How much influence can phonology have on a surface syntactic form? An answer at one extreme is that morpheme choice and order are fully determined before the phonological content of those morphemes is accessed; all the phonological module can do is fill in their sound content– that is, phonology is late. At the other extreme– early phonology– phonological considerations have the same qualitative ability as syntactic considerations to influence morpheme choice and order, and even syntactic structure.
In practice, both theoretical and processing models have adopted a variety of intermediate positions. For example, Embick and Noyer (2001), working within Distributed Morphology, adopt an essentially late-phonology stance, but with Local Dislocation, a restricted mechanism by which morpheme order can change in response to phonological information. By contrast, Anttila, Adams, and Speriosu (2010), building on Bresnan et al.’s (2007) work on the English dative alternation, propose an Optimality Theoretic (Prince and Smolensky 1993) grammar in which the syntactic constraint “The goal NP forms an XP with its head” competes directly with phonological constraints such as “No stress clashes within a phonological phrase” (p. 962). While the syntactic constraint is one of the highest-ranked in the grammar, it is not qualitatively prior to phonological constraints. This short article will not attempt a full history of this debate. Pullum and Zwicky (1988 and others) argue for late phonology, and exclude “preferences and tendencies” (p. 274) from being possible counterexamples. Within the field of phonology, however, preferences and tendencies have been treated as qualitatively similar to categorical grammaticality; it has been argued that the same factors that can render a form phonologically ungrammatical in one language or context can have a gradient effect in another, or even that apparent categoricalness is merely an extreme form of variation (e.g., Pierrehumbert 1994; Boersma 1998; Cohn 2006; Hayes and Londe 2006 for a variety of perspectives). Shih (2014) provides a recent survey of the issue, including new cases from English variation that support early phonology.
The goals of this paper are empirical and methodological. On the empirical side, it argues that phonological information must be allowed to influence, albeit probabilistically, whether French de ‘of’ appears once or twice within a coordinated structure. We will consider whether the French case must reflect phonological influence on choice of syntactic structure, or could be accounted for by giving the phonology the power to delete the second de. This case thus adds to the body of data whose implications for the syntax-phonology interface can be debated.
On the methodological side, this paper strives to meet the evolving standards of reproducible research (Stodden et al. 2014 is a recent handbook). Replication of a study typically begins from data collection, such as running a new experiment, or gathering a new corpus. If another researcher can perform these steps and obtain similar results, confidence in the original finding increases. The term reproducible research, on the other hand, refers to research that makes it easy for a reader to start from the same raw data as the original study, and re-generate and scrutinize the original analysis– that is, only part of the original research pipeline is repeated. Reproduction is useful as a substitute for replication when it would be infeasible to run a new experiment, gather a new corpus, etc.: with much less effort, many of the benefits of replication can still be obtained. This paper begins from a publicly available corpus, and the accompanying web materials, archived at TROLLing (The Tromsø Repository of Language and Linguistics, opendata.uit.no/dvn/dv/trolling), include all scripts used for data processing and analysis, with ample documentation, so that the interested reader can reproduce all results, statistical tests, and figures; make changes to the assumptions and methods; and look for errors, with only one departure from the script for some hand-coding, which the reader can either use as-is or perform independently.
1 Background: French de and coordination
The French preposition de, meaning roughly ‘of’, has two properties crucial to this study: it can optionally take wide scope in coordination, and it has different allomorphs depending on whether the following word begins with a consonant or a vowel. This section describes each of these properties in turn.
French de can optionally take wide scope over a coordinated structure. Examples are given in (1), with de in boldface, brackets marking selected constituents, and the two coordinands underlined.
‘pieces of tomatoes and (of) carrots’
Miller (1992) appears to be the earliest formal treatment of wide scope versus repetition for French de (among other function words). (The terms wide scope and reduction will be used interchangeably below to describe data like (1)b; repetition will describe cases like (1)a.)
This study examines whether phonological properties of the words involved affect the frequency of wide scope. I assume the syntactic bracketings suggested in (1): for repetition, constituents [de X] and [de Y] are coordinated; for wide scope/reduction, de is outside a constituent [X & Y]. The conclusion will consider (and reject) the alternate possibility that de never actually takes wide scope, and (1)b has the structure [[detomates] et [de carottes]].
