University of New Hampshire

David Heap

University of Western Ontario

The distinction between languages where a subject either is or is not obligatorily expressed has been descriptively captured by the Null Subject Parameter, presented as follows:

The cluster of properties to which Chomsky (1981:240) refers is :

(1) i. missing subject

However, such binary parametric analyses have several shortcomings (some of which have been discussed in Safir (1985), Roberge (1989), Wanner (1993):

1. It is not possible to set a binary parameter to "sometimes" to capture inherent synchronic variation (i.e., different possible surface forms of one sentence), let alone correlations between this variation and other factors.
2. It is not possible to set a binary parameter to "sometimes" to capture the fact that many languages exist in which the parameter seems to be set [+] for some types of sentences and [-] for others (Heap 1997:39).
3. The parameter is arbitrary: one might just as well posit a Null Predicate Parameter or a Null Object Parameter.
4. The cluster of properties in (1) does not appear in all languages which exhibit some of the properties (Wanner 1993).

These shortcomings are inherent to any parametric approach which allows for only a binary analysis. Generative approaches with rules saying that the (underlying) Subject either must or must not be represented on the surface have the same shortcomings.

A very brief review of some relevant literature highlights the complexity of the issue. Brandi & Cordin (1989) show that subject clitics in Northern Italian varieties interact in complex ways with object clitics and with negation. They propose that subject clitics, unlike object clitics, are not part of the VP. Rizzi (1986) uses similar facts to argue for two functional heads for agreement, corresponding to two clitic clusters: one including subject clitics and negation, and the other direct and indirect object clitcs. Mair Parry (1997) shows, however, that the distribution of negation in Val Bormida varieties contradicts the categorical predictions of Rizzi's hypothesis. In these varieties, negation appears before or after subject clitics, and between object clitics, depending on grammatical person. These facts strongly suggests that the ordering of subject pronouns and other clitic morphemes, while strongly conditioned by (quasi-) universal principles, nonetheless remains flexible enough for these principles to be bent (if not broken) in specific instances.

The state of affairs described above more plausibly reflects a model based on violable constraints than one based on rigid principles and parameters. Furthermore, a constraint-based approach:

While a complete inventory of universal constraints has yet to be articulated, certain general types, including alignment and faithfulness constraints, recur with sufficient frequency in the phonological Optimality Theory (OT) literature that we can be fairly confident they will play an important role in constraint-based morphological analyses (McCarthy & Prince 1993). Competing constraints have already successfully accounted for many morphological issues (cf. Anderson 1994, 1995, Prince & Smolensky 1993, Bonet 1994, Anttila & Cho in press, Heap 1998).

While evidence exists to support this approach in other relevant Romance varieties, we restrict our discussion here to variable subject pronoun presence in a closely related group of Francoprovençal varieties. The geolects of the Gallo-Italo-Romance linguistic continuum provide an excellent site for comparing and evaluating the categorical parametric and the probabilistic constraint-based approaches. This group of languages exhibits considerable variation in subject pronoun usage (Heap 1997). Contrary to the predictions of the Null Subject Parameter these grammars have subject pronoun paradigms that are not only "partial," (subject pronouns are regularly used in some but not all grammatical persons) but also variable, conditioned by a number of linguistic factors.

In order to be able to evaluate these theories for their diachronic predictive power, we extend the scope of our study to include the Faetar (FT) speech community. This is a small but linguistically thriving community of Francoprovençal (FP) speakers who have survived in comparative isolation in Southern Italy for six centuries, since the original settlers emigrated from the Gallic FP region. Faetar has a variable subject pronoun system which is typologically more similar to the FP varieties of southeastern France, western Switzerland and northwestern Italy than to its southern Italian "Null Subject" neighbors. Its subject pronoun morphology varies between FP-like (subject pronoun usually present) and Italic types (subject pronoun usually absent). We show evidence of a historical change in subject pronoun usage by comparing contemporary Faetar speakers of different ages and by contrasting the patterns found by Jaberg & Jud (1928-40) in the early part of this century with patterns established by the more recent Faetar fieldwork (Nagy 1996).

