2.1 Merchant’s data of nominal ellipsis

Merchant (2014) claims that there are three groups of human-denoting MASCULINE-FEMININE noun pairs in Greek, according to whether nominal ellipsis with GENDER mismatch is possible. More concretely, based on data from predicative uses of relevant nouns, he puts forward the following classification:

• For Class I nouns like adherfos ‘brother’ vs. adherfi ‘sister’, nominal ellipsis with mismatching GENDER is unacceptable with a MASCULINE or FEMININE antecedent;

• EPICENE nouns like jatros ‘doctor’, which constitute Class II, allow for nominal ellipsis with GENDER mismatches with a MASCULINE or FEMININE antecedent; EPICENE

• For Class III nouns like dhaskalos vs. dhaskala ‘teacher’, nominal ellipsis with GENDER mismatch is grammatical when the antecedent is MASCULINE but not when the antecedent is FEMININE.

Relevant examples are reproduced below from Merchant (2014) (his (9), (22) and (25), respectively). In these examples, the intended gender of the elided noun is visible on the determiner (D) and adjective (A). The supposed elided phrases are indicated throughout the paper as ELLIPSIS (although as we will see in Section 4, Merchant claims that these cases involve a proform, rather than deletion). The judgments are as reported in Merchant (2014).⁴

Noting potential inter-speaker variation (see also Bobaljik & Zocca 2011 for a related remark about Brazilian Portuguese), Merchant (2014) lists the following nouns as examples of the three classes.⁵

We would like to point out, however, that the judgments of these sentences do not seem to be as clean as reported in Merchant (2014). Upon finding some variability among speakers, we closely consulted 8 native speakers of Greek, and also conducted an informal questionnaire with 16 native speaker linguists, but failed to reliably replicate the above patterns. In particular, while the judgments about Class I and Class II nouns seem to be clear and replicable to some extent, the asymmetry between the MASCULINE and FEMININE antecedents with Class III nouns was not constantly found, and for many of our consultants, even when there is a contrast, it does not seem to be very prominent. Furthermore, as noted above, there seems to be inter-speaker variation as to which nouns belong to Class I or Class III, and we found no pair of MASCULINE-FEMININE nouns that showed a reliable asymmetric pattern that persisted across speakers.

Part of the reason why we failed to replicate these results might be due to the fact that none of the sentences above are perfectly natural, which is largely due to the presence of an indefinite article only in the second conjunct, a factor Merchant himself acknowledges as a potential confound. He explains that the indefinite article is included in these and other examples in his paper, despite the somewhat degraded status, because without it the second sentence would preferentially receive a predicative adjectival reading (which is akin to ‘Mary is bad’, rather than ‘Mary is a bad one’) and would not involve nominal ellipsis. Also, adding an indefinite article in the first sentence would result in an identity reading, rather than the intended predicational reading, and is also somewhat unnatural.

We therefore suspect that the unnaturalness of the examples contributed to our failing to replicate the judgments he reports, but at the same time, we have reasons to think that that does not explain everything. Specifically, we will see in Section 3 that nominal ellipsis is possible in argument position as well, contrary to Merchant’s claim, where we do not have analogous concerns about unnaturalness. Nonetheless, we failed to find a pair of gendered nouns that exhibited a reliable contrast between MASCULINE and FEMININE antecedents even with such sentences.

For these reasons, we think Merchant’s (2014) data involving nominal ellipsis do not constitute conclusive evidence for his classification of noun-pairs with interpreted GENDER in Greek. This of course does not mean that his classification needs to be rejected, as there are several possible reasons for our failure to replicate his results. Also, Bobaljik & Zocca (2011) present similar patterns in other languages, which suggests that what Merchant observed had some truth to it. While we are unable to pin down the exact reasons why the data turned out to be not as clear as we expected, we will raise further data involving focus constructions that support essentially the same classification of noun pairs with interpreted GENDER in Greek and that received relatively stable judgments among the 16 speakers we consulted.

2.2 Focus constructions

To show that there are indeed three classes of nouns with GENDER in Greek, we will make use of focus constructions. Focus constructions involve reference to alternatives, and depending on the noun, the alternatives might or might not have to refer to the gender inference. Consider first the following examples involving Class I nouns, adherfos and adherfi, and observe that they lack inferences about the opposite gender.