There are some non-phonological restrictions on wide scope. Miller cites claims from a classical and still popular descriptive grammar (Grevisse 1980) about the conditions under which de can take wide scope, but he does not agree with all of them. Miller concludes that repetition is the unmarked case, occurring more frequently than wide scope and imposing no special semantic restrictions, whereas wide scope requires a group reading, at least in some cases (p. 162).
Tseng and colleagues, in a series of articles (Tseng 2003; Abeillé et al. 2004, Abeillé et al. 2005, Abeillé et al. 2006) make several further observations. They divide de into what they call oblique and non-oblique versions, where only the oblique de can take wide scope. Non-oblique de occurs with partitive the+N’, bare N’, and VP; oblique de occurs with NPs, PPs, APs, AdvPs, or lexical Vs (Abeillé et al. 2006, p. 151): 1
|Non-oblique– repetition required|
‘you need flour and baking powder’
‘a lot of bread and wine’
‘I dream of reading this book and explaining it to my son’
|Oblique– wide scope allowed|
‘I need this flour and this baking powder’
‘He’s coming back from Paul or Marie’s’
‘someone good at math and strong in gym’
‘I dream of reading and explaining this book to my son’
This study looks only at the coordination of single words. Unfortunately, it is not possible to tell whether a single N is itself an NP, or whether a single V is a VP, so inevitably the data contains some oblique instances (where repetition of de is claimed to be obligatory), mixed in with the non-oblique instances (where the variation of interest would lie).
Some of the phonological hypotheses to be tested below arise from the well-known fact that de has two allomorphs: de [də] is used before a consonant-initial word, and d’ [d] before a vowel-initial word. Certain complications arise in glide-initial words and some words that are spelled with initial silent h, but those cases will be excluded from the data here. Some examples, from Abeillé et al. (2004 p. 16):
|CC-initial:||de clients||[dəkliã]||‘of clients’|
|C-initial:||de livres||[dəlivʁ]||‘of books’|
For phonological information to influence the choice between repetition and wide scope, there mustbephonological differences among coordinands that cause omitting de to have different consequences. 2 That is, reduction of de has to produce a better result for some words than for others, because of phonological properties of those words. The sense in which the result is worse (or better) could be phonological (relying on known properties of French phonology or of phonological typology generally), could involve processing, or could even involve the syntax-phonology interface.
Some of those consequences arise from the de/d’ allomorphy. The following four cases arise for coordinating words X and Y, where C stands for ‘consonant-initial word’, V stands for ‘vowel-initial word’, and & stands for et ‘and’ or ou ‘or’:
|C&C: de X et/ou (de) Y||de biens et (de) services||‘of goods and (of) services’|
|C&V: de X et/ou (d’)Y||de juillet et (d’)août||‘of July and (of) August’|
|V&C: d’ X et/ou (de) Y||d’aller et (de) venir||‘of going and (of) coming’|
|V&V: d’ X et/ou (d’)Y||d’Adam et (d’)Eve||‘of Adam and (of) Eve’|
The allomorphy suggests the following hypotheses:
|Hiatus avoidance: 3Et [e] and ou [u] both end with vowels, so reduction when the second conjunct begins with a vowel causes a V-V sequence (e.g., et Eve [e ɛv] vs. et d’Eve [e dɛv]). Assuming this is penalized, reduction is less likely before V. (E.g., Prince and Smolensky’s (1993) Onset; for the role of hiatus avoidance in French specifically, see Tranel (1996), among many others)|
|→ more reduction in X&C than in X&V|
|Dependent d’: Although the allomorph de has schwa as its sole vowel, and therefore does not qualify as a phonological word of French, it can nonetheless stand alone as the citation form of the morpheme, whereas d’ cannot– it can only be uttered when cliticized to another word. Assuming that de is a more independent unit than d’, de is more likely to take wide scope. That is, there is some penalty incurred by the structure [d’ [X & Y]], where d’ forms a prosodic but not a syntactic unit with X.|
|→ more reduction in C&X than in V&X (where X is either C or V)|
|Parallelism: As mentioned above, we will consider below the possibility that wide scope is actually the result of deleting the second de. Deletion in conjunction is normally of a repeated element, so it should be more likely when the two des, if both present, would appear in the same allomorph. That is, reduction is more likely when the conjuncts would take the same allomorph.|
|→ more reduction in C&C and V&V than in C&V and V&C|
All three hypotheses require phonological information about the two conjuncts to, at least sometimes, be available before the final decision about reduction-versus-repetition is made
A final hypothesis requires not only prosodic information (C-initial or V-initial, syllable count) but even segmental information to be available:
We will see, however, that this last hypothesis is not supported.
