1.1 Preface

The inspiration for this paper comes from the observation that gender agreement can sometimes follow different criteria cross-linguistically, and more crucially, in different contexts within a language and between individuals. This is not a new observation; this paper elaborates on it by examining typological variation and contextual variation to propose a system for discussing different types of “gender” that are linguistically encoded and how they affect the form that agreement takes. The novel contribution is a framework for integrating general and linguistic cognition as related to gender, broadly construed. This framework will allow theoretical and experimental work in this area to more clearly identify and navigate issues relating to human gender as both a categorical and gradient phenomenon.

To begin, I lay out my proposed terminology for discussing gender in a principled way. This sets the stage for examining data from English, which does not overtly mark gender agreement outside of third person singular pronouns, and comparing it to observations from a variety of other languages that have richer gender inflection systems. I also examine how some lexical innovations which encode nonbinary gender fit into the wider picture of coreference.

Finally, these observations provide the foundation for a proposal which places languages (or, potentially individual speakers of those languages) along a gradient of permissiveness in gender agreement and relates this to how different types of gender, including nonbinary identities, are conceptualized and learned. The intention of this structure is to organize formal, empirical, and philosophical evidence to support the claim that gender is represented and accessed at different levels and to different degrees during the process of coreference resolution.

1.2 Gender as a complex phenomenon

The term gender is fraught in part because definitions given in the linguistics literature can vary dramatically across subfields or even specific works and are sometimes left as tacit assumptions, even within contexts like coreference resolution. This paper aims to clarify what kinds of gender might be relevant for real-time processing of syntactic agreement and coreference between a pronoun and a referring expression, noting proper names and genders outside of the ‘masculine-feminine’ (or ‘male-female’) binary. It develops the hypothesis that the type of gender involved in coreference checking in English, and possibly other languages, is primarily a domain-general categorical representation of the referent which a formal syntactic or semantic feature can draw upon during agreement and checking operations. That is, the mechanism for categorization of gender which is used to check gender congruency between a pronoun and the expression with which it corefers relies fundamentally on a general cognitive mechanism for classifying and checking congruency of gender rather than relying on a mechanism specific to linguistic processing, in line with the ‘mental model’ framework (e.g., Garnham & Oakhill 1990; Garrod & Terras 2000). Finally, I suggest some lines of research into individual variation that would be able to inform the questions brought up herein.

I will explicitly and precisely define several types of gender in order to provide consistent and unambiguous terminology to the study of coreference and pronouns. These definitions of “gender” include grammatical gender, conceptual gender, gender identity, gender expression, and biosocial gender. I iteratively develop a criterion for checking gender congruency (whether or not two lexical items ‘match’ in gender), then suggest a gradient way in which languages might employ the final formulation of the criterion to result in the typological variation observed. I also describe a three-tiered schema for formalizing the process of gender checking during coreference resolution. While English is the primary focus of this paper, I will demonstrate that motivation for these three categories can be found cross-linguistically. I draw on biological, social, cognitive, and grammatical evidence for how gender is conceptualized and used in human interaction in order to argue that coreference resolution (in English) relies primarily on a non-syntactic property, conceptual gender, for determining whether or not a pronoun and coreferring expression match or mismatch, which is domain-general in origin.

The relative difference in acceptability between sentences (1-a) and (1-b) (indicated by a #) illustrates that English coreference is influenced by discourse-level information and world knowledge. In order to develop a felicitous context for (1-a), one almost must assume the speaker is communicating their disapproval of the referent through misgendering. That is, although the referent’s gender remains ambiguous without further context, a salient interpretation would be that the speaker is intentionally discussing the referent using gendered words (either pronouns or definitional nouns) that are incongruent with the gender identity (defined in Section 2) and wishes of the referent.¹ In contrast, (1-b) provides a context that immediately allows for a felicitous and not necessarily transphobic interpretation since the gender of the costume-wearer and the gender of the pronoun may ‘mismatch’ without qualifying as misgendering the referent, as the costume is intended to mask their gender identity.