These examples can be understood as cases where the focus alternatives must involve the gender inference. That is, mono ‘only’ negates the focus alternatives in these sentences, and the focus alternative all look like ‘Maria is Janis’s male sibling’ and ‘Petros is Janis’s female sibling’, respectively. Consequently, if the gender of the subject of an alternative sentence does not match the gender of the noun, then the falsity of the alternative sentence is trivial.

On the other hand, Class II nouns do give rise to inferences about the opposite gender in the same construction.

Here the focus alternatives are oblivious to the gender information, and consequently alternatives like ‘Maria/Petros is a good doctor’ are negated, regardless of the GENDER marking in the sentence.

Crucially, we observe an asymmetry with Class III nouns like dhaskalos-dhaskala where the MASCULINE gives rise to an inference about female individuals, while the FEMININE does not give rise to an inference about male individuals.

Other focus constructions point to the same conclusions, including superlatives, ordinals and nominal only. In order to save space, we will only present data involving superlatives here. In the superlative construction of the form the best N, Class I nouns give rise to readings that only compare individuals with the same gender.⁶

On the other hand, Class II nouns give rise to readings where the comparison is across individuals of both genders.

With Class III nouns, we observe a crucial asymmetry such that the MASCULINE nouns give rise to a reading with gender-neutral comparison like Class II nouns, and the FEMININE noun gives rise to a reading with gender-separated comparison like Class I nouns.

These are consistent with the patterns we observed with mono ‘only’ above.

Using these tests, we can identify essentially the same three classes of nouns with interpreted GENDER in Greek that Merchant (2014) proposes. It seems to us that our diagnostics with focus constructions are more reliable and yield results that are more consistent across speakers, although, as we remarked above, the reason why the ellipsis data are not so clear remains a mystery for now. Furthermore, these data involving focus constructions not only show the same empirical point as Merchant (2014) but also pose a challenge for his theory, as we will discuss in Section 4.

4.1 Assessment

To summarize, the following are crucial ingredients of Merchant’s (2014) proposal.

1. There are two strategies for nominal ellipsis, (i) PF-deletion and (ii) a pro-form e_N.

2. PF-deletion requires total semantic identity, including the presupposition. This bans all cases of deletion involving GENDER mismatch.

3. The proform e_N needs to be licensed by a local D. When e_N is in argument position, the φ-features of D will not be valued, which results in ungrammaticality.

4. Some nouns have gender presuppositions in their lexical semantics (i.e. Class I nouns and Class III FEMININE nouns), while others (Class II nouns and Class III MASCULINE nouns) don’t.

Given our observation from the previous section that nominal ellipsis with GENDER mismatch is actually possible in predicative and argument positions, Merchant’s (2014) analysis cannot be retained in its full form. However, there is a very easy fix that makes it compatible with nominal ellipsis with GENDER mismatches in argument position, namely, simply by dropping 3. above. That is, it could be assumed that φ-features on determiners and adjectives need not be syntactically licensed and are simply interpreted (or licensed by a φ-head above the DP; Sauerland 2003; see Section 5.2.2 below for a concrete implementation). With this fix, nominal ellipsis with GENDER mismatch becomes possible in both predicative and argument positions under his theory.

However, there are remaining issues. Firstly, there is now no reason to postulate two separate mechanisms for nominal ellipsis in Greek. Rather, we will demonstrate below that all the relevant data can be accounted for just with PF-deletion (we will also discuss and reject the alternative possibility of only using e_N in Section 6.3). Thus, we will drop Merchant’s (2014) assumptions 1. and 3. above.

Notice that in order for this alternative theory to work, we will also have to drop 2. above, according to which deletion requires complete semantic identity, because that is what prevents deletion from applying for ellipsis with GENDER mismatch. As we will see immediately below, there is actually independent reason to think that ellipsis (and focus constructions) systematically ignore presuppositions triggered by φ-features (‘φ-presuppositions’) including GENDER features, and ellipsis involving such mismatches in φ-features are generally licit (Ross 1967; Fiengo & May 1994; Spathas 2010; Johnson 2014). For instance, the following examples involving ellipsis with GENDER mismatches are all grammatical in English.

In other words, Merchant’s assumption 2. can be shown to be problematic on independent grounds.