Statistical analysis (using GoldVarb 2.0, Rand & Sankoff 1990) indicates that linguistic factors including grammatical person, type of nominal subject (noun, demonstrative, or empty), the subject's position in the clause, and the presence of object and/or reflexive proclitics have significant conditioning effects on subject pronoun usage. Similar effects have been shown in other Gallo-Italo-Romance varieties (Heap 1997). The effects of several of these factors converge to favor forms with the verb at the leftmost edge of its clause, a pattern easily captured in a constraint-based hierarchy, but one which does not fall out of a rule-based or principles and parameters approach.

We analyze the FP subject pronoun system using data from four sources:

The atlas data selected are from the atlas points listed in Table 1. All these varieties have (quasi-)categorical subject pronoun presence. The numbers in the rightmost column indicate the number of grammatical persons which normally have subject pronouns present, as determined by approximately 100 tokens (atlas maps) per town.

Subject constituents (both in Gallic FP varieties and Faetar) may appear in a number of different forms, as shown in (5). The first four types exist in FP and all types exist in Faetar.

These seven variants constitute the values of the dependent variable in the present study. There may be no surface subject, or there may be one subject pronoun present, which may be either the strong or the weak form (the latter including the generic [o]), and may occur in addition to a nominal subject. In Faetar, both the strong and the weak form may appear adjacently, without emphatic effect, but this pattern will not be addressed here. We also consider sentences with the ambiguous 'ki,' which seems to serve as a generic [+human] pronoun in main clauses, but as the COMP of a relative clause.

Seven linguistic factors were coded for each token (clause). They are:

(6) Linguistic factors

Factors i-v will be considered here, as they allow us to evaluate the relationship between subject pronoun presence and edge-alignment. They are directly relevant to the morphological analysis presented in Section 6.

In the analysis of the contemporary Faetar data (cf. (4.iii), three social variables are considered: age, sex, and intensity of contact with Italian. The speaker sample is balanced for these factors. The speakers range in age from 11 to 77 and are divided into 4 groups, roughly by generation: <20, 21-40, 41-60, >60. Four means of determining the intensity of contact with Italian were considered: amount of schooling (in Italian), language/place of work, presence of non-Faetani in the family, and residence elsewhere in Italy (Nagy 1996:47-8). These intensity of contact factors were not found to have significant effects and so will not be discussed further. No analysis of social factors is possible for the atlas data because there is only one speaker per atlas point.

For the FP data as a whole, 19% (251/1317) of the clauses examined have Ø-subjects and 81% have subject pronouns present. In contrast, in the contemporary Faetar data, 50% (1056/2092) of the clauses have Ø-subjects. The older Faetar data (Jaberg & Jud 1928) strongly resembles the FP atlas data (Gilliéron & Edmont 1902): 18% (16/88) of the sentences have Ø-subjects. The distribution of types of pronoun forms in the contemporary Faetar data is shown in Table 2.

Table 2: Distribution of types of subject pronouns in contemporary Faetar

Table 3 indicates the effects of the linguistic variables for the FP and Faetar data. Factor weights are from a Goldvarb 1-level binomial analysis. Each factor group is labeled as significant or non-significant, as determined by Goldvarb step-up/step-down binomial analysis. The factor groups which are significant in both the FP and the Faetar data are discussed below. Larger numbers indicate a higher probability weight/frequency of Ø-subjects for clauses with those variants.

FP (Input = 0.100) Faetar (Input = 0.230)

Factor Weight Freq. Weight Freq.

1: Person SIG. SIG.