The methods section of this paper is in two parts: the brief version here, and the full details given in the much-longer HTML file that accompanies the paper. The HTML file gives a full narrative of the method, from downloading the initial corpus files to the final regression models. The narrative is interspersed with extensively commented code and the output of that code, including figures. The HTML file is more than a lab notebook, though, because it is generated by an.Rmd (“R markdown”) file, which contains the narrative and the R code (R Core Team 2014); the.Rmd file runs the R code and generates the HTML file using the knitr package (Xie 2014a, Xie 2014b). This makes it easy to alter and re-run the analysis. Also downloadable are a Python script and helper file that the R code calls to run part of the analysis, and various intermediate versions of the data files for readers who wish to skip steps.
Data come from the Google n-grams corpus (Michel et al. 2011; Lin et al. 2012) for French. The raw data (not available to the public and not used here) are scans of books published in French, digitized through optical character recognition. The downloadable n-gram corpus used here compiles word sequences (such as d’ aller et de venir) and reports, for every publication year, how many tokens of that sequence occurred.
As the accompanying HTML file details, relevant 4- and 5-grams were extracted. The data used come only from the year 1900 and later, exclude conjuncts that begin with glides or orthographic h, exclude numerals, and include only conjuncts that are semantically parallel (e.g., two places, two materials, two occupations).
To test the hypotheses outlined above, each word1/conjunction/word2 combination’s reduction rate is calculated. This will be used as the dependent variable in the plots below. For example, d’aller et de venir occurs 7,616 times and d’aller et venir occurs 12,696 times, so the reduction rate for aller/et/venir is 0.63. The reduction count used in the Poisson regression model below would be 12,696 in this case, and the exposure would be 7,616+12,696.
Some independent variables used below were calculated from the n-gram dataset: total frequency of word1/conjunction/word2, initial sound types of word1 and word2. Others were looked up in the Lexique 3.80 database (Matos et al. 2001; New et al. 2004): word1 and word2 frequency, syllable count, and part of speech.
After various types of data cleaning, the final data set includes 8,332 word1/conjunction/word2 triples, representing 10,061,917 tokens.
4.1 Descriptive statistics
The overall reduction rate is strongly skewed, with most items tending not to undergo reduction. This is in line with Miller’s 1992 treatment of repetition as the unmarked case. The following histogram illustrates the distribution of reduction rates:
Note that there is a poor-get-poorer artefact here: the Google n-grams corpus includes only items that occur in at least 40 books (over the whole timespan of the corpus; the number of occurrences used here, since 1900, may be smaller). Suppose that de X et de Y occurs 270 times (with hits coming from 40 or more books) and reduced de X et Y occurs only 30 times in the whole corpus, falling below the threshold. The true reduction rate is 10%, but in the Google n-grams corpus it will appear to be 0%. The real distribution must be slightly less skewed than what we see in Figure 1.
To test the hypotheses concerning the allomorphs de and d’, the following bar plot breaks down mean reduction rates by the conjuncts’ initial consonant type, separately for et and ou. Each category is coded as word1 type, conjunction, word2 type. For example, C et V means that the first conjunct is a consonant-initial word, the second is vowel-initial, and they are coordinated with et.
Figure 2 shows that, within both et and ou (which differ in their overall reduction rates), there is more reduction if word1 begins with C than if it begins with V (consistent with the Dependent d’ hypothesis), and more if word2 begins with C than V (consistent with Hiatus avoidance/Lapse avoidance). It is hard to see from the plot whether Parallelism is supported – e.g., whether there is more reduction for V&V than would be expected from the overall rate of reduction for V in each position.
Reassuringly, the pattern holds within each part of speech, as plotted in Figure 3. Omitting adverbs, for which data are sparse, we can see the same pattern within adjectives (ADJ), nouns (NOM), and verbs (VER), though the overall rates differ by part of speech. This makes it unlikely that the effects are merely driven by particular words or constructions.
The lower rates seen for nouns and verbs could be because of the mixture of N’ and NP, and V and VP mentioned in Section 1.