This leaves us with 4. about the distinction between nouns with and without gender presuppositions. While we adopt this idea to make sense of the three classes of noun pairs in Greek, we also claim that it needs to be revised to account for the data involving focus constructions discussed in Section 2.2. In particular, recall that for Merchant, gender inferences are all presuppositional. However, it is widely observed that focus alternatives are generally oblivious to φ-presuppositions, including gender presuppositions, just like structures under ellipsis (see Spathas 2010; Jacobson 2012; Sauerland 2013 for relevant discussion).¹⁴ For instance, consider the following examples, under the bound readings of the possessive pronouns.

Suppose that the relevant students are the speaker, John and Mary. Then, (42a) entails that Mary and John didn’t do their homework, (42b) that the speaker and Mary didn’t do their homework, and (42c) that the speaker and John didn’t do their homework. What is of importance here is that the φ-features (person and GENDER features here) of the bound possessive pronoun seem to have no semantic effects in the focus alternatives. For instance, what is negated in (42c) looks like the following, and the third person and FEMININE features do not figure here.

On the other hand, assertions are never ignored. For instance, (44) does not entail that John is not an athlete, but only that John is not a female athlete.

If all gender inferences of Greek nouns with gender inferences are presuppositional, the data like (9) and (11), repeated below, will remain puzzling.

That is, the gender inferences in the focus alternatives here should be ignored, and consequently, (9a), (9b) and (11b) are expected to have entailments about the opposite gender as well, just like (11a) (and also (10) with Class II nouns).

In order to account for this, we will claim that Class I nouns like aderfos and aderfi and Class III FEMININE nouns like dhaskala have gender inferences in both assertion and presupposition, while Class II nouns like jatros and Class III MASCULINE nouns have no gender inferences in either dimensions of meaning. Furthermore, when showing FEMININE agreement, jatros has a gender inference only as a presupposition, and when showing MASCULINE agreement, it receives a gender inference in competition with the FEMININE version of the noun. We will raise evidence for this analysis and give a more detailed explanation about how this accounts for the above data in the next section.

4.2 Merchant’s argument from extraction

Before leaving this section, we should mention Merchant’s (2014) data involving extraction from the ellipsis site, which he raises as support for his postulation of the proform e_N. He observes that extraction from the ellipsis site is with matching GENDER, but not with mismatching GENDER.¹⁵ Here, what is extracted is the genitive phrase, and dhikigoro ‘lawyer’ is an EPICENE noun.

Merchant argues that this observation speaks in favor of his analysis for the following reason. Recall that according to him, GENDER-matching ellipsis can be derived via PF deletion, while GENDER-mismatching ellipsis always involves a null pronominal e_N. It is then expected that extraction in the latter case should be impossible, given that extraction from a pronominal should be generally banned (cf. Fiengo & May 1994; Schwarz 2000; Johnson 2001; Saab to appear). The above contrast in fact accords well with this prediction.

While the observation seems to be solid, we would like to disagree with the conclusions Merchant draws. Firstly, the unacceptability of (45b) is actually not as grave as one might expect for a violation of the ban on extraction from a pronominal element. In fact, the contrast appears to be comparable to the mere degradation associated with ellipses with GENDER mismatches across the board. Secondly, we observe that the baseline cases without ellipsis already exhibit the same kind of contrast (at least for some speakers):¹⁶

Although we need to leave an analysis of these data for another occasion, they indicate that whatever is responsible for the contrast is not the mechanism of ellipsis, contrary to what Merchant (2014) assumes. Rather, ellipses with matching GENDER and non-matching GENDER seem to show mild contrasts across the board (cf. fn.4). The nature of these contrasts are not very clear to us at this moment, but that they are not very sharp suggests that they are likely to be caused by non-syntactic factors, and given that they are observed with the overt versions of the sentences, they have little, if any, to do with ellipsis per se.¹⁷

5.1 Unmarked GENDER and gender competition

For Merchant’s (2014), [MASCULINE] and [FEMININE] are symmetric at least in the semantics in the sense that both features give rise to gender presuppositions. However, since Roman Jakobson’s seminal work (Jakobson 1984), it is well recognized that GENDERs in natural languages are often not equal. And as far as human-denoting nouns in Greek are concerned, [MASCULINE] is less marked than [FEMININE] in the semantic sense, i.e. [MASCULINE] can often be used in a gender-neutral manner as an ‘elsewhere GENDER’, so to speak (see Spathas 2010 for Greek and Corbett 1991; Bobaljik & Zocca 2011; Heim 2008; Kramer 2015; Percus 2006; 2011; Sauerland 2008b for other languages).²⁰ This is most clearly observed with items that have gender inferences in the domain of presupposition, e.g. pronouns.

