Using Statistical Distributions to Learn Abstract Phonotactic Constraints

French liaison has long been a favourite testing ground for phonological theories, a situation which can undoubtedly be attributed to the complexity of the phenomenon, involving in particular phonology/syntax, phonology/morphology, phonology/lexicon interfaces. Dealing with liaison requires stepping into all the components of the grammar, while tackling at the same time the quicksands of variation. The data on which a number of formal analyses are based are a source of concern since liaison, in part because of its intrinsic variable character, requires extensive and robust data. In the wake of the results from the study of other corpora, we present here extensive results based on the PFC database (Phonologie du français contemporain: usages, variétés et structures: Durand, Laks, Lyche 2003, Durand 2005) and point to their implications for models of linguistic structure. On the basis of a study of 195 speakers (28893 liaison codings) from a wide-range of survey points and three registers (reading a text aloud, formal and informal conversations), we first of all provide global results concerning liaison. We show that liaison consonants are reduced to five segments which fall clearly into two sets [z, n, t] vs. [r, p]: (1) Frequency of liaisons : [z]: 4544 > [n]: 3689 > [t]: 1665 > [r]: 13 > [p]: 9 Apart from the fact that [r] and [p] are not attested for the majority of speakers, they only occur in variable environments. These results also show that [n] is more frequent than [t], as against the opposite ranking given by most earlier specialists (e.g. Léon, 1992). We also demonstrate that, while a number of contexts show categorical behaviour in European French (e.g. Det + vowel initial X within NPs: ses[z]amis, ses[z]autres amis), other contexts traditionally assumed as environments for ‘obligatory’ liaison are seriously threatened (Adj + N, Prep + NP). Moreover, within each context we show that the phenomena cannot be dealt with category by category but often word by word or more precisely wordform by word-form taking into account prosodic weight. Thus inflectional endings do not have a uniform liaison behaviour and for the majority of informants only forms of the auxiliaries avoir and être (e.g. était, avait) exhibit liaison behaviour. The same inflections in other verbs (e.g. voulait, devait, etc.) do not trigger liaison at all. These observations and others underline that usage-based models (e.g. Bybee 2001, 2005, 2007) cannot be dismissed too lightly as often assumed in more ‘theoretical’ treatments. Finally we examine the three following assumptions made in the classical generative tradition: (2) (a) Enchaînement (systematic forward linking petit ami = [p ti-tami]) (b) Strong and regular link with phrasal syntax (e.g. X-bar theory) (c) The liaison consonant is underlying and belongs to the linking word Assumption (2)(a) can be supported from our data but, of course, in other styles of speech studied by Encrevé (1988), ‘liaison non enchaînée’ is well attested. We examine the consequences of these observations in the light of developmental data provided by Chevrot, Dugua & Fayol (2005) and Brau & Wauquier (2005) who show that the liaison consonant, if present in the speech of pre-school children, is always linked forward. The question of literacy therefore needs to be further studied to make sense of Encrevé’s data. We show that (2)(b) is highly questionable: liaison behaviour does provide evidence for syntactic structure but not for any strongly articulated model. As far as (2)(c) is concerned, there is an abundant literature advocating all the possible solutions as to the source of the liaison consonant (coda of the linking word, floating, epenthetic, prefixal). Morin in a series of insightful papers (1986, 2003, 2005) has argued that liaison consonants have been reinterpreted as prefixal. Our