The difference in apparent acceptability between these two sentences indicates that the property of gender relevant for coreference is, at the very least, more complex than a formal syntactic feature. This observation by itself is not novel (e.g., Joseph 1979; Hess, Foss & Carroll 1995; Garnham, Oakhill & Reynolds 2002; Duffy & Keir 2004; Nieuwland & Van Berkum 2006; Gygax et al. 2008; Pyykkönen, Hyönä & van Gompel 2010; Collins & Postal 2012; Frazier et al. 2015). Thus this paper develops a formal treatment of how certain “types” of gender can match or mismatch during coreference dependency resolution and what this means for the linguistic encoding of gender identity across languages.

One possible model to explain how gender is conceived and applied to linguistic referents is described in the following sections. It represents a self-consistent, comprehensive model that can be tested empirically. Furthermore, it provides a starting point for interdisciplinary research into the many linguistic facets of gender. In particular, I anticipate this approach will benefit linguistic work which examines phenomena where an individual’s gender identity and/or gender expression is relevant, as well as work which makes use of biosocial gender, including phenotype and hormonal profiles. I especially hope to encourage linguists who make use of psycholinguistic properties of gendered pronouns in their research to be aware of the issues surrounding the various ways in which gender broadly construed and cognition may interface.

2.1 Grammatical gender

Grammatical gender comprises formal morphosyntactic features. They are the properties of words that allows the formal grammatical process of agreement to be carried out. This includes agreement of grammatical gender categories such as masculine, feminine, neuter, common, etc.² These features are properties of the morphemes themselves, and may be independent from the real-world biosocial genders associated with the referents. However, Corbett (1991) notes that there is a tendency for languages to correlate grammatical gender with the gender of the referent, particularly if human. Moreover, Comrie (2005) adds that there is a tendency for personification of animals and inanimate objects in languages with grammatical gender to correlate with the grammatical gender of the noun phrase. This is further supported experimentally by Konishi (1993), who suggests that perception of inanimate referents are semantically influenced by grammatical gender cross-linguistically. Finally, it may be noted that languages that use different noun classes for subdividing humans almost always divide along a male-female category line independently of how many other noun classes are present or what other types of nouns are included in those two classes. Subdivision of humans across noun classes is a crucial point here, as noun classes that use animacy as a distinction will group humans in the animate category, independent of human gender. I am not aware of any language that encodes more than two human genders grammatically.³ Even languages of people whose culture encodes more than two human genders do not seem to encode those genders grammatically, as illustrated in Section 3.3.

In (2), the Dagestani language Tsez places animals in a noun class that is distinct from ones that include humans, and Comrie (2005) reports that grammatical gender does not change to reflect the gender roles of personified animals. This contrasts with languages like English, in which personification or anthropomorphization can result in the use of gendered third person pronouns to refer to non-human animals and inanimate objects that would otherwise be referred to with it, the inanimate/non-human pronoun. I argue that the variation in use of grammatical gender points to a deeper, more complex system of gender categorization both grammatically and conceptually. That is, grammatical gender is in principle independent from other types of gender but the way it is deployed and the way it influences non-grammatical interpretation suggests it is not entirely decoupled.

The extracts in (2) come from a story in which a rooster (definitionally male) and a hen (definitionally female) are married, but the rooster has another romantic partner (a frog, no specified gender explicitly or grammatically) thus causing strife in the rooster and hen’s relationship (Comrie 2005). Although all animals fall into the third noun class (III) in Tsez, the words for rooster (mamalay) and hen (onoču) still have defined conceptual or semantic genders despite this not being reflected in the grammatical features. That is, the grammatical gender of the frog, the hen, and the rooster are all obligatorily noun class III, with agreement marked on the verb, which is not used for humans of any gender.

In (2), the gender roles of the three characters are inferred through cultural norms, e.g., marriage, and expectations, e.g., housework and romantic liaisons, rather than solely through grammatical gender such as noun class morphology. In the case of łoħro (frog) there is no lexical distinction between the males or females of the species. Thus, the interpretation that the frog is a female interloper in the birds’ marriage is not linguistically encoded. Comrie (2005) reports that Tsez speakers uniformly interpret the frog to be female and not male, although it would not be ungrammatical for the frog to be male. Thus, the interpretation of the frog as female must come from the cultural expectations of the speakers rather than from their language.