There are independent reasons to believe that interpreted GENDERs (and possibly other interpretable φ-features) on pronouns are presupposition triggers (Cooper 1983; Heim & Kratzer 1998; Kratzer 1998; 2009; Heim 2008; Jacobson 2012; Sudo 2012). In English, for example, her is used to refer to a female individual, and if it is used to refer to a male individual, it gives rise to presupposition failure. More specifically, assuming that pronouns denote variables, Heim & Kratzer (1998) propose the following treatment, based on Cooper (1983) (see Spathas 2010 for Greek pronouns).²¹

One might be tempted to give a similar analysis to MASCULINE pronouns, but unlike FEMININE pronouns, MASCULINE pronouns in Greek can be used as gender-neutral pronouns.²² This asymmetry between MASCULINE and FEMININE can be observed, for example, when the pronoun is bound by a quantifier with individuals of both genders in the domain of quantification. In such contexts, a MASCULINE pronoun is felicitous but not a FEMININE pronoun, as shown in (51). Here, the intended reading is one where the pronoun tu/tis is bound by the disjunctive subject.

As we will explain more precisely below, the unacceptability of (51b) shows that FEMININE pronouns are exclusively used for FEMININE referents (see the references cited above for similar facts in other languages). We will present below further evidence of the gender-neutrality of MASCULINE in Greek with data involving gender features on determiners and adjectives (namely, (64), (65), (68) and (69); see also Spathas 2010).

In sentences like (51), MASCULINE pronouns behave like gender-neutral pronouns, but in other contexts, they are not completely void of gender inferences. For instance, a free MASCULINE pronoun is typically used to refer to a male individual, and it is simply infelicitous to use it to refer to, say, Hilary Clinton, while knowing her gender. Concretely, (52) cannot be used to mean ‘I voted for Hilary Clinton’.

Also, the following example is unacceptable under the bound reading of tu.

To reconcile these facts with gender-neutral uses, we follow Percus (2006; 2010; 2011) and Sauerland (2008a; b) and assume that [MASCULINE] is actually semantically completely gender-neutral in languages like Greek, but systematically excluded when [FEMININE] could be used felicitously instead. In order to make this idea more concrete, we postulate a principle forcing the use of the more specific form of the MASCULINE and FEMININE pair, whenever possible (cf. the literature on anti-presupposition/implicated presupposition, e.g. Heim 2008; Percus 2006; 2011; Sauerland 2003; 2008a; b).²³

The Principle of Gender Competition accounts for the data in (51) as follows. By assumption, the FEMININE pronoun tis presupposes that the referent is female. In (51b), this presupposition makes one of the disjunctive possibilities unable to be true. Generally, a disjunction is only felicitous if all the disjuncts are possibly true (Gazdar 1979). Since this is not the case for (51b), the example is unacceptable. On the other hand, the MASCULINE pronoun in (51a) has no presupposition by assumption. The Principle of Gender Competition says that (51a) is only felicitous in contexts where (51b) is not felicitous. Since (51b) is infelicitous everywhere, (51a) can be used in any context.

In cases where the FEMININE counterpart could be used felicitously, on the other hand, a MASCULINE pronoun ceases to be gender-neutral, as predicted by the Principle of Gender Competition. Suppose that we know that Hilary Clinton is female. Then, in order to refer to her, the use of a FEMININE pronoun is forced by the Principle of Gender Competition, and the use of the MASCULINE pronoun is consequently banned, although it is semantically coherent. The example in (53) is explained similarly: In this sentence, the FEMININE pronoun tis could be felicitously used instead, which blocks the use of the MASCULINE pronoun tu.

5.2 Technical remarks

At this point, we would like to digress a bit and remark on some technical details.