Finally, one might wonder whether these phonological effects are frozen in high-frequency items, or apply productively to new combinations. To test this, word1/coord/word2 triples are divided into four frequency quartiles (frequency of 8–128, 129–474, 475–1,237, and 1,238–110,818), and the data are plotted separately for each quartile in Figure 4. The only place where the pattern fails to hold is for V et C in the lowest frequency quartile. Overall reduction rates are lower in the lower frequency quartiles (as confirmed in the regression model below). It seems plausible that the more frequent a combination, the more likely the shorter form (e.g., Bybee 2001 on reduction of all kinds in frequent items). We can also observe that the et/ou difference is driven by the lowest frequency quartile.
4.2 Regression models
Because reduction rates are so strongly skewed, with most tuples having a value near zero, Poisson regression is used. (See HTML file for full regression details and results.)
Each observation is one word1/conjunction/word2 combination. The dependent variable is the countofreduced tokens, and the log total number of tokens (exposure) is used as an independent variable. 4A random intercept is included for word2– that is, the model allows that each word2 could have its own idiosyncratic baseline rate of reduction (there is no random intercept for word1 because of convergence problems; it seems less plausible that idiosyncratic properties of word1 would have a large influence). The fixed effects are as follows:
log frequency of the word1/conjunction/word2 combination
log frequencies (from Lexique) of word1 and word2
part of speech: baseline is adjective, with adverb, noun, and verb being contrasted
word1 type (initial sound is C or V) and word2 type
whether word1 and word2’s initial sound types match 5
conjunction: et or ou 6
whether word2 starts with [d]
|Poisson regression model (fixed effects)|
|post.mean||l-95% CI||u-95% CI||eff.samp||pMCMC|
|log word1 frequency||0.0549||–0.2027||0.3104||605.23||0.664|
|log word2 frequency||–1.2920||–1.6523||–0.8781||586.41||<0.001||***|
|POS = adv. (vs. adj.)||–0.8498||–2.2353||0.2852||633.15||0.198|
|POS = noun (vs. adj.)||–0.8954||–1.3151||–0.4978||806.91||<0.001||***|
|POS = verb (vs. adj.)||1.5921||0.7178||2.4504||684.17||<0.001||**|
|w1 is V-initial (vs. C-init.)||–2.2124||–2.5491||–1.8514||246.40||<0.001||***|
|w2 is V-initial (vs. C-init.)||–1.7439||–2.2163||–1.1508||342.31||<0.001||***|
|w1 type (C or V) = w2 type||0.2865||0.0065||0.6134||293.61||0.056||.|
|conjunction = ou||–0.7711||–1.0476||–0.5205||356.52||<0.001||***|
|word2 begins with d||–0.0584||–0.7841||0.8069||346.72||0.892|
Because the regression model fits logs of reduction counts, the results are interpreted differently than in other types of regression. For example, when comparing two w1/conj/w2 combinations, if the log frequency of word1 is greater in one of the combinations by 1 standardized unit (i.e., 2 standard deviations), then, all else being equal, the difference in the logs of their reduced-form counts is predicted to increase by 0.0549. As in other types of regression, a positive coefficient estimate means a factor favors de-reduction and a negative value means it disfavors it.
Considering non-phonological predictors first, combinations with higher frequency are subject to more reduction, as we have seen; the frequency of word2 has a smaller effect, in the other direction (a more-frequent word2 has less reduction). As we saw in the plots above, ou has a lower rate of reduction than et.
Initial regression models found more reduction when word2 begins with [d], but using the final Poisson approach, this factor made no significant contribution.
As for allomorphs of de, as we saw in the plots above, reduction is less frequent if either word1 or word2 begins with a vowel. There is also a small effect of matching: if word1 and word2 have the same status (C&C or V&V), then reduction is slightly more frequent than would be expected from the statuses of word1 and word2 independently.
5 Summary, discussion, and conclusion
The following Table 1 summarizes the results obtained for the hypotheses put forth in Section 2:
|Hiatus avoidance||Strongly supported:|
Less reduction when word2 is V-initial, and thus would produce hiatus if d’ is omitted
|Dependent d’||Strongly supported:|
Less reduction when word1 is V-initial, and thus would take d’
|Parallelism||Supported (significant, but small coefficient):|
Slightly more reduction when word1 and word2 would take the same allomorph of de.