Compare rooster and hen in English and Tsez to languages like German (masculine Hahn and feminine Henne, respectively) and Russian (masculine petux and feminine kurica, respectively), in which the grammatical gender of the words and the real-world sex of the animals is congruent. In Russian, the word for frog (ljaguška) happens to be grammatically feminine, thus congruent with the anthropomorphic gender role of the frog character. However, in German the word for frog (Frosch) is grammatically masculine. Comrie (2005) reports that this makes it difficult, potentially bordering on ungrammatical, to use Frosch in translation, since the grammatical gender is incongruent with the anthropomorphic gender role of the frog character. According to him, the way to translate this story without indicating a homosexual relationship between the rooster and the frog would be to change the species of the interloping character to a feminine word like toad (Kröte). This suggests that the grammatical gender of a word and the gender role of the character are conceptually connected, even though this need not be the case (Konishi 1993; Irmen & Kurovskaja 2010). On the other hand, what might be called grammatical gender in English, which is restricted almost entirely to third person pronouns, appears to be fully coupled to conceptual gender since the pronoun used would determine how the character’s gender role is interpreted. This leads us to the question: what role does grammatical gender play in English, if any?

It is unclear whether or not grammatical gender plays a role in English syntactic operations or psycholinguistic processes. It has been argued that English has completely lost grammatical gender, based on historical changes and loss of productive gender morphology (Baron 1971). Certainly, there is no overt gender agreement between nouns, adjectives and articles. However, Bjorkman’s recent treatment of gender agreement between names and pronouns makes a case for a limited grammatical gender system in English, in which sentences like (3) display a contrast in acceptability (Bjorkman 2017).

Bjorkman observes that sentences like (3-a) are more acceptable than (3-b), even when the surgeon is known to all parties, and suggests this is due to names having grammatical gender (i.e., a ϕ-feature) in English, which must then agree with the pronoun, at least for some speakers. A reviewer points out that (3-b)’s acceptability is contextually dependent, as Johnathan’s gender identity and the interlocutors’ knowledge of this will affect the acceptability of the sentence. For instance, however, consider people like anti-bullying activist Jeffrey Marsh who is nonbinary and whose pronouns are they/them, but whose forename is strongly biased as masculine. In this case, it is unlikely that speakers will have a lexical entry for Jeffrey that doesn’t have a masculine ϕ-feature, but this does not change that Jeffrey Marsh’s pronouns are they/them and using other pronouns would be misgendering. Speakers would then need to have explicitly acquired the knowledge of which pronouns are appropriate in order to avoid misgendering a person whose gender identity is not immediately inferred from culturally specific cues in gender expression and gender role.

Whether or not English makes use of grammatical gender to determine gender congruency between coreferring elements, an argument for ϕ-features on names must account for how gender (conceptual and/or grammatical) is associated with their referents, since gender bias of names is wildly variable and mutable, more akin to cultural shifts than language change (Barry & Harper 1982; 1993; Van Fleet & Atwater 1997; Lieberson, Dumais & Baumann 2000; Hahn & Bentley 2003; Barry & Harper 2014). Thus, for grammatical gender to play a role in English, it would need to be the case that names and a limited number of nouns have ϕ-features for gender, but that agreement with a coreferring pronoun is optional in cases where the antecedent does not have a ϕ-feature for gender. To this end, I will set aside the status of grammatical gender in English for the time being and return to it in Section 4.1.3.

2.2 Conceptual gender

Conceptual gender encompasses a large number of closely related terms currently in use in the literature. This includes semantic gender (e.g., Asarina 2009), definitional gender (Kreiner, Sturt & Garrod 2008) and notional gender (i.e. natural gender, but see McConnell-Ginet (2015) for why the term ‘natural’ is inappropriate), which are ways of associating lexical items with masculine or feminine properties, but without necessarily attributing formal features to them.

This may be illustrated by the strong gender biases of many English occupational terms (e.g., Garnham, Oakhill & Reynolds 2002; Kennison & Trofe 2003; Duffy & Keir 2004; Gygax et al. 2008; Kreiner, Sturt & Garrod 2008). These biases, although in principle mutable, seem to hold consistently and for large swathes of the population. This bias underpins the confusion caused by the “riddle” cited in Reynolds, Garnham & Oakhill (2006) (originally from Sanford 1985: 311):

A man and his son were away for a trip. They were driving along the highway when they had a terrible accident. The man was killed outright but the son was alive, although badly injured. The son was rushed to the hospital and was to have an emergency operation. On entering the operating theatre, the surgeon looked at the boy, and said, “I cant do this operation. This boy is my son.” How can this be?

