• Aucun résultat trouvé

CHAPTER I. AN INTRODUCTION TO PHONOLOGICAL VARIATION, FRENCH SCHWA,

I.4 Speech production models and phonological variation

I.4.5 Further issues in speech production

The models described above provide the general features of the word production process. On this basis, further studies have investigated and documented more specific aspects of the production process. Some of these aspects, such as differences in lexical selection for open versus closed-class words, competition effects and contextual influences, are not trivial with regard to our research questions or methods. We will therefore briefly sketch the main discussions and findings of these studies in the following sub-sections.

39 I.4.5.1 Lexical selection: open versus closed-class words

In the models sketched above, the lexical selection process is described as if it were identical for all words. However, a distinction is traditionally made between open and closed-class words with regard to the lexical selection process (see Alario, Ayora, Costa & Melinger, 2008 for a review). A first difference lies in the kind of information driving the selection of these words. Open-class words are thought to be selected on the basis of semantic information only, whereas closed-class words are selected on the basis of diverse types of information, including the linguistic context (Alario & Caramazza, 2002).

Furthermore, several studies have documented that the process of open-class word selection entails competition at the semantic level while the situation for closed-class word selection is less clear7. Schriefers, Jescheniak and Hantsch (2002) suggest that closed-class words are selected via the same principle as open-class words, i.e., through activation levels and competitive selection. In contrast other authors suggest that closed-class words selection is not competitive. In Alario and Caramazza (2002, see also Alario et al., 2008) for instance, the selection of determiners depends on their activation level but not on the activation level of other candidates.

Additional evidence in favor of the assumption that open and closed-class words are processed differently comes from a corpus study by Bell, Brenier, Gregory, Girand and Jurafsky (2009). These authors show that these two classes of words are affected differently by lexical frequency and predictability. Open-class words appear to be shorter if they are more predictable or more frequent, whereas the length of closed-class words is less affected by frequency and predictability.

Many studies have been interested more particularly in determiner selection. Using a picture word interference paradigm (i.e., participants are presented with a picture (target) to be named with its determiner, and with a word (distractor) they have to ignore), it was shown that latencies increase when target and distractor have different grammatical genders.

Interestingly, this “gender congruency effect” is present in Dutch (e.g., Schriefers, 1993), but

7 Such evidence has been gathered through the interference paradigm in picture naming. In this task, participants have to name the picture and ignore a distractor, presented auditorily or visually. Latencies are usually longer if the distractor is semantically related to the noun to be produced (see Alario et al., 2008 for a review).

40

not in several Romance languages (e.g., French or Italian). This absence of effect was related to the fact that in these languages, grammatical gender is not sufficient to select the right determiner. In French, for instance, the singular masculine demonstrative determiner ce

‘this’ has two phonological forms, ce [s] and cet [st]. Ce is used when the following word starts with a consonant (e.g., ce chien ‘this dog’) while cet is used when the following word starts with a vowel (e.g., cet arbre ‘this tree’). Similarly, in Italian, information about the immediate phonological context is necessary in some cases. For instance, the masculine singular determiner has two forms, il [il] and lo [l]. The choice of the appropriate determiner depends on the phonological properties of the following noun (lo is appropriate when the following noun starts with a vowel, an affricate, the cluster /!n/ or any cluster of the form “/s/+ consonant”, see Miozzo & Caramazza, 1999). In French and Italian, several experiments have failed to find a gender congruency effect (see for example Alario &

Caramazza, 2002 for French and Miozzo & Caramazza, 1999 for Italian). Miozzo and Caramazza (1999) suggest that for these languages, the noun’s grammatical gender and number activate a set of determiner allomorphs, but that the selection of the specific allomorph has to wait for the insertion of the following word’s phonological form into the phonological phrase (late selection hypothesis).

I.4.5.2 Competition at the lexeme level

In Levelt et al. (1999) lexemes do not enter into competition with each other. In this respect, the production process differs crucially from the recognition process. In contrast, Baayen (2007) argues in favor of a mechanism of word form selection similar to what can be found in recognition: lexical competitors could be co-activated together with the lexeme to be selected. According to this author, a word’s phonological word form should not be seen as a static representation. Instead it should be seen as the outcome of a dynamic competition process. Arguments in favor of such a competitive mechanism are found in phonetic and psycholinguistic studies.