1 sg. 0.859 0.18 0.037 0.04

1 pl. 0.638 0.10 0.029 0.07

2 sg. 0.216 0.01 0.200 0.17

2 pl. 0.525 0.03 no data

3 sg. masc. 0.385 0.34 0.563 0.36

3 sg. fem. 0.035 0.37 0.677 0.46

3 pl. masc. 0.234 0.44 0.494 0.22

3 pl. fem. no data 0.456 0.23

noun 0.998 0.88 0.716 0.30

demon. pronoun 0.983 0.73 0.896 0.72

none 0.231 0.09 0.309 0.25

start of MC 0.553 0.24 0.213 0.30

elsewhere in MC 0.506 0.27 0.211 0.20

start of EC 0.337 0.05 0.00

elsewhere in EC 0.581 0.97 0.619 0.11

relative clause 0.247 0.07 no data

affirmative 0.491 0.22 0.486 0.33

post-verbal negative 0.852 0.26 0.729 0.39

reflexive 0.668 0.09 0.843 0.57

none 0.442 0.22 0.467 0.31

"en"-type partitive 0.580 0.50 N/A

direct object 0.608 0.18 0.862 0.67

indirect object 0.820 0.29 1.000

new N/A 0.672 0.41

old 0.394 0.25

embedded clause 0.106 0.14

future 0.467 0.24 0.105 0.22

present 0.492 0.24 0.524 0.34

any other tense 0.513 0.19 0.246 0.25

For both the FP and the Faetar data, there is a strong effect of subject person, and this factor plays an important role in our constraint-based analysis. The other significant linguistic variables relate to the presence of other preverbal material, i.e., other arguments in the subject's surface position. Subject pronouns are less likely to surface when other material fills the same part of the linear surface order.

The presence of a [+Argument] subject (noun or demonstrative pronoun) has a significant effect in both data sets: doubling of overt NPs with subject pronouns is disfavored, while subject pronouns are favored where there is no overt NP. This is particularly true in FP where sentences with noun and demonstrative pronoun subjects have factor weights of 0.99 vs. 0.23 for sentences without a [+Argument] subject. Less strongly, the same pattern is found in Faetar: 0.89 (for demonstrative pronouns) and 0.72 (for nouns) vs. 0.31 (if no subject present). That is, some doubling is permitted in Faetar but is virtually categorically proscribed in Gallic FP.

In both FP and Faetar, the absence of object proclitics favors subject pronoun usage, and any non-subject proclitic form favors the use of Ø-pronouns. Again, this supports the hypothesis that subject pronouns surface less frequently when their argument position is otherwise occupied.

The non-significance of the negation factor in FP also supports the hypothesis that preverbal material tends to block subject pronoun usage (Heap 1997:208). Since all the negation considered here is post-verbal, we expect it to be "invisible" to the subject pronoun. We have no explanation for the slightly significant effect of this factor in contemporary Faetar.

In sum, there is a conspiratorial effect to force the verb toward the left-edge of its clause. A subject pronoun is less likely to surface when there is any other material intervening between the left edge of the clause and the head verb. This type of pattern is nicely handled in a constraint-based theory.

Rather than having an invariant Null Subject parameter or subject spell-out rule, we propose that the effects described above are the result of an interaction between two familiar constraints: ALIGN and PARSE (Prince & Smolensky 1993; McCarthy & Prince 1993a, b). These violable constraints are defined as follows:

This instantiation of ALIGN is a universal rule of right-branching languages such as the Romance languages. Any head is optimally positioned at the left edge of its constituent (or the right edge, in the case of left-branching languages).

We motivate this "explosion" of PARSE as follows. Accounts of the different behaviors of subject pronouns of different persons and reasons for these differences abound. For example, Dubois & Dubois-Charlier (1970:34-7) make a distinction between les pronoms personnels: or third person pronouns, and les noms personnels: or first and second person pronouns, proposing that the former, like nouns, have a determiner in the underlying form, but the latter do not. They find this distinction necessary to account for why only third person pronouns are marked for gender and as being somehow related to the fact that first and second person pronouns must have human (or at least animate) referents.

Mauger (1968:178) attributes the difference in behavior of pronouns of different person (in spoken French) to the fact that first and second person pronouns needn't refer to a noun that has already been expressed; while third person pronouns, derived from Latin demonstatives, usually represent a previously enounced noun or pronoun. Therefore, he posits, these pronouns which do not carry essential information, may be omitted more freely than first and second person pronouns in informal contexts (ibid 311).

Brandi & Cordin (1989:113) cite different behaviors for pronouns of different persons, in the Fiorentino and Trentino Italian dialects: Fiorentino requires a subject pronoun for all persons except first singular; Trentino requires a subject pronoun for all persons except first singular and plural and third plural. They attribute no theoretical significance to this pattern.

Hyams (1989:211) in contrast, notes that for English-speaking children in the early stages of language acquisition, omission of the subject pronoun is not restricted by person or other semantic features, but no quantitative evidence is provided. Harvie (1997) finds that English has many sentences with null subjects, but that subject type does not correlate significantly with null subject presence in English, although there are more null subjects for third person. Type of referent (same or switch) also correlated significantly.