The difficulty of interpreting the surgeon as being either the son’s mother or any other parental figure besides the previously mentioned father is reflected in the enduring nature of this riddle. In either case, surgeon is demonstrated to have a strong male bias despite there being no definitional requirement for surgeons to be men. While gender is not overtly morphologically or grammatically marked in English, there is still some sort of conceptual bias that can be difficult to override.

In Russian, conceptual gender and grammatical gender sometimes clash. Asarina (2009; 2011) observes that doctor (vrach) is in the first noun class (I), which typically includes human male nouns, among other things. However, when referring to a doctor who is a woman, there are a few strategies that may be employed in different registers.⁵ See also King (2015) for another detailed account of mixed agreement in Russian. This is a particularly clear case of a clash between grammatical and conceptual gender because there are two loci that agreement could target and the different structural positions each target a different locus.

The explanation Asarina gives for how Russian can have mixed case agreement is that there is a structural representation of the grammatical feature in the syntax (as opposed to in the semantic representation). This means that an unpronounced functional projection encodes something about conceptual gender. For example, in Russian, there is a functional projection in sentences like (4), i.e. <wmn>, and the agreement is triggered by the closest class feature in the tree, i.e. noun class II. Thus the adjective agrees with the grammatical gender of the noun (masculine/noun class I, because ‘doctor’ is in the first/masculine noun class), but the verb agrees with the conceptual gender of the noun phrase (feminine/noun class II, because the doctor is a woman).

In this representation, it’s argued that ‘dental’ agrees with ‘doctor’ because the masculine ϕ-feature from vrach is the closest target of agreement in the tree, whereas the verb agrees with the (unpronounced) functional head <wmn> as it is the closer target of agreement. This requires the functional head be tied to the discourse context, thus is more flexible and potentially more defeasible than if such a functional head were absent or unavailable in the language. In fact, this type of functional head only seems to be available for human referents and not animals, even when the animals are anthropomorphic (Comrie 2005). This suggests that there is some super-level of categorization in Russian that distinguishes animals and humans even in contexts where animals are filling human-like gender roles. I will set aside the question of distinguishing animals and humans grammatically, but I will also suggest that the categories could be cognitively structured in a manner similar to gender.

On the other hand, this is not also the case in formal registers of European French.⁶ In (5), the form of the noun (masculine) does not change to match the gender of the referent, although this is at least partly for orthographic reasons.

In formal European French the form of the noun does not change. All gender agreement must match either the grammatical gender of the head noun or the conceptual gender of the referent. Thus, any mixed agreement should only occur when the conceptual gender of the referent mismatches the grammatical gender of the head noun. In this case, the <wmn> features Asarina proposed would be located above N but below any of its projections, which is prima facie counter-evidence for a syntactic head that governs gender agreement in French.

Responses by Francophone colleagues to my informal queries indicate that mixed agreement in formal French is marginal in some speakers, since there is often an alternative form of the noun that would match the conceptual gender of the referent. Thus, further investigation into the nuances of mixed agreement in French is warranted. Further investigation into agreement with nonbinary conceptual gender will also become a viable line of research, as users of French (much like Russian) are in the early stages of developing gender-neutral or nonbinary grammatical solutions to conceptual gender (Shroy 2016).

Returning to sentence (1) for instance, cowgirl is definitionally female, but can be used for a male/masculine referent in certain circumstances. The feminine definition associated with cowgirl is thus defeasible, since gender agreement between cowgirl and his should be impossible if the property being checked is a morpho-syntactically defined ϕ-feature. This is not incompatible with English having formal gender features for some words, but I argue that it is strong evidence that what is primarily relevant for coreference resolution is not the morphosyntactic feature. This argument will be elaborated upon in Section 4.1, below.