5.3 The semantics of GENDERed nouns

Recall our implementation of Merchant’s (2014) idea about nouns with and without gender inferences, repeated here.

We argued above that [MASCULINE] is generally semantically gender-neutral in Greek, but note that we are allowing Class I MASCULINE nouns to have a gender inference. Consequently, there are two types of MASCULINE nouns in our analysis: semantically MASCULINE ones like adherfos and semantically neutral ones like dhaskalos. Although their semantics is distinct, both types of MASCULINE nouns trigger MASCULINE agreement. Similarly, there are two types of FEMININE nouns, namely, ones with a gender inference like adherfi and dhaskali and ones that do not have gender inferences but that occur with determiners and/or adjectives with gender presuppositions.

We will now present two sets of evidence motivating this distinction between nouns with and without gender inferences in their semantics, namely, plural nouns (Section 5.3.1) and negative existential sentences (Section 5.3.2).

5.4 GENDERed nouns in focus constructions

Strictly speaking, the above two phenomena only motivate the distinction between nouns with and without gender inferences in their semantics, and do not constitute evidence for our claim that gender inferences in the nouns themselves manifest themselves in the assertive component of meaning. This part of the analysis is actually crucial for accounting for the behavior of nouns in focus constructions, which we saw in Section 2.2.

As discussed in Section 4.1, it is independently observed that φ-presuppositions are generally ignored in focus alternatives, whereas asserted meaning cannot be. What our analysis entails, then, is that nouns with gender inferences in their semantics, having asserted gender inferences, should require focus alternatives to also have gender inferences as well, while nouns without gender inferences in their semantics do not require focus alternatives to have the same gender inference. This is exactly what we observed in Section 2.2.

Let us review the crucial data. According to our analysis, Class I nouns all involve gender inferences and indeed, there is no entailment about the opposite gender in the following examples.

This is because the gender inferences in the alternatives make the inference trivial when individuals of the opposite gender are subjects in focus alternatives, e.g. that Maria is not Janis’s male sibling is trivially true, and so is that Petros is not Janis’s female sibling.

On the other hand, Class II nouns do not have gender inferences in their semantics, and as expected, we do observe entailments about the opposite gender in the following examples.

In these cases the gender inferences are ignored in the alternatives, so the sentences have these entailments.

Finally, for Class III nouns, we observe an asymmetry, as expected.

Data involving other focus constructions can be explained in similar ways.

6.1 Internal syntax of nouns with GENDER

We took our observation to motivate the particular semantic variation we proposed among nouns with interpreted GENDER in Greek but we think it shows us very little about their internal syntax. Throughout the paper we have been speaking as if relevant nouns are listed in the lexicon as they are, but this is not to exclude the possibility that these nouns are decomposable into smaller components. For instance, it is perfectly compatible with our proposal to assume that adherfos decomposes into the root adherf- and some functional node (e.g. n) that hosts the GENDER feature. We remain uncommitted to the internal syntax of the nouns under discussion, especially the exact location of the GENDER feature, as this is currently an actively debated topic (see Wechsler & Zlatić 2003; Sauerland 2008b; Matushansky 2013; Pesetsky 2013; Kramer 2014; 2015; Puškar 2015; 2017; 2018; Fathi & Lowenstamm 2016 and references therein). Luckily, none of the arguments we make in this paper hinges on a particular theory of it. We discuss related points in another paper, Spathas & Sudo (to appear).

6.2 Nouns under ellipsis

Recall now that we mentioned in Section 4.1 that ellipsis also ignores φ-presuppositions, similarly to focus constructions, in the sense that mismatches in φ-presuppositions are tolerated. Furthermore, it can be observed that assertive meaning is not ignored in ellipsis, as in (69).

Generally, identity up to φ-presuppositions is required across various elliptical phenomena.³¹ Our denotations of nouns with interpreted GENDER make the predictions that partially match Merchant’s (2014) data of nominal ellipsis.

That is, Class I nouns like adherfos and adherfi have gender inferences in the assertive dimension of their sematnics, so GENDER mismatches should not be tolerated. On the other hand, Class II nouns have no asserted gender inferences, so GENDER mismatches should be tolerated. Recall that the judgments of these cases are relatively clear and we take it that our analysis make good predictions here. The relevant data are repeated from Section 3.

