*Introduction:* In a lexical decision “megastudy”, researchers collect
responses to a wide range of words and non-words (e.g. differing in
morphological complexity) in order to produce a massive database via which
others can subsequently test novel hypotheses by analyzing a subset of the
total dataset. Megastudies circumvent many of the shortcomings of
traditional experiments and have grown in popularity in recent years
(Keuleers and Balota 2015): visual lexical decision megastudies have been
conducted on English (Balota et al. 2007), French (Ferrand et al. 2010),
Malay (Yap et al. 2010), Dutch (Brysbaert et al. 2016), and Cantonese (Tse
et al. 2017). To date, no megastudy has focused on a language which
productively uses nonconcatenative morphology, which poses novel challenges
for theories of word recognition (e.g. Frost et al. 1997). A Semitic
language, Maltese not only uses nonconcatenative morphology, but its
speakers have borrowed extensively from Indo-European languages such that
roughly half the lexicon comprises Sicilian, Italian, and English loanwords
which largely use concatenative morphology (Bovingdon and Dalli 2006),
creating further challenges for theories of lexical processing. We report
the creation of a database of Maltese visual lexical decision responses and
demonstrate its use in replicating an analysis of the effect of etymology
(Semitic vs. non-Semitic borrowings) on lexical decision speed.
*Methods:* In total, we have collected approximately 237,000 lexical
decision responses from 104 native or near-native speakers of Maltese (*M*
Age = 24.0 years, range = 18−77 years; 53 participants identified as
female, 51 as male; 87 identified as right-handed, 17 as left-handed) to
11,000 real Maltese words and 11,000 non-words. The real-word targets were
randomly selected from Korpus Malti v3.0 (Gatt and Čéplö 2013), a
250-million-token corpus of written Maltese that we trimmed to remove
non-Maltese texts and non-words (e.g. URLs), and supplemented by other
written sources. The selected words were also checked against Ġabra, a
Maltese lexical database containing 16,593 lemma-based entries (Camilleri
2013), and vetted by a native speaker. The final set of real-word targets
consisted of 6,451 Semitic Maltese words, 4,439 non-Semitic words, and 110
words of uncertain etymology (Aquilina 1987-1990), and included both
uninflected and inflected forms. Real-word targets ranged in frequency from
0−20,385.4 occurrences per million words in Korpus Malti (*M* = 36.1
occurrences per million words) and in length from 2−21 letters (*M* = 7.1
letters). For each real-word target, we constructed a non-word matched in
length and in frequency-weighted neighborhood density (*M*Real = 92.9, *M*
Nonce = 88.5 occurrences per million; Welch’s *t*-test: *t*(21,998) =
−0.47, *n.s.*). A native speaker vetted all potential non-word targets.
Participants completed 1−35 total sessions (*M* = 5.8 sessions), up to
three sessions per day, during each of which they judged the lexicality of
200 visually-presented real words and 200 non-words to produce a total of
9−13 lexical decisions per item (*M* = 10.7 decisions). We excluded data
from participants whose average RT exceeded 1,500 ms or accuracy rate fell
below 80.
*Analysis of lexical stratum:* In a visual masked priming lexical decision
study, Geary and Ussishkin (2018) found that Maltese readers responded
faster to Semitic-origin words than to non-Semitic borrowings, independent
of frequency, neighborhood density, and word length. We replicate this with
a larger dataset (10,890 versus 96 different words) by analyzing log RTs to
real-word targets on trials where participants responded correctly, using
the lme4 package (Bates et al. 2015) in R (R Core Team 2019) to fit an LMER
model and assessing significance using the lmerTest package (Kuznetsova et
al. 2016) to simulate Satterthwaite approximations for degrees of freedom.
The model included lexical stratum (Semitic vs. Non-Semitic; reference:
Non-Semitic), log frequency, log frequency-weighted neighborhood density,
age, trial number, session number, and same-day session number as fixed
effects; subjects and targets as random effects; and by-subjects random
slopes for lexical stratum. The effect of lexical stratum was significant (
*t*(191.2) = −7.13, *p* < 0.001), with participants reliably faster to
judge Semitic words (*M* = 847 ms) than non-Semitic words (*M* = 852 ms).
While the effect size is considerably smaller (5 versus 30 ms), the
replication of this effect motivates further analyses which will assess
whether, for instance, the “lexical stratum” effect may in fact reflect
targets’ overall morphological complexity.
