basilio.calderone
Phonotactics refers to the sequential organization
of phonological units that are legal in a language
(Crystal 1992). However, legal sound sequences
do not all occur with the same probability in a
language. Phonotactic probability is most often
measured in terms of transitional probabilities
(TPs) of biphones and has been shown to
influence a large range of processes, including
infants’ discrimination of native language sounds,
adults’ ratings of the wordlikeness of nonwords
In this paper we focus on the role of phonotactic probabilities in priming morphologically simplex and complex words in Italian. We investigate whether biphone TPs affect the recognition of word targets after exposure to fragment primes differing in the probability with which the fragment-final consonant predicts the consecutive segment in the target.
We opted for a non-factorial, regression
design including lexical and sub-lexical frequency
and distributional variables as predictors
Experiment
Forty-two native Italian speakers participated in a speeded lexical decision task in a fragment priming paradigm. Thirty bi- or tri-syllabic Italian nouns containing a biphonemic consonant cluster in internal position (e.g. borsa, ‘bag’) served as targets. Each target was primed by a sequence corresponding to an initial fragment of the target (e.g. bor-borsa). The fragment prime could consist of 3 o 4 phonemes and always ended with the first consonant of the cluster. The average length ratio between prime and target was 0.49. The clusters were different across words and each cluster could occur in only one target (although more than one fragment could end in a given consonant). 12 were heterosyllabic (e.g. bor-sa ‘bag’), 12 tautosyllabic (e.g. degrado ‘decay’) and 6 ambisyllabic clusters (e.g. dis-tanza ‘distance’).
Another set of 30 Italian nouns matching for average length, frequency and prime/target length ratio, in which the fragment prime ended in a syllable onset consonant followed by a vowel (e.g. tuc-tucano ‘toucan’). The same proportion of fragment-final consonants was maintained in the two sets of words.
Sixty pseudowords matching for average length and properties of the fragment were added. Pseudowords were obtained by changing one letter of existing words (belonging to the same frequency range of the experimental words), for 1/3 in their initial part, 1/3 in their central part and 1/3 in their final part. The 30 clusters used for pseudowords did not appear in the words’ list.
In the lexical decision task, participants were asked to press a button corresponding to their dominant hand as soon as the orthographically presented target was judged as a word, and a different button for targets judged as nonwords. All the stimuli appeared in Courier New font, 18 point size in the center of the computer screen. In order to avoid allographic effects, primes were displayed in uppercase and targets in lowercase. The fixation was 200 ms, followed by a 50 ms pause. Primes appeared for 150 ms, followed by a 50 ms pause. The targets remained on the computer screen for a maximum of 1 sec. If the participants did not produce any answer within that time, the feedback Fuori tempo (‘Out of time’) appeared on the screen. Reaction times (RTs) and the number of errors (Nerr) constituted the dependent variables. The reaction times were measured from target onset to subject’s response, and responses given after the deadline were scored as errors.
The Experiment was preceded by a practice session. When the participants reached the 70 % of valid responses the experiment started. 2.2
Several statistical and distributional properties of
word primes, targets and clusters were derived
from the CoLFIS corpus
For each prime-target pair, we calculated (i) the token frequency of the target (‘TargetFreq’), (ii) the N of words beginning with the prime fragment (‘PrimeTypeFreq’), (iii) the cumulated frequency of the words in (ii) (‘PrimeTokenFreq’), (iii) the length of the target (in N graphemes), (iv) the length of the prime (in N graphemes), (v) the prime/target length ratio.
For each cluster, we calculated (vi) the TP value, i.e. the probability with which the first consonant of the cluster predicts the occurrence of the following consonant, calculated over the corpus word tokens (‘BigramTP’), (vii) the N of words containing the cluster (‘BigramTypeFreq’), (viii) the cumulated frequency of the words in (vii) (‘BigramTokenFreq’), (ix) the TP between the fragment prime and the second consonant of the cluster, e.g. P(s|bor) in borsa ‘bag’ (‘SequenceTP’), (x) the N of words containing the sequence of the prime followed by the second
C of the cluster (‘SequenceTypeFreq’), (xi) the cumulated frequency of the words in (x) (‘SequenceTokenFreq’). 2.3
Fixed and mixed models with subject and prime as random variables were used.
For the purposes of the present study, we tested two different models, both including frequency variables and phonotactic probability variables; they are shown in Table 1. The two models differed for the presence, in model II, of a measure of prime frequency, which was not included in model I, and for being focused either on sequence and bigram token frequencies (model I), or on sequence and bigram type frequencies. Both models were tested for CC items (e.g. borsa, ‘bag’) and CV items (e.g. tuc-ano ‘ toucan’) separately.
Model II TargetFreq PrimeTokenFreq LengthRatio SequenceTypeFreq BigramTypeFreq SequenceTP BigramTP Subject Fragment prime Fixed effects
Model I TargetFreq LenghRatio SequenceTokenFreq BigramTokenFreq SequenceTP
BigramTP Random effects
Subject
Fragment prime
The results of the fixed effects analyses for the relevant models are summarized in Table 2 (dependent variable: RTs) and Table 3 (dependent variable: Nerr).