On the other hand, our analysis does not straightforwardly predict the asymmetric licensing with Class III nouns dhaskalos-dhakala that Merchant (2014) reports. While we failed to replicate such a pattern, it would still be instructive to discuss such data for two reasons. Firstly, our failure to replicate Merchant’s data might be due to some independent factors that we failed to control for. In fact, we have no evidence that directly contradicts the data he reports. Secondly, Bobaljik & Zocca (2011) report similar data in languages like Brazilian Portuguese and we certainly expect our analysis to carry over to these languages (see Kučerová 2018; Murphy et al. 2018; Sprouse et al. 2018 for more data).

The reason why our analysis does not straightforwardly predict asymmetric licensing is because ellipsis involving asymmetric entailment is generally unacceptable, as shown by (70).

According to our analysis the Class III FEMININE noun dhaskala asymmetrically entails the Class III MASCULINE noun dhaskalos both in the assertion and in the presupposition. Then, it is expected that GENDER-mismatching ellipsis with Class III nouns should be generally unacceptable, regardless of which one is the antecedent.

We would like to suggest one possible fix of our analysis to address this potential issue. The idea is to assume that the Principle of Gender Competition does not apply to elided nouns.³² Then, we can analyze the data as in (71).

The ellipses here involve total identity, and the DP with an elided noun has DP-internal GENDER mismatch. What’s crucial here is the assumption that such mismatch is tolerated, because the Principle of Gender Competition is not active.

Unsurprisingly, the overt counterparts of these sentences are both utterly unacceptable, as in (72).

In fact, such DP-internal GENDER mismatches with overt material are generally banned in Greek. Our analysis explains the unacceptability of the sentences in (72) as follows. Although (72a) is semantically perfectly coherent, due to the gender-neutral denotation of dhaskalos, it is ruled out by the Principle of Gender Competition. The other example, (72b), is ruled out by the same principle, because the FEMININE determiner mia is available here. But, crucially, notice that there is simply no way that (72b) could receive the intended reading, because dhaskala is false of a male individual, due to its semantic entailment.

Coming back to the elliptical cases, the infelicity of (71b), then, is not at all surprising. The explanation is the same as for (72b). This just cannot mean what it should mean. What is surprising is the felicity of (71a). The minimal difference between (71a) and (72a) is whether the noun is overt, and the reason why (72a) is infelicitous is because of the Principle of Gender Competition. Conversely, if the Principle of Gender Competition is made inactive under ellipsis, the sentence should be acceptable. With this auxiliary assumption, therefore, we could explain the asymmetric licensing that Merchant (2014) reports.

Furthermore, it is crucial here that what is elided in (71a) is a MASCULINE noun dhaskalos, which is assumed to have no lexically encoded gender inference. This makes a prediction that when an elided noun with a FEMININE determiner occurs in a focus construction, the interpretation should not be restricted to female individuals. Consider, for example, the following examples:

The crucial point about (73b) is that it entails that the other relevant people have no teacher, male or female, in Katerini, and, therefore, should be judged false if it turns out that Petros has a male teacher, for example. Furthermore, the following sentence with an overt FEMININE noun is not judged false in such a scenario.

A contrast of the same nature should arise with other focus constructions too, e.g. superlatives:

(75a), unlike (75b), should entail that Maria’s female teacher is better than anybody else’s teacher, including male teachers. Thus, if Petros has a male teacher and if he turns out to be better than Maria’s, (75a) is judged false, while (75b) stays true.

These data would constitute strong support for the analysis that what is elided in (71a) is a MASCULINE noun with gender-neutral meaning, even when other items like determiners are overtly marked as FEMININE, and an important consequence of this would be that underlyingly such GENDER mismatching nouns phrases are allowed, if not overtly, in Greek. It should be remarked in this connection that such GENDER mismatching nouns phrases are overtly attested in other languages like Russian (Pesetsky 2013) and Croatian/Serbian/Bosnian (Wechsler & Zlatić 2003; Puškar 2017; 2018). Importantly, for the purposes of this paper, we do not have to make commitments about syntactic assumptions as to how such mismatches arise in the syntax, and we simply refer the reader to the works cited here.

