Predicate order and coherence in copredication

ABSTRACT

This article proposes that predicate order and coherence relations are the two major determining factors in copredication licensing, resolving a long-standing puzzle over the criteria for constructing acceptable copredications. The effects of predicate ordering are claimed to be anchored around semantic complexity, such that copredications with semantically Simple–Complex predicate orderings are more acceptable than the reverse. This motivates a parsing bias, termed Incremental Semantic Complexity. Particular ways of implementing this parsing bias are discussed. The effect of predicate coherence is claimed to be anchored around a sense of causality and featural commonality. Lastly, a hierarchy of possible copredications is outlined (the Copredication Hierarchy), helping to delimit the modelling of copredications to a greater extent than has previously been possible.

KEYWORDS:

• Copredication
• polysemy
• internalism
• sense order
• incremental semantic complexity

1. Introduction

The act of ascribing multiple, semantically distinct types to a single nominal has come to be known as copredication. If a book is deemed ‘interesting’ and also ‘blue’, informational and physical senses are simultaneously being referred to without resort to multiple uses of the nominal (Asher 2011). Given its potential significance for a considerably large number of topics in psycholinguistics and philosophy of language, copredication is a relatively neglected topic.¹ It is standardly defined as ‘a grammatical construction in which two predicates jointly apply to the same argument’ (Asher 2011, 11; Cruse 1986, 2000). More specifically, copredication occurs when ‘two or more equally connected predicates have different selectional restrictions for the same argument’ (Kinoshita, Mineshima, and Bekki 2018, 155).

Mixing an abstraction with a concrete meaning appears to be part of the productive nature of the lexicon. Indeed, an informal definition of copredication might be ‘category mistakes which are nevertheless felicitous’. Unlike Fregean and Russellian logicians, who were primarily interested in language’s role in judgment formation, nineteenth-century semanticists and semasiologists saw these forms of polysemy as an integral part of linguistic creativity: Bréal (1897) saw it as reflecting most clearly the conceptualising capacities of humans.

A number of researchers have cleared the ground for a formal account of certain aspects of copredication. Yet, much remains unsystematically explored, such as the factors involved in copredication acceptability. This will be the major focus of the present article. I will be primarily concerned with cases in which an argument shifts between polysemous senses in copredication, and the apparent constraints on this shifting.

To illustrate, consider the copredications below (the examples in (1) are taken from Asher 2011, 63, with ‘#’ marking putatively degraded semantic acceptability):²

• (1) a. The city has 500,000 inhabitants and outlawed smoking in bars last year.

  b. #The city outlawed smoking in bars last year and has 500,000 inhabitants.

• (2) a. The White House is being repainted and issued a statement concerning taxes.

  b. #The White House issued a statement concerning taxes and is being repainted.

I will largely focus on nominals that have one physical sense, with their other senses differing in their abstract sense (Falkum 2011). For instance, book, lunch and city all host the same concrete sense of physical (physical artefact, food or building; all sub-types of physical), but they vary in their abstract sense (information, polity/people, event). I will be concerned with lexical items permitting various combinations of the four most commonly discussed sense types in the literature: physical, information, event and institution (Arapinis 2015; Asher 2011; Jezek 2016; Pustejovsky 1995).³

After focusing on apparent constraints on sense shifting, I will introduce ‘coherence relations’, which concern the degree of semantic coherence that discourse segments exhibit. In particular, I will focus on two specific types of coherence relations: causal connection (as in, ‘The newspaper was sued and it was deeply offensive’, where one can readily infer a causal relation) and extensional overlap (as in, ‘The school banned videogames at lunch and had approximately 200 gamers’, where a shared feature of game/gaming helps make the predicates cohere). I will argue that coherence is a major factor in copredication acceptability, connecting this component to a number of existing claims in the literature, and re-evaluating them.

2. Defining copredication

While polysemy may be widespread, what is commonly and synonymously termed ‘logical’, ‘complex’ or ‘inherent’ polysemy has been the subject of a more focused set of investigations in philosophy of language and linguistics (Apresjan 1974; Carston 2021; Leivada and Murphy 2021; Pustejovsky 1995; Ostler and Atkins 1992). This form of polysemy occurs when different senses are deemed an essential, inherent part of an entity and are not ‘accidental’ (Asher 2011):

• (3) a. Mary doesn’t believe the book.

  b. John sold his books to Mary.