No increase in reduction when word2 starts with [d], so that retaining de would yield [dəd] sequence
It is hard to see how the French data could be consistent with a theory in which phonological information has no influence at all on the choice to repeat or not repeat de. Whether the data are consistent with a late, limited mechanism or require an earlier or more active role for phonology will depend on the syntactic analysis of wide-scope de.
One possibility is that there really are two syntactic structures, as given in (1): [de X] & [de Y] and [de[X&Y]]. In that case, the syntax could produce multiple candidates, which are then operated on by the phonological component, and the final choice is made at least partly on grounds of phonological well-formedness. Or, still assuming two different syntactic structures, phonological information could simply be available while syntactic decisions are being made. From a processing point of view, access to phonological information would most likely unfold gradually, so that it couldn’t be depended on for a categorical syntactic rule, but on occasions when phonological information was available before the speaker had committed to a syntactic structure, it would have a say in the decision (see Wagner 2011 on the idea that allomorphy is constrained by availability of phonological information during production planning). In both proposals, phonology is early.
Another possibility is that the syntax outputs a structure [[de X] & [de Y]], without reference to phonological information, but the phonology can delete of the second de, presumably without subsequent syntactic reorganization. That is, the syntactic structure of reduced forms would be [[de X] & [de Y]], and the first de would not in fact take wide scope. This proposal would maintain that phonology is late.
Several authors have weighed in on the question of coordination vs. deletion when it comes to coordination in compounds or affixed words, as in English half-(siblings) or step-siblings, step-sons and (step)-daughters. Booij (1985), working mainly with Dutch and German data, argues that these structures arise through deletion rather than coordination. His reasons include: it is possible to “coordinate” non-constituents, as in the difference between a third- and a sixth-grader; it is possible to “coordinate” items of different syntactic category, as in Dutch wis- en sterren-kunde ‘math and astronomy’, literally ‘sure- and stars-knowledge’; and it is not only conjunctions that allow this kind of deletion, but also other constructions, including prepositions, as in how to distinguish neuro- from psycholinguistic claims (Chaves 2008, p. 266). Chaves (2008) adds to Booij’s arguments the observation that utterances like the following are possible: Pre- and post-revolutionary France were very different from each other (p. 264). Two different versions of France are in question (pre-revolutionary France and post-revolutionary France), not a single version of France that is [pre- and post-]revolutionary. Artstein (2005), by contrast, argues for coordination, in part because of a claimed semantic difference between Mary and Bill are orthodontists and periodontists (they must each practice both specialties) and Mary and Bill are ortho- and periodontists (Mary and Bill may each have a different dental specialty). Artstein also points out that it seems strange for a phonological deletion rule to apply only next to conjunctions (and some prepositions): why not *My step-brother likes his half-brother? More generally, allowing the phonology to have a deletion rule that is sensitive to repetition in coordinated structure pushes at the boundary of the types of syntactic information that should be available to the phonology.
But the main challenge for treating de reduction as like phonological deletion is that it would be a rather different rule from within-word deletions. Usually in languages that allow deletion or coordination of word parts, both the remnant and the deleted part have to be at least a phonological word (or, in Artstein’s 2005 analysis, a separate foot). Addressing the size of the remnant, Abeillé (2006) says, “Coordination of prefixes is marginally available in French, but only for prefixes with some phonological autonomy” (p. 3), and gives these examples: sur- ou sous-évaluer ‘over- or under-evaluate’, but *re- ou dé-faire ‘re- or un-do’; [ʁə] lacks full vowel and is presumably sub-minimal as a French word. Addressing the size of the deletedpart, Booij 1985 contrasts examples such as regel-ordening en –toepassing ‘rule-ordering and –application’ with *ge-hijg en –puf ‘gasping and puffing’, where ge- is not a foot or a phonological word. In French de X & Y, the remnant of deletion would be Y, which poses no prosodic problem, because Y will always be at least a full phonological word. The deleted portion, however, is the subminimal clitic de or d’. Thus, de omission would be different from the within-word coordination/deletions discussed by Booij (1985), Artstein (2005), Chaves (2008), and Abeillé (2006). It is possible that within-word coordination really is deletion, but that de omission instead reflects a different, wide-scope syntactic structure. 7
In summary, the choice between French de X & de Y and de X & Y depends on phonological properties of both X and Y, including their influence on what allomorph of de is used. This could mean that the choice is itself phonological one (whether to phonologically delete the second de), but deletion would have to be phonologically different from deletion that has been argued for within words. Or, it could mean that phonological information is available early enough to influence syntactic structure.