However, since the judgments of GENDER-mismatching ellipsis with Class III nouns are generally not stable with our informants, we could not test these predictions. If the judgments of examples like (71) are stable in languages like Brazilian Portuguese, the above prediction should be testable there. But we have to leave this open for now.

6.3 On the proform eN and some remarks on uninterpreted GENDER

Notice at this point that we assumed above without argument that the mechanism for nominal ellipsis in Greek is PF-deletion. Recall that Merchant (2014) postulated another mechanism, namely, the proform e_N. We could actually account for the main data with e_N instead, for the following reasons.³³ Recall that we assumed total identity under ellipsis, in particular total semantic identity. But e_N would also force total semantic identity, because by assumption it has the same denotation as the antecedent noun.

That said, there are several reasons to favor the PF deletion account. Firstly, as we saw in 4.2, extraction from nominal ellipsis site is generally possible, which suggests that the ellipsis site has an internal syntactic structure, and so PF-deletion is at least possible.

Our second argument comes from grammatical gender. Greek has several neuter nouns that refer to humans, e.g. koritsi ‘girl’, melos ‘member’, pedhi ‘child’, agori ‘boy’ (cf. Spathas 2010; see also Wurmbrand 2016). As mentioned briefly in Section 1, we treat them as involving uninerpreted GENDER. As Merchant (2014) observes, ellipsis with GENDER mismatches is not possible at all with these nouns (for any speaker), as demonstrated by (76) (see also Merchant’s 2014 (71)).

The unacceptability of (76) does not immediately follow from our analysis above, because the structure of the sentence is essentially identical to cases like (71). Notice also that all the gender presuppositions should be satisfied in this sentence. However, it seems to us that there is an independent constraint that specifically targets uninterpreted GENDER and forces DP-internal concord even under ellipsis. This rules out (76), because the second conjunct here involves a NEUTER noun but the determiner and adjective in the same DP bear FEMININE.

Now, if e_N is available, this analysis will be no longer available. Since e_N must be compatible with any GENDER in principle, it should be able to appear with a determiner and/or adjective carrying a FEMININE feature, given that nominal ellipsis with GENDER mismatches is possible with EPICENE nouns and with dhaskalos-dhaskala when dhaskalos is the antecedent. Moreover, there’s no reason why it cannot refer back to the meaning of koritsi ‘girl’ in (76). Then, (76) is wrongly ruled in. In fact, Merchant (2014) leaves (76) as an open problem that is unaccounted for in his theory. On the other hand, our analysis with PF deletion has at least something to say about this data point, namely, syntax treats interpreted GENDER and uninterpreted GENDER differently, although morphology seems to conflate them. Analyses along these lines have in fact been put forward by some previous studies, such as Alexiadou (2004); Kramer (2014; 2015) and Kučerová (2018). We also propose a view based on the proposal in this paper in Spathas & Sudo to appear), and we refer the interested readers to these works.

6.4 Classification of GENDERed nouns

We have largely put aside the issue of which noun pairs with interpreted GENDER fall into which of the three classes. There are several reasons for this. For one, apart from the robust generalization that all EPICENE nouns lack gender inferences in their semantics, there does not seem to be other stable morphological or syntactic cues as to which nouns have gender inferences and which nouns don’t in Greek (see, however, Bobaljik & Zocca 2011 for the potential relevance of morphology in Brazilian Portuguese and other languages). For instance, adherfos and dhaskalos share the suffix -os, but their GENDERs do not have the same semantic status. In fact, which nouns belong to Class I vs. Class III seems to be a point of inter-speaker variation (see Bobaljik & Zocca 2011 and Merchant 2014: fn.6 for related remarks).

Furthermore, there does not seem to be a solid semantic predictor, either. As Bobaljik & Zocca (2011) point out for other languages, there seems to be a tendency for kinship terms to belong to Class I, but there are some exceptions to this in Greek (at least for some speakers), e.g. thios-thia ‘uncle, aunt’ seems to be in Class III. Thus, the semantics, although arguably relevant, is not the only factor, either (pace Bobaljik & Zocca 2011). At the moment, we have nothing insightful to say about this issue.