Furthermore, there is evidence from developmental psychology and language acquisition that young children acquire labels for gender categories before they are able to consistently sort people into those categories (Fagot & Leinbach 1993; Welch-Ross & Schmidt 1996; Bussey & Bandura 1999; O’Brien et al. 2000; Zosuls et al. 2009; Waxman 2010; Fausto-Sterling 2012). At this point in development, (at least) two gender categories are present but not enough input has been received to develop a consistent rubric for evaluating the massive variation present in the population. For instance, children may be able to use the proper pronouns for common and canonically gendered referents (e.g., “mommies” and “daddies”) but fail to generalize identification criteria to novel referents that deviate in one or more ways (e.g., men with long hair, women wearing collared shirts) (Taylor & J. A. Hall 1982; Fagot & Leinbach 1993; Armann & Bülthoff 2012; Ansara & Hegarty 2013). This may indicate that gender categories are developed and refined by repeated exposure to exemplars and top-down societal reinforcement. The acquisition of gender category labels could conceivably support the acquisition of the conceptual categories. I am unaware of any cross-linguistic differences in age of acquisition of gender categories, but should such differences exist, this would support my claim that linguistic labels feed into non-linguistic categorization behaviors.

2.3 Biosocial gender

Biosocial gender is, fundamentally, an individual’s gender as it is experienced internally. In addressing this type of gender, a few terminological clarifications are necessary. I will assert a distinction between sex and gender, which are widely confounded terms in linguistics and psychology (Cheshire 2002; Ansara & Hegarty 2013). Herein, sex refers to biological properties such as karyotype (XX, XY, etc.) and phenotype (e.g., internal and external anatomy, circulating hormonal milieu). Even in biological terms, sex is not a binary property since the physical traits contributing to an organism’s sex can vary along multiple dimensions. See Fausto-Sterling (2019) for a recent review. As an example of an edge case, people with Complete Androgen Insensitivity Syndrome (CAIS) may have XY chromosomes but a predominantly female phenotype (e.g., Hughes et al. 2012). However, sex is still often used as a shorthand for distinguishing the bimodal nature of the male-female spectrum (Lorber 1996; J. L. Johnson & Repta 2012).

This definition of sex overlaps with biosocial gender. More precisely, biosocial gender is the multidimensional property of an individual as determined by their biology and cultural norms of identity expression. What distinguishes biosocial gender from other types of gender is that, as an external observer, one’s accuracy of categorization is impossible to assess without input from the individual’s introspection and medical history. That is, biosocial gender may not be something that can be doubtlessly determined without detailed anthropological, introspective and potentially invasive medical analyses. This is because social pressures and societal norms can contribute to an individual representing themself in a way that is inconsistent with the way they categorize themself (Fausto-Sterling 2012; Ansara & Hegarty 2013; Zimman 2017). One clear illustration is the case of transgender people who are “in the closet” or otherwise representing themselves as the binary gender category to which they were assigned at birth, despite not identifying as this gender. Here, an individual’s biosocial gender might be in direct conflict with the gender with which other people would categorize them, that is, the conceptual gender other people attribute to them.

Our current census data suggests that the majority of people have a gender identity that fall into a bimodal distribution of biosocial genders (0.4% of respondents in a UK survey reported thinking of themselves as a way other than ‘male’ or ‘female’; Glen & Hurrell 2012). But many individuals do not categorize themselves with a discrete binary label, and it would do the science and the individuals a disservice to gloss over the often subtle and diverse variations in gender identity present in the population at large, even within male and female categories (J. L. Johnson & Repta 2012). Despite the potential complications in identifying the precise biosocial gender of an individual, it is still an important factor for phenomena involving social identity and certain physiology relevant to linguistic processes such as auditory brainstem responses (Liu et al. 2017). One’s biosocial gender can affect mental, emotional, and social well-being outcomes, indexical properties of speech, and perception of in-group versus out-group (Rubin & Greene 1991; Zimman 2017; K. Johnson et al. 2019). Therefore, it is important to explicitly define biosocial gender as distinct to ensure it is not confounded during investigation of phenomena associated with either grammatical or conceptual genders.