In Scarborough (2004), for instance, more confusable words (i.e., words with lower frequencies than their phonological neighbors) exhibit more coarticulation. This increase in coarticulation is interpreted as a way to facilitate listeners’ perception. Similarly, Van Son and Pols (2003; see also Kuperman, Pluymaekers, Ernestus & Baayen, 2007) found that the fine phonetic details of a given segment in the word reflect the information load of that segment, i.e., the extent to which the segment, in a given situation, contributes to the

41

reduction of uncertainty in the cohort of candidates. These studies suggest that the more intense the process of competition, the more the unique properties of a word become relevant to distinguish it from its competitors: greater lexical competition leads to greater articulatory precision.

Evidence in favor of a competitive mechanism in lexeme selection can also been found in psycholinguistic studies. For instance, Jones (1989), using a tip-of-the-tongue elicitation task, showed that the ability to retrieve the target word decreases when a distractor is presented that is phonologically similar to the target word. Sevald and Dell (1994) found that sequences of words with similar onsets (e.g., “cat, cab, car”) are produced with longer latencies than sequences with different onsets (e.g., “cat, bat, mat”). In the same vein, Yaniv, Meyer, Gordon, Huff and Sevald (1990) found longer latencies in the production of pairs of syllables when the two syllables had similar vowels than when they had different vowels. In Spanish, Vitevich and Stamer (2006) showed that picture naming latencies were longer for objects whose Spanish names have many phonological neighbors than for those with few.

All these studies show inhibition effects caused by phonological similarity between word forms, and thus suggest that these forms are in competition during word form selection.

There is, however, also some evidence suggesting that phonologically similar words facilitate word form activation and retrieval. Vitevich (2002), for instance, showed that words with fewer phonological neighbors are produced with longer latencies and more errors than words with many phonological neighbors in picture naming and speech error elicitation tasks. Similarly, Jescheniak and Schriefers (2001) using a cross-modal picture-word interference paradigm, showed that naming latencies are shorter when the distractor is phonologically related to the target word than when it is unrelated.

According to Vitevich (2002), studies showing facilitory and inhibitory effects of phonological similarity between words all suggest that several lexemes are activated simultaneously, with the one lexeme corresponding to the intended word. Inhibitory and facilitative effects of these other word forms probably arise at different levels of representation.

I.4.5.3 Context effects in speech production

We know from linguistic studies that contextual factors affect the output of the production process. For instance, several studies have shown that the predictability of a word has an impact on the acoustic realization of its segments (Torreira & Ernestus, 2009; Pluymaekers,

42

Ernestus & Baayen, 2005; Kuperman et al., 2007). Other studies have documented an effect of speaking style on vowel articulation (e.g., Harmegnies & Poch-Olivé, 1992).

Psycholinguistic studies, on the other hand, have not yet systematically investigated the effects of contextual factors on word production. As a consequence, the role of context, how it comes to influence the production process and the time course of this influence are not formalized in detail in word production models. It is likely that context does not only affect a word’s realization but also its activation and/or selection. A similar line of reasoning is found in Pluymaekers and colleagues (2005). In order to account for predictability effects on segment reduction, they provide several alternative explanations. One of these explanations assumes a direct link between the amount of activation of a word and the rapidity of its articulation. A higher predictability leads to a higher degree of activation. This higher degree of activation allows for the activation to spread to the word’s constituent segments more quickly, resulting in a quicker preparation of the speech sounds and thus a shorter duration.

Some mentions of contextual influences on word production processes can also be found in Levelt (1989), who discusses the influence of register and speech rate. Concerning the influence of register, he suggests that in casual speech, speakers are biased into using a particular subset of their lexicon (e.g., cop rather than policeman) but also into using particular allomorphs (e.g., I’ve rather than I have). As for speech rate, it may affect phonological encoding, leading for instance to assimilation and reduction phenomena. It may also affect segments after syllable access.

Empirical on-line studies should now investigate these proposals in detail. This issue is especially important for the study of phonological variation. We know that contextual factors are crucial in determining the choice of the variant to be produced. How this influence can be formalized in current psycholinguistic models of speech production remains, however, to be resolved. Furthermore, the design of experiments on phonological variation has to take account of the fact that contextual factors influence variant choice and possibly the time course of the production of words with several variants. Firstly, these influences must be considered in the choice of the linguistic context in which the phonological variants will be produced. We know for instance that in isolation, schwa variants are more likely to be produced than non-schwa variants. This means that whenever possible, a more neutral context should be chosen for a production experiment. Secondly, one must be aware that the overall contextual characteristics of the experimental setting may strongly influence the study’s outcome. Psycholinguistic experiments are formal situations and will therefore call

43

for a higher register than everyday conversations. This will probably influence variant choice. These issues are further examined in the next section, where methodological questions concerning the study of phonological variants are discussed.

I.5 Methodological tools to study the production of phonological