As mentioned repeatedly, there are other languages that seem to have the same three classes, e.g. Brazilian Portuguese (Bobaljik & Zocca 2011). It is of particular interest here that the Brazilian Portuguese counterparts of Greek EPICENE nouns have separate MASCULINE and FEMININE forms, e.g. médico ‘male doctor’ and médica ‘female doctor’. What this entails is that there can be MASCULINE-FEMININE noun pairs where both nouns lack lexically specified GENDERs (at least in the assertion) in some languages, although no such nouns are found in Greek. Bobaljik & Zocca’s data also suggest that class assignment is crosslinguistically variable and is unpredictable from the semantics alone. For instance, tio-tia ‘uncle, aunt’ in Brazilian Portuguese seems to belong to Class I, but the corresponding nouns in Greek belongs to Class III (although the data are potentially muddled by the possibility of inter-speaker variation, as noted above). These issues lead also to important questions concerning the acquisition of nouns and their class assignment, but questions like these are far beyond the purview of the present paper (see Sprouse et al. 2018).

Relatedly, there seems to be a gap in the paradigm, at least, in Greek in that there is no MASCULINE-FEMININE pair that has a gender inference only on the MASCULINE noun (i.e. the opposite of dhaskalos-dhaskala). This is likely to be due to the general unmarkedness of MASCULINE relative to FEMININE in Greek, but nothing in our analysis excludes such a pair. Although we will not try to capture the absence of such pairs of nouns in Greek, we would like to note that there seem to be languages where FEMININE is less marked than MASCULINE (see Corbett 1991: Ch.7, Kramer 2015: Ch.5; see also Bosnian/Serbian/Croatian data reported in Puškar 2017; 2018; Murphy et al. 2018), and there might well be pairs where the MASCULINE noun has a gender inference and the FEMININE noun does not (see Percus 2011 for related discussion). Whether this is so is left unanswered here as well.

6.5 Cases with no gender competition

We mentioned in fn. 25 the existence of gender-neutral nouns that have gender-specific counterparts but do not seem to compete with them. For instance, for many speakers of contemporary English, actor can be used gender-neutrally, although it has a morphologically related FEMININE form actress. Thus, for these speakers, we observe the following pattern of judgments.

If the Principle of Gender Competition applied to actor-actress, (77a) should be ruled out. Thus, this indicates that actor somehow does not compete with actress.

In Greek, one can find such nouns in the domain of animal-denoting nouns (although not among human-denoting nouns). For instance, jata is grammatically a FEMININE noun, but commonly used to describe both female and male cats. In addition to this noun, there is a gender-specific noun jatos, which is exclusively used for male cats. Concretely, from (78), one cannot infer the gender of the cat that Maria bought, but one can infer that Kostas bought a male cat.

Just as in the case of actor-actress, if jata and jatos competed, one should infer from (78a) that Maria bought a female cat. Thus, again, jata does not seem to compete with jatos.

The immediate question that arises is why these nouns are exempt from gender competition. We would like to suggest here that the gender-neutral member of these pairs has uninterpreted GENDER, and the Principle of Gender Competition does not apply to uninterpreted GENDER. Specifically, we analyze jata ‘cat.F’ as having uninterpreted GENDER. In fact, we observe that the jata-jatos pair does not allow nominal ellipsis with GENDER mismatches, just like in the case of koritsi ‘girl.N’ in (76).

We spell out more details of this idea in Spathas & Sudo (to appear).

Extending this idea to English, actor (for certain speakers) could be analyzed as having uninterpreted GENDER. There are, however, some more complications here. In particular, actor can bound a FEMININE pronoun, as in (80).

One might take this as showing that actor has an interpreted GENDER. However, as Spathas (2010) observes, GENDER mismatches in pronominal binding are possible, at least in some languages, if not universally. As shown in (81), to koritsi ‘the girl’, which is NEUTER, can bind a FEMININE pronoun in Greek (see Wurmbrand 2016).

Under the bound reading of the pronoun, this entails that other relevant people (e.g. boys) did not leave their houses. Thus, the example in (80) does not convincingly show that the GENDER on actor is not uninterpreted.

Thus, there is a puzzling contrast between koritsi ‘girl’ and jata ‘cat’. The latter does not allow for GENDER mismatching pronominal binding.

This observation suggests that there is something fundamentally different between uninterpreted GENDER on human-denoting nouns and animal-denoting nouns, at least in Greek. We address this and related issues in a separate paper (Spathas & Sudo to appear).