Thanks to Stephanie Shih for discussion of this case, including some of the hypotheses tested below; to Bruce Hayes for discussion of the data set and suggesting an additional hypothesis; to Michael Wagner for discussion of compounds. For general discussion, thanks to Morgan Sonderegger and participants in the UCLA phonology seminar and the Boston University linguistics colloquium, especially Byron Ahn, Carol Neidle, Neil Myler, and Jon Barnes. Thanks to a reviewer and the associate editor for thorough comments. Input from participants in the University of Leipzig Institut für Linguistik guest lecture series came after the paper had been finalized, but special thanks to Matías Guzmán Naranjo for suggesting TROLLing. Thanks to Christine Wells, Phil Ender, Joni Ricks, Brian Kim, Alex Whitworth, and especially Andy Lin, all of the UCLA Institute for Digital Research and Education, for statistical advice and assistance.
Abeillé, Anne. 2006. In defense of lexical coordination. In Olivier Bonami & Patricia Cabredo Hofherr (eds.), Empirical Issues in Syntax and Semantics 6, 7–36. www.cssp.cnrs.fr/eiss6/index_en.html
Abeillé, Anne, Olivier Bonami, Danièle Godard & Jesse Tseng. 2004. The syntax of French de-N’ phrases. In S Müller (ed.), HPSG 2004: Proceedings of the 11th international conference on head-driven phrase structure grammar, 6–26. Stanford: CSLI Publications. Google Scholar
Abeillé, Anne, Olivier Bonami, Danièle Godard & Jesse Tseng. 2006. The syntax of French à and de: an HPSG analysis. In Patrick Saint-Dizier (ed.), Syntax and semantics of prepositions, 147–162. Dordrecht: Springer. Google Scholar
Bates, Douglas, Martin Maechler, Ben Bolker & Steven Walker. 2014. lme4: Linear mixed-effects models using Eigen and S4. R package version 1.1-6. http://CRAN.R-project.org/package=lme4
Boersma, Paul. 1998. Functional phonology: Formalizing the interaction between articulatory and perceptual drives. The Hague: Holland Academic Graphics. Google Scholar
Booij, Geert E. 1985. Coordination reduction in complex words: A case for prosodic phonology. In Harry Van der Hulst & Norval Smith (eds.), Advances in nonlinear phonology, 143–160. Dordrecht: Foris. Google Scholar
Bresnan, Joan, Anna Cueni, Tatiana Nikitina & Harald Baayen. 2007. Predicting the dative alternation. In G Boume, I Kraemer & J Zwarts (eds.), Cognitive foundations of interpretation, 69–94. Amsterdam: Royal Netherlands Academy of Science. Google Scholar
Bybee, Joan. 2001. Phonology and language use. Cambridge: Cambridge University Press. Google Scholar
Cohn, Abigail. 2006. Is there gradient phonology? In Gisbert Fanselow, Caroline Féry, R Vogel & Matthias Schlesewsky (eds.), Gradience in grammar: Generative perspectives, 25–44. Oxford: Oxford University Press. Google Scholar
Green, Thomas & Michael Kenstowicz. 1995. The Lapse constraint. In Leslie Gabriele (ed.), Proceedings of the sixth annual meeting of the formal linguistic society of mid-America, vol. I, 1–14. Bloomington, IN: Indiana University Linguistics Club.Google Scholar
Grevisse, Maurice. 1980. Le bon usage. Gembloux: Duculot. Google Scholar
Lin, Yuri, Jean-Baptiste Michel, Erez Lieberman Aiden, Jon Orwant, Will Brockman & Slav Petrov. 2012. Syntactic Annotations for the Google Books Ngram Corpus. Proceedings of the ACL 2012 System Demonstrations, 169–174. (ACL’12). Stroudsburg, PA, USA: Association for Computational Linguistics. http://dl.acm.org/citation.cfm?id=2390470.2390499
Löfstedt, Ingvar. 2010. Phonetic effects in Swedish phonology: Allomorphy and paradigms. UCLA Ph.D. Dissertation. Google Scholar
Matos, Rafael, Ludovic Ferrand, Christophe Pallier & Boris New. 2001. Une base de données lexicales du français contemporain sur internet : LEXIQUE. L’année psychologique 101(3). 447–462. CrossrefGoogle Scholar
