Exemplar dynamics: Word frequency, lenition and contrast
Janet B. Pierrehumbert
Northwestern University
Evanston, IL
June 16, 2000
To appear in J. Bybee and P. Hopper (eds.), Frequency effects and the emergence of linguistic
structure. Amsterdam: John Benjamins.
1 Introduction
Over the last decades, a considerable body of evidence has accumulated that speakers have detailed
phonetic knowledge of a type which is not readily modelled using the categories and categorical
rules of phonological theory. One line of evidence is systematic differences between languages in
fine details of pronunciation. For example, it is known that Spanish and English differ systematically
in the exact formant patterns typical of their point vowels (Bradlow 1995). Canadian French differs
from both Canadian English and European French in the distribution of VOT times of voiced and
voiceless stops (Caramazza and Yeni-Komshian, 1974). These are just two of many examples, with
more reviewed in Pierrehumbert (in press) and Pierrehumbert et al. (in press); at this point, it
is not possible to point to a single case in which analogous phonemes in two different languages
display exactly the same phonetic targets and the same pattern of phonetic variation in different
contexts. Exact phonetic targets and patterns of variation must accordingly be learned during the
course of language acquisition. The usage-based framework readily accomodates such findings by
proposing that mental representations of phonological targets and patterns are gradually built up
through experience with speech.
A particularly interesting and challenging result is the discovery that learned phonetic detail
may be associated not just with languages or dialects, but even with specific words in the lexicon of
a given dialect. This observation is made most convincingly in a series of studies by Bybee which
explore the relationship of word frequency to lenition. Bybee (Hooper 1976) explored the process
of schwa reduction and desyllabification which applies variably before sonorants such as /r/ and
/n/ in English. She found that in high frequency words, such as every and the noun evening , the
schwa was completely absent and the syllable it originally projected had vanished. In mid-frequency
words, such as memory and salary , the modal outcome is a syllabic /r/. In rare words, such as
1

mammary and artillery , the modal outcome is a schwa plus /r/. Another example is provided
by so-called t/d-deletion, which is widely acknowledged to be a case of variable undershoot of the
coronal articulation of the /t/ or /d/. Bybee (2000) found that deletion – defined as the inability of
the transcriber to hear the stop on a tape-recording – is more prevalent in high-frequency words than
in low-frequency words. The set of double-marked past tense verbs (such as told and left) provides
a way to control for the morphological factors which could play a part in this pattern. Within
the set of double-marked pasts alone, Bybee’s data showed a statistically significant relationship of
word frequency to the rate of /t/ deletion, with the most frequent word (told ) having /d/ deleted in
68% of cases while the least frequent (meant ) never had the /t/ deleted. Further documentation of
the association between word frequency and leniting historical change is provided in Phillips (1984,
this volume).
Although these frequency effects will be the main focus in this paper, is is also important to
acknowledge that word-specific allophony has been found in a number of other situations as well.
For example, Yaeger-Dror and Kemp (1992) and Yaeger-Dror (1996) demonstrate that words in a
particular cultural/semantic field in Montreal French have resisted a historical shift in the vowel
system and as a result display idiosyncratic vowel quality. Hay (2000) also presents data relating
degree of morphological decomposibility to degree of /t/ lenition in words such as ”shiftless”.
These results challenge standard models of phonology and phonetics at two levels. First, in all
standard models, the lexicon is distinguished from the phonological grammar. The exact phonetic
details of a word’s pronunciation arise because the word is retrieved from the lexicon, and processed
by the rules or constraints of the grammar whose result (the surface phonological form of the word)
is fed to a phonetic implementation component. The phonetic implementation component computes
the articulatory and/or acoustic goals which actualize the word as speech. The phonetic implemen-
tation component applies in exactly the same way to all surface phonological representations, and
the outcome depends solely on the categories and prosodic structures displayed in those represen-
tations. As a result, there is no way in which the phonetic implementation can apply differently to
some words than to others. If a phonetic implementation rule is variable and gradient, then the same
probability distribution of outcomes would arise for all words which meet the structural description
of the rule. This generic feature of modular generative models with phonetic implementation rules
is developed at more length in Pierrehumbert (1994).
A second challenge arises from the fact that the differential phonetic outcomes relate specifically
to word frequency. Standard generative models do not encode word frequency. They treat the word
frequency effects which are so pervasive in experiments involving priming or lexical decision tasks
as matters of linguistic performance rather than linguistic competence. Thus the intrusion of word
frequency into a traditional area of linguistics, namely the conditioning of allophony, is not readily
accommodated in the classical generative viewpoint.
If each word corresponded to a completely idiosyncratic phonetic signal, then results such as
Bybee’s could be readily formalized in a highly transparent scientific model. We would simply
assume that holistic gestural or acoustic templates are associated with word meanings. The real
challenge arises from the fact that the classical view does provide important insights about the
mental representation of phonology. Although a word may have idiosyncratic phonetic properties,
it is perceived as made up of units of sound structure which are also shared with other words. The
2