases (rheumatoid arthritis, alcoholic beverages) were approximately as common as the least common adjective (decadent) and noun (grasslands) in our dataset.
The phrases used in Experiment 1 were taken from the Google n-gram corpus as described. While this tells us that they occurred on the Internet with some frequency, many of the infrequent phrases (e.g. “chrome throttle" or "psychic nephew”) would not be encountered frequently in daily life, and consequently we cannot be sure that they are meaningful to participants. This may put them at an encoding disadvantage, as has been seen in recognition memory for pseudowords (e.g. Diana & Reder, 2006). We therefore tested whether our key effects hold for another set of phrases where even the "low frequency" phrases are likely to be familiar and meaningful to participants.
To that end, we developed an additional stimulus set from the spoken portion of the Corpus of Contemporary American English (COCA; Davies, 2008), which consists primarily of publicly broadcast material from the news, on talk shows, etc. These phrases therefore represent more easily recognizable phrases. We gathered a total set of 112 phrases (56 in a high frequency phrase list and 56 in a low phrase frequency list) meeting several criteria, which we discuss below.
All the phrases we gathered from COCA were compositional (nonidiomatic) adjective-noun phrases varying in their frequency of occurring in the subset of the database containing spoken English. We calculated the spoken frequencies of these phrases from the years 2009 to 2012, which represents more a recent and ecologically valid sample of the language the typical freshman undergraduate would experience while watching the news from the beginning of middle school through the most recent collection of data in COCA. Noun and adjective length did not significantly correlate with phrase frequency (Pearson's r = -.11, p = .28 and r = -0.14, p = .16).
Nouns and adjectives were deliberately selected to be higher in frequency than in Experiment 1 in order to increase the chances that the participants actually knew all of the words within the phrase, with the least common adjective and noun occurring 200 times more often than the least common phrase. Frequencies for the adjectives and nouns were restricted to the same range, from 1031 to 4021 and from 1026 to 4037, respectively out of the entire corpus from 2009 to 2012. As such, all nouns and adjectives were within a single power of 2 in COCA frequency. The lowest frequency phrases were "poor credit", "southern food", "fantastic panel", and "nice hair". The highest frequency phrases were "foreign language", "presidential candidate", "middle class", and "grand jury". Log frequencies of the counts ranged from 2.32 to 9.57.
Eighty-eight nouns from a set of ninety-six nouns used by Balota et al. (2002) served as the stimuli, which had been controlled for concreteness/imageability and word length. The full set of nouns was not used because there are additional constraints based on phrase construction that will be clarified when we introduce Experiment 3b. Words in the dataset spanned a continuous log frequency range of 16.9-28.4 and included, for example, tree, wizard, and anvil. All nouns were concrete with the exception of nation.
Phrases from this experiment were a subset of the 112 phrases used in Experiment 3 of Chapter 1. These phrases were taken from the Corpus of Contemporary American English (COCA; Davies, 2008) and included items such as "critical condition", "horrible mistake", and "impossible dream." To ensure that our assessment of the influence of phrase frequency on recall was not the result of any confounding between frequency and compositionality or concreteness, we conducted a norming study on Qualtrics in which University of Illinois undergraduates rated the items along several dimensions and completed the questionnaire at their own pace.
In this norming study, 30 participants provided responses to a number of questions on a five-point Likert scale from "Strongly Disagree" to "Strongly Agree". First, familiarity with the component words of each phrase and the phrase itself was assessed; participants answered whether they knew the meanings of, for example, the word "impossible", "dream", and the phrase "impossible dream." Then, to rate the imageability of the phrase, participants rated whether they could easily picture what this phrase describes. Finally, as a measure of compositionality, participants rated whether "impossible dream" had the same meaning as a dream that is impossible. Ratings were averaged across all participants and then centered and scaled with respect to all items for inclusion in the analyses. In the final stimulus set, phrases were restricted to just those where the average imageability and compositionality scores fell within a narrow range in order to decorrelate imageability and compositionality from phrase frequency (r = .11, t(70) = 0.89, p = n.s. for imageability; r = -.14, t(70) = -1.19, p = n.s., for compositionality). After these requirements were met, 72 phrases remained.