This differential behavior of the pronouns reported for different languages (and observed in our data) provides support for a theory which treats pronouns of different persons in a discrete yet unified way, lending support for the three different PARSE constraints that we propose.

Of course, allowing for unconstrained free variation in the ranking of constraints is not an attractive feature. To constrain possible grammars, we assume that if more than one combination of rankings is permitted in a grammar, then all possible rankings of those constraints must be admitted as possible grammars and be motivated by data and constrained by theory (Anttila & Cho to appear, Nagy & Reynolds 1997). This restriction is implemented by systematically varying the rank order of the constraints in a subset of the constraint hierarchy. The system we implement includes Floating Constraints. One constraint, called the Floating Constraint (FC), may appear, with equal frequency, in any number of adjacent positions in a subsection of the hierarchy. This subsection is the domain of the FC. The position of the FC in its domain will determine the optimal candidate: different positions will favor different candidates. These different rankings account in a unified way for both synchronic and diachronic variation.

We may understand synchronic variation (different ways of producing an utterance at one point of time in a language's evolution) as the result of the FC appearing in different positions in its domain. In our account, ALIGN is the FC and its domain is the PARSE family's position in the constraint hierarchy. In FP, the members of the PARSE family are invariably ranked with respect to each other:

(8) PARSE family in FP: PARSE (II) >> PARSE (I) >> PARSE (III).

As will be discussed below, this accounts for all FP dialects having subject pronouns present most in second person and least in third, although different FP dialects have different frequencies for different grammatical persons. Faetar, the variety which has contact with Italian rather than French, has a different ranking of the PARSE constraints:

(9) PARSE family in FT: PARSE (I) >> PARSE (II) >> PARSE (III).

ALIGN, the FC, falls variably in four different positions in its domain (above, between, and below the three PARSE constraints) and will occur in each position in one-fourth of the utterances. The relevant part of the constraint hierarchy for FP is shown in (10) and for FT in (11).

(10) Constraint ranking forFP < ALIGN >

(11) Constraint ranking forFT < ALIGN >

The four possible rankings predict frequencies of each optimal candidate (with subject pronoun or with null subject) that differ according to subject person, as shown in the following tableaux. These tableaux indicate which of the two candidates incurs the highest ranking violation under each of the four possible rankings, by marking it with a star in the appropriate column. (12) illustrates the ranking for FP and (13) for FT.

This model accounts for each of the forms that surface and prohibits those forms that never surface. Furthermore, the number of occurrences of each surface form is correlated to the number of rankings that produce each form, as shown in Tables 4 and 5. The predicted percentages of null-subject pronouns for each grammatical person are determined by calculating the ratio of the number of columns in the tableaux above that do not indicate a violation for the null subject candidate to the total number of columns (= 4).

Grammatical Observed Predicted

Person (3 constraints) (7 constraints)

1st person 15% 50% 25%

2nd person 02% 25% 13%

3rd person 38% 75% 38%

Grammatical Observed Predicted

Person 3 constraints 7 constraints

1st person 13% 25% 13%

2nd person 18% 50% 25%

3rd person 33% 75% 38%

Two models are shown under "Predicted." The one labeled "3 constraints" illustrates the predicted values given just the three motivated PARSE constraints in the domain of the FC as illustrated in (12) and (13). While these predictions rank the persons in the appropriate order and predict all and only the observed forms, the predicted frequencies are not particularly close to the observed frequencies. If there were other constraints in the domain of the FC, the actual values could match better than with the three constraints described above; the match is very close with seven constraints in the domain of the FC. We have yet to determine what those other constraints might be, but it seems plausible to hypothesize that they will turn out to involve the other properties often associated with the NSP, such as that-trace violations and subject inversion. In this latter case, for example, there is likely a constraint which selects between the inversion and the deletion of a subject in order to avoid violation of ALIGN: inversion turns our to be (categorically?) prefered to deletion in certain contexts, such as questions. Clearly, this is an area requiring further research. Another possibility is that there are more than three PARSE constraints: either six, one for each person; or five, the three we’ve listed and PARSE(singular) and PARSE(plural). However, neither further explosion seems particularly illuminating based on current data.