According to model I, with RTs as the dependent variable, the sequence’s TP (i.e., the TP between the fragment prime and the second consonant of cluster) turned out to be the most significant predictor, even outranking the contribution of frequency values (for the target, the sequence and the bigram), which all concurred to the intercept. A different picture emerged however for the CV items, for which no probability variables turned out to significantly predict the subjects’ response times; on the contrary, the target frequency, with the secondary contribution of the frequency of the cluster, appeared to play a role for this subset of items.
According to model II, for CC items the role of the target frequency turned out to be very important, and the only additional effect was generated by the sequence’s TP. Thus the two models were similar in emphasizing the role of the probability with which a given C follows the prime sequence. As for CV items, model II returned a picture very similar to the one that emerged in model I, with target frequency and bigram type frequency as the only significant predictors.
When subject and prime were included as random factors, the pairwise comparison in the likelihood ratio test confirmed that the contribution of the sequence’s TP increased significantly the predictability of the RTs patterns: χ2(1)= 11.184, p= 0.0008 in model I, χ2 (1)= 5.4403, p= 0.019 in model II.
The average reaction times and the number of errors were positively and significantly correlated, though with an intermediate correlation coefficient (r = .648, p < .01). We thus tested the two models with Nerr as the dependent variable, in order to determine if the error rate was influenced by frequencies and probabilities to a different extent than response latencies.
With Nerr as the dependent variable, R2 values were consistently lower than in the RTs simulations (Table 3), thus indicating that the error patterns were accounted for by our frequency and probability variables to a more limited extent. In particular, both model I and model II emphasized for the CC items the role of target frequency as the only significant predictor of errors, while for CV items an additional role of bigram frequencies (by token and by type, respectively) was found. Thus for the CV items, RTs and error rate produced consistent results. This work aimed to shed light on the role of TPs in a so far unstudied experimental environment, i.e., a lexical decision task with fragment priming. As the large part of studies on phonotactic probabilities focused on English, this work also added to the field with evidence from a poorly investigated language, Italian.
Fragment priming is known to be modulated
not only by word frequency and the frequencies
of words matching the fragment but also by
topdown information conveyed by the prime: a
fragment prime matching a unique
morpholexical family is as effective as a stem prime,
thus showing that priming acts as a cue for the
properties displayed in the target
This study has shown that the priming effect when an initial fragment is available is influenced also by bottom-up variables; in particular, it depends on the probability with which the segments composing the fragment or the fragment-final consonant predict the occurrence of the consecutive consonant. Although to a lesser extent, the frequency with which bigrams and sequences occur (as types or tokens) in the lexicon also predict the subjects’ behavior. Phonotactic probabilities thus turned out to predict the subjects’ response to a large degree for many of the phonological environments tested in the current experiment, sometimes outperforming target frequencies, and consistently overtaking the contribution of the prime/target length ratio and of the prime frequency.
The results however suggested that the
phonotactic probabilities in the case of consonant
clusters were overall more important than in the case
of consonant-vowel sequences; thus it must be
concluded that the contribution of TPs in lexical
recognition is not the same across phonological
environments. Consonant clusters might play a
particularly relevant role in lexical access,
compared to CV sequences, as contemporary theories
based on the principles of phonological and
morphological naturalness also seems to predict
Additionally, for CC sequence the token frequencies (of the bigram and of the prime + C sequence) turned out to be relatively more important than the corresponding type frequencies, thus suggesting that the exposure to the number of occurrence of a cluster or of a segment sequence may be more important in lexical access than the exposure to the individual items containing them.
An additional issue concerns the role of TPs in
morphologically complex words. According to
some models, morphological parsing is necessary
for lexical access and the prefix (in the case of
prefixed words) has to be stripped away in order
for the word to be recognized
A follow-up experiment will therefore test the
contribution of phonotactic statistical knowledge
in native speakers’ access to complex word
forms (specifically, prefixed nouns). Prefixed
and pseudo-prefixed words will be used for that
purpose. In particular, fragment primes will be
selected according to two different conditions: in
condition a) the targets are prefixed words and
the fragment prime coincides with the prefix
(e.g. bis-bisnonna ‘grandmother’); in condition
b) the targets are pseudo-prefixed words and no
morphological boundary occurs between the
initial fragment and the second part of the word
(e.g. per-perdente ‘loser’). Together with the
current experiment, the experiment on prefixed
and pseudo-prefixed words will determine
whether or not the role of TPs is different when
the target is a simplex word compared to when it
is a prefixed word, and to when it is a
pseudoprefixed word. Different hypotheses may be put
forward here, according to whether or not
morphological boundaries affect the processing of
consonant clusters
By describing phonotactic probability and
frequency effects during word recognition, this
study offers arguments to models of lexical
access based on bottom-up processes such as
cohort models for orthographic stimuli (see e.g.
Johso