Michel, Jean-Baptiste, Yuan Kui Shen, Aviva Presser Aiden, Adrian Veres, Matthew K. Gray, Joseph P. Pickett, Dale Hoiberg, et al. 2011. Quantitative analysis of culture using millions of digitized books. Science 331(6014). 176–182. CrossrefWeb of ScienceGoogle Scholar
Miller, Philip H. 1992. Clitics and constituents in phrase structure grammar. New York: Garland. Google Scholar
New, Boris, Christophe Pallier, Marc Brysbaert & Ludovic Ferrand. 2004. Lexique 2: A new French lexical database. Behavior Research Methods, Instruments, & Computers: A Journal of the Psychonomic Society, Inc 36(3). 516–524. CrossrefGoogle Scholar
Pierrehumbert, Janet. 1994. Knowledge of variation. Papers from the Parasession on Variation, 30th meeting of the Chicago Linguistic Society. Chicago: Chicago Linguistic Society. Google Scholar
Prince, Alan & Paul Smolensky. 1993. Optimality theory. Manuscript, published 2004, Malden, MA: Blackwell. Google Scholar
Pullum, Geoffrey & Arnold Zwicky. 1988. The syntax-phonology interface. In F. Newmeyer (ed.), Linguistics: The Cambridge survey, vol. 1, 255–280. Cambridge: Cambridge University Press. Google Scholar
Raffelsiefen, Renate. 1999. Phonological constraints on English word formation. In Geert E. Booij and Jaap van Marle (eds.), Yearbook of morphology 1998, 225–287. (Yearbook of Morphology 8). Dordrecht: Kluwer. Google Scholar
R Core Team. 2014. R: A language and environment for statistical computing. Vienna: R Foundation for Statistical Computing. www.R-project.org
Shih, Stephanie. 2014. Towards optimal rhythm. Stanford University PhD dissertation. Google Scholar
Stodden, Victoria, Friedrich Leisch & Roger D Peng (eds.). 2014. Implementing reproducible research. (Chapman & Hall/CRC The R Series). Boca Raton: Chapman and Hall/CRC. Google Scholar
Tranel, Bernard. 1996. French liaison and elision revisited: A unified account within Optimality theory. In Claudia Parodi, Carlos Quicoli, Mario Saltarelli, Maria Luisa Zubizarreta, Claudia Parodi, Carlos Quicoli, Mario Saltarelli & Maria Luisa Zubizarreta (eds.), Aspects of Romance Linguistics, 433–455. Washington, DC: Georgetown University Press. Google Scholar
Tseng, Jesse. 2003. Phrasal affixes and French morphosyntax. In G Jäger (ed.), FGVienna: Proceedings of formal grammar 2003, 177–188. Vienna: Technische Universität Wien. Google Scholar
Wagner, Michael. 2011. Production planning constraints on allomorphy. Journal of the Canadian Acoustical Association 39(3). 160–161. Google Scholar
Xie, Yihui. 2014a. Knitr: A general-purpose package for dynamic report generation. R package version 1.6. cran.r-project.org/web/packages/knitr/index.html
Xie, Yihui. 2014b. Knitr: A comprehensive tool for reproducible research in R. In Victoria Stodden, Friedrich Leisch & Roger D Peng (eds.), Implementing reproducible computational research. Boca Raton: Chapman and Hall/CRC. Google Scholar
Yip, Moira. 1995. Repetition and its avoidance: The case in Javanese. In Keiichiro Suzuki & Dirk Elzinga (eds.), Proceedings of South Western Optimality theory workshop 1995, 238–262. Tucson: University of Arizona, Department of Linguistics. Google Scholar
Abeillé et al. (2006) also claims that in French, prefixes can’t take wide scope at all: *Il ne faut pas surévaluer et surnoter ‘One should not overevaluate and over-rate’. That is, the utterance is legal, but cannot have the meaning of over-rate, only of rate. In English, however, word-initial deletion requires a strong supporting context to be interpreted as such, and perhaps the same is true in French. In isolation, step-sons and daughters would probably not be interpreted as involving step-daughters, but in the context When Maria married Captain von Trapp, she acquired several step-sons and daughters, the step-daughters reading is salient if the listener is familiar with the movie The Sound of Music, in which then-childless Maria marries a widowed man with several sons and daughters.