3.1 Personal names as antecedents

Personal names comprise a large portion of antecedents used in empirical investigations and syntactic judgments of English coreference, presumably due to their intuitive gender-specificity, although this has been identified as an issue in stimulus design (Kasof 1993; Merritt & Kok 1995; Van Fleet & Atwater 1997; Lieberson, Dumais & Baumann 2000; Gabriel et al. 2008). However, English lacks overt morphological marking on names to unambiguously distinguish a correct assessment of the gender identity of the referent, where a ‘correct assessment’ would result in a conceptual gender that is congruent with the referent’s gender identity. A clear example of this problem is illustrated in (6-a), in which the two given pronouns can corefer with the name equally well in the absence of disambiguating context (such as whether the Taylor in question is Taylor Swift, a woman, or Taylor Lautner, a man. As for Taylor Mason, a nonbinary character played by the nonbinary actor Asia Kate Dillon, (6-b) is the appropriate construction (Dillon 2017), although the processing cost and intuitive acceptability of this linguistic structure, in terms of linguistic judgments, is currently a subject of investigation and may vary in reported ‘acceptability’ (Konnelly & Cowper 2017; Ackerman 2018; Conrod 2018; Prasad, Morris & Feinstein 2018).

One possibility is that the name Taylor is stored in the lexicon as discrete entries (e.g., Taylor_<masc>, Taylor_<fem>). The possibility of the lexicon containing Taylor_<nonbinary> is a logical possibility but cannot be discussed in much more detail at this point without introducing speculation because of the current dearth of empirical studies on nonbinary gender perception and its influence on lexical categories. If we consider the two binary grammatical genders, a comprehender may retrieve one of the two entries initially, but have to revise the selection if conflicting information is received at a later time during comprehension. The presence of different lexical entries for each string-identical name, each with a distinct valuation of a gender ϕ-feature, makes testable predictions regarding the learning and application of new lexical entries. One can quickly learn a new name or a new use of a common name, but if extensive previous experience with a common name (e.g., Michael_<masc>) influences the processing of a newly encountered and rare version of the name (Michael_<fem>), this might be observable in behavioral or psychophysical measures. If this is the case, it would need to be determined how names most often used by nonbinary people are stored in the lexicon, and if these entries are associated with a specifically nonbinary feature or another configuration of grammatical gender. If names are stored generically with some gender label determined by stereotypicality or statistical probability, then by familiarizing a naïve participant to an uncommon or novel pairing between a name and gender (e.g., a woman named Michael or a boy named Sue), there should still be a detectable processing cost to forming a coreference dependency between the pronoun and name. However, if names instead receive gendered properties from domain-general or world knowledge, then retrieval of the uncommon entry should be facilitated more by the context and less processing cost should be observed (Pyykkönen, Hyönä & van Gompel 2010). See Cai et al. (2017) for examples of how long- and short-term learning can be tested.

Another possibility is that “unisex” or names which are not strongly associated with a particular gender category are morpho-syntactically underspecified for gender (e.g., Taylor_<0>), and whatever gender assumptions are made about the referent are done so without reference to the lexicon or morphosyntactic features. However, it is not immediately clear what the implications of this configuration would be or how this could be tested. At the very least, it would be necessary to conduct extensive evaluation of each individual participant’s experience with the target names and gender nonconformity and examine effects from the perspective individual differences (Barry & Harper 1982; 1993; Van Fleet & Atwater 1997; Lieberson, Dumais & Baumann 2000; Barry & Harper 2014).

3.2 English as a leader of change

More than just a language of convenience, English has certain properties that allow dissociation of the three proposed types of gender. English marks gender (broadly construed) on its third person pronouns (she, he), but it does not have consistently overt or productive morphological agreement for gender. Numerous studies demonstrate strong gender biases of certain noun phrases (e.g., surgeon, pilot, nurse, babysitter), but these are defeasible which indicates the biases are tied to conceptual gender rather than grammatical gender (Garnham, Oakhill & Reynolds 2002; Kennison & Trofe 2003; Duffy & Keir 2004; Kreiner, Sturt & Garrod 2008; Pyykkönen, Hyönä & van Gompel 2010). Furthermore, English has some remnants of gendered morphology (actor/actress, aviator/aviatrix) and definitionally gendered nouns (mother, father, cowgirl, bellboy). It is conceivable that the morphologically gender-marked words do have grammatical gender. At least those marked as <fem> are the most likely to have retained grammatical gender in English, as those are distinctly non-default and definitional. As for the definitionally gendered words, I have already demonstrated that it is possible to find contexts where the gender is defeasible. This suggests that these words are not grammatically gendered, or at least the relevant type of gender is conceptual gender and not grammatical gender.