A further benefit of the OT approach with FC’s is their ability to account for interspeaker variation, both synchronic and diachronic. Diachronic variation, the way a language systematically changes over time, can be accounted for by a gradual repositioning of a FC. The FC moves from being permanently anchored at one end of a domain to being permanently anchored at the other, by passing through a stage in which it is a FC (Nagy 1996:226-30).

To illustrate, we compared data from Faetar speech recorded in the 1990's to Faetar utterances transcribed in the 1930's. The contemporary speech shows a much higher frequency of null subjects. This may, however, be an artifact of the context of elicitation. The 1990's data consists of recorded narratives and conversations addressed to a person familiar to the community; the 1930's data comprises isolated elicited translations of Italian utterances spoken to an outsider.

Additionally, the higher rate of null subjects in the newer data could be attributed to the increased degree of contact that Faetani now have with Italian speakers. That is, their speech may reflect some degree of code-switching between Faetar and Italian, and there would be more of it as time passes and more Faetani learn and use Italian in more contexts.

However, setting aside these possibilities for a moment, we might interpret our data as contradicting the proposal that [+Null Subject] is the unmarked setting for a language, and that, over time the parameter may be reset to [-Null Subject] (Hyams 1989:227). Our data suggests a change in exactly the opposite direction. The FC-enhanced OT theory allows for motion in both directions and does not need to stipulate one grammar as more (or less) marked than any other. This may be seen as either an advantage or a disadvantage of OT.

In order to overcome the problems caused by the differences between the two sets of Faetar data, we compare the speech of younger and older Faetar speakers in the 1990's data. We assume that this snapshot in time provides a true reflection of the diachronic progress of the language (following Bailey et al. 1991). As shown in Figure 2, younger speakers show a much lower rate of null-subject usage. This supports Hyams' proposal about the possible direction of change for this part of the grammar (though we are not proposing that a person’s grammar is still being determined throughout adolescence and adulthood). The differences across these eight subsets of speakers is significant, according to a Goldvarb 2.0 one-level regression analysis.

As another means of overcoming the inequality between the two data sets, a small corpus of contemporary Faetar elicitation/translation data was compared to the atlas data. This data set shows virtually categorical usage of subject pronouns: only 2 of 271 sentences have a Ø-subject, indicating a change toward a [-Null Subject] language, again supporting Hyams' claim.

The model presented her differs from previous proposals in several important ways:

These changes allow us to account for diachronic and synchronic variation in a unified way, and to describe a grammar using constraints only, rather than a combination of parameters, rules, and constraints. We summarize the motivation for these innovations in this section.

In the most straightforward instantiation of OT with FC's, which is the model we are working with here, the FC will occur in each position in its domain with equal frequency. Hence the name "Floating Constraint"-- the constraint is not anchored and so it floats about, falling randomly, and with equal frequency, in each possible space in its domain. This approach avoids the need for weighting constraints or other stipulative restrictions on the grammar. Because good estimates of actual linguistic production have been achieved in this manner for a Faetar phonological variable's synchronic variation (Nagy & Reynolds 1997) and diachronic variation (Nagy 1996), as well as for the data considered here, it seems reasonable to conclude that an accurate model of grammar can be achieved without stipulated probabilistic weights.

Invoking a family of PARSE constraints, as opposed to a single PARSE constraint, highlights the fact that while different FP dialects have different frequencies of subject pronouns for different pronominal persons, an invariable implicational hierarchy is in place within each language variety. All dialects of FP have subject pronouns present least in third person and most in second person. The one exception is Faetar, the Italian FP dialect, which has most subject pronouns present in first person rather than second. This different ordering of the PARSE family may be due to Italian influence, but quantitative analysis of the Apulian Italian dialect which which Faetar is in most direct contact is necessary before such a conclusion can be drawn.

Finally, the "bundling" of properties originally associated with the Null Subject Parameter (that-trace violations, subject inversion, long WH movement, etc. Chomsky 1981:240) appears to be more of a robust generalisation than an absolute universal (Safir 1985, Wanner 1993). In a framework which invokes categorically binary parameters, such evidence would invalidate the original hypothesis; in our model, such "soft universals" are not only tolerated but in fact predicted as part of the variation inherent to constraint interaction. Thus, the strongest motivation for a theory of Floating Constraint, which entails a constraint-based model, is its ability to account for many types of variation found in the data.