Finally, in cultural terms, English has been at the international forefront of informal, community-based development of nonbinary language and so-called “neopronouns”. Examples of neopronouns include Spivak pronouns introduced by Spivak (1990: xv) and gender variant neologisms described in Centauri (2013), Hord (2016) and Bradley et al. (2019), among others. The combination of linguistic innovation, on-going sociological research, and prominence of media exposure makes the English language uniquely situated (in the present moment) to development and linguistic change regarding gender categories inclusive of nonbinary gender(s) and gender neutrality (Page 2013; Brutt-Griffler & Kim 2018).

3.3 Other gender paradigms

Many cultures around the world have established and traditional nonbinary, queer, and third-gender categories. Navajo people called nádleehí are traditionally characterized as participating in gendered behaviors of the “opposite sex” (Epple 1998). However, the Western concepts of being ‘transgender’, ‘queer’, or ‘homosexual’ do not quite capture the Navajo cultural concept. To this end, the terms ‘alternate gender’ and ‘two spirit’ have been used to describe nádleehí. While these cultural concepts seem to provide potential for investigating concepts of gender categories and language, the Navajo language does not mark grammatical gender on human pronouns. Furthermore, the strategy for speaking about nádleehí in English is to use standard binary pronouns in a similar manner to how binary trans men and trans women use English pronouns, and “not neuter pronouns or pronouns specific to nádleehí” (Epple 1998: 279).

This seems to be very similar to how Māori culture and language encodes gender outside of the binary (Murray 2003). The terms whakawāhine and whakatāne are “terms which translate roughly to ‘becoming’ or ‘making’ woman or man, indicating a transcendent or permeable gendered identification” (Murray 2003: 240). However, as in Navajo, Māori grammatical gender does not distinguish conceptual gender on pronouns.

The hijras of India are similarly difficult to quantify in Western terms, considering themselves to be “‘deficiently’ masculine and ‘incompletely’ feminine” (K. Hall & O’Donovan 1996: 229). Linguistically, they use the grammatical (and conceptual) gender system of Hindi to express their relationship to their gender identities and their affiliation to the community with a mix of grammatical gender and fluid interaction with binary gender roles.

In Buginese, a language spoken on Sulawesi, Indonesia, by approximately five million people, there are distinct lexical items for each of the five recognized genders, but gender is not otherwise encoded grammatically (Graham 2004). The five genders can roughly be translated into Western concepts as feminine woman (makkunrai), masculine man (oroani), feminine man (calabai’), masculine woman (calalai’), and nonbinary (bissu).⁷ Importantly, people who identify as calalai’ and calabai’ do not wish to conform to feminine or masculine standards for the makkunrai or oroani, respectively, but rather have their own standards for gender expression. Furthermore, people who identify as bissu are considered to have both masculine and feminine elements in their souls and thus serve spiritual roles in the community (Graham 2004). Much like in Navajo and Māori languages, Buginese does not distinguish pronouns for these five conceptual gender categories.

Generally, pronouns are more likely to mark animacy as a ϕ-feature in these example languages. When a language does mark grammatical gender, nonbinary gender categories can be indicated through shifting use of standard binary gender agreement (e.g., K. Hall & O’Donovan 1996). Investigation of gender perception, category acquisition, and development in other cultural paradigms will bring crucial supplementary information to our understanding of how different types of gender are mentally represented and how they influence each other during linguistic and non-linguistic cognitive behaviors.

4.1 Checking for congruency

If formal morphosyntactic gender features are present in a language like English, but cannot be used to model how the parser checks for congruency in coreference dependency formation, what purpose do they serve? I will not argue for or against English having formal grammatical features for gender, but rather that such features are irrelevant during coreference dependency formation. Instead, English and languages with similar gender systems rely on conceptual gender for evaluating gender congruency in real time. In order to describe how such a system operates, a three-tiered scheme of linguistically and cognitively encoding gender is posited below.

The three tiers comprise an exemplar tier, a category tier, and a feature tier (Figure 1). These tiers are not meant to represent actual processing mechanisms or structures in the mind. Rather, they are abstract categories of processes or representations that can be used to map behaviors and empirical observations to theoretical properties of grammars and other mental mechanisms and modules. Thus, each tier is designed to be as theory-agnostic as possible to provide the most utility across the various popular frameworks.

The first tier, the exemplar tier, is represented by a strongly bimodal continuum indicative of how biosocial gender and conceptual gender can vary within a population. Although only color and height vary in this diagram, one may imagine that this tier has many more dimensions that could align with variation in gender role, gender expression, and overt biosocial properties. The second tier, the category tier, comprises two discrete, non-overlapping spaces overlaid on the exemplar tier. These categorically distinct spaces represent the binary genders as might be conceptualized by someone from a society that reinforces a strictly binary gender schema. However, even so, one might not be able to categorize all individuals into one of these spaces, so the gap between the categories allows for ambiguous, nonconforming, and ‘other’ instances to exist outside the binary. If the exemplar tier is, indeed, multidimensional beyond what can be represented on paper, I request that the reader accept that these two categories are not as simple as the rectangles depicted, and their apparent shapes simply due to the limitations of the medium. For instance, if one dimension encodes hair length as a gendered property of appearance, then a man who otherwise fits all other stereotypically masculine traits but has long hair would align predominantly but not completely with the category tier’s binary (masculine) category. Finally, the third tier is the feature tier which, unlike the previous two, comprises labels associated with spaces rather than spaces themselves. In this illustration, the labels are the grammatical features <fem> and <masc>, corresponding to a language that has two noun classes. A language with more noun classes (or fewer) would have a different configuration for the labels.

4.2 Typological evidence

Together, these tiers describe three levels of encoding of gender, broadly construed, that a language (or an individual) may draw upon in order to determine the gender congruency of a pronoun and candidate antecedent during real-time coreference resolution. In (11), I describe three possible configurations of languages based on the broad criterion and the three tiers of mental representation of gender. These three configurations are points within a hierarchy of how rigidly a language (or individual) adheres to matching the gender encoded on the feature tier to assess gender congruency. While I list languages as examples of these points in the hierarchy, I also suggest that individuals may vary within what an individual language permits. That is, a speaker of French who finds any mixed agreement to be unacceptable would be applying the description of a Strict feature rather than where I have categorized French on the whole (Mixed feature). Similarly, English speakers who find singular they difficult to learn or use may be employing more of the ‘mixed’ matching strategy than ‘absent’. Furthermore, English speakers who do use the ‘absent’ strategy may also vary in the shape and adaptability of their category tiers, thus introducing intra-language variation in acceptability.

Tsez exemplifies a strict matching strategy, with (2) demonstrating a rigid grammatical gender system for anthropomorphic animals and other noun phrases (Comrie 2005; Corbett 2015). Thus, no matter what the conceptual genders of the characters in the story are, the agreement is consistent with the morphosyntactic features of the lexical items. This may also be the case for some speakers of French for whom maire is necessarily <masc> and mairesse (mayor_<fem>) is a viable alternative. While I am being careful to avoid a neo-Whorfian claim that language shapes or limits our thought, I think it is reasonable to posit that the categories present in one language and not in another could draw attention to different non-linguistic properties of the members of those categories, thus creating subtle distinctions in the boundaries and shape of the categories. In this way, we might explain how grammatical gender can limit conceptual gender in practical translation without necessarily claiming that German speakers think toads are necessarily feminine (Konishi 1993; Irmen & Kurovskaja 2010). Languages with intermediate strategies like French and Russian would then show some mixed properties wherein formal features are checked during coreference resolution, but may be overridden given contextually appropriate information (Asarina 2009; 2011). Moreover, languages without grammatical gender would then rely entirely on the conceptual categorization of the antecedent to evaluate coreference feasibility.

Where does English fit into this hierarchy? As it has been claimed that English no longer has grammatical gender (except, possibly on pronouns) (Baron 1971), it might be an absent feature language. However, Bjorkman (2017) suggests that English does have limited use of grammatical gender agreement, particularly when referring to named individuals. If so, we might expect such cases to elicit psycholinguistic/cognitive behaviors that are similar to those observed in languages that make use of the feature tier. However, testing this is made difficult by the limited circumstances in which English could have grammatical gender. The potential environments for detecting grammatical gender in English overlap with environments where conceptual gender (as determined by the category tier) could be an alternative source for checking during coreference resolution. That is, words that could have formal gender features (as Bjorkman suggests, personal names) should also typically receive a gender property from the cognitive gender of the referent, encoded in the category tier.