Cases of complex polysemy appear to have two core diagnostics: (a) copredication involving coordinating distinct senses, and (b) anaphoric binding involving pronominal reference to the same nominal but involving a distinct sense (Norrick 1981, 112–115):

• (4) a. The book was interesting and weighed a ton.

  b. That book was brilliant. Put it back on the shelf.

Copredication involves associating semantically related senses, but senses of distinct semantic types, which are associated with a single polysemous word to derive its meaning, such as in lunch, which can last for a given duration whilst also being a substance. Lunches therefore seem to be physical-event composites, while books would be physical-information.

Perhaps above all else, what copredication is can most readily be captured in the following way: Copredication is about conflict. The intuitive notion of conflict is deeply rooted in all definitions of, and perspectives on, copredication.⁴ Semantic theories see copredication as a conflict in type selection, whereas pragmatic and philosophical theories see it as a conflict in referential relations. This may be what Jezek and Melloni (2011, 5) have in mind when discussing ‘an ambiguity available by virtue of the semantics inherent in the noun itself’. For each pair of ‘self-contradictory’ (Elbourne 2011, 26), or ‘contradictory in type specification’ (Jezek and Melloni 2011, 4) or ‘incompatible’ (Jezek and Vieu 2014 and Xue & Luo 2012) senses, there is a relation connecting them (see also Evans 2006, 2013, 2015). In brief, copredication has been analysed as an unusual, complex form of polysemy, involving constructions which ‘ascribe two properties that … can’t be jointly instantiated’ (Liebesman and Magidor 2019, 1).

While the most standard copredication pattern is [V[Det Adj N]], such as burned the offensive books (Jezek and Vieu 2014), copredication can also involve (i) splitting the copredication between a main clause and a subordinate clause (‘The building, which started yesterday, will be very nice’; Jacquey 2001, 15), (ii) a modificational structure (‘The most provocative newspaper of the year has been sued by the government’), (iii) anaphora (‘He paid the bill and threw it away’; Asher 2011, 63; Cruse 1986), or (iv) temporal disjunction between predicates denoting distinct situations (‘The translation was completed last week and is on the table’) (see Frisson 2009).⁵

Raising briefly a point of context, Cruse (1986) originally used copredication purely as a diagnostic for complex polysemy. During the 1990s, interest shifted from using copredication merely as a diagnostic to viewing it as an independent phenomenon worthy of theoretical investigation (e.g. Pustejovsky 1995). By the 2000s, a small number of semantic theories had tried to accommodate it and also use it to object to traditional externalism (referentialism) in philosophy of language (e.g. Chomsky 2000), and within the past few years copredication has also infiltrated certain strands of psycholinguistic research. Perspectives from across these domains will be consulted below.

3. Accounting for copredication

This section briefly reviews some theoretical treatments of copredication, focusing on accounts most relevant to my focus on sense shifting dynamics. Even though the terrain across accounts will differ, I will argue that the underlying theme is that particular discourse effects appear to have a major impact on the licensing of copredication.

3.1. Formal semantics accounts

While polysemy and pragmatics have been explored in only relatively minor detail in relation to copredication, the most comprehensive and widely discussed account of complex nominals like book and committee comes from formal semantics. The most influential account is Pustejovsky’s (1991, 1995) Generative Lexicon theory. Pustejovsky (1991, 1995) proposes that word meaning is structured by four ‘Qualia roles’ which capture human understanding of objects and their relations:

formal: The basic category that distinguishes an object within a larger domain.

constitutive: The relation between an object and its constituent parts.

telic: The object’s purpose and function.

agentive: Factors involved in the object’s origin or ‘coming into being’.

Pustejovsky (1995) uses these ideas to argue that the different meanings of book arise from the foregrounding of its distinct aspects (see also Srinivasan and Snedeker 2011).

While traditional semantic theories have focused on selectional restrictions and domain-specific constraints encoding a ‘static’ (Pustejovsky and Boguraev 1993) representation, the Generative Lexicon attempts to explain why new senses can emerge compositionally through the above system of semantic types (Pustejovsky 1991, 426–427). Less emphasis is placed on extra-grammatical and extra-semantic processes such as pragmatics, and the explanatory burden is typically placed on semantic type structure, whereby complex polysemous nominals are represented as ‘dot-types’ or ‘dot-objects’ (e.g. book would be of the type physical • information). These are at the core of all copredications; their meaning ‘combines two or more different semantic types, which are equally important in the word’s definition’ (Pustejovsky and Batiukova 2019, 175).

Overall, natural language permits rich combinations of semantic senses (Pietroski 2012, 2018), as seen in copredication via dot-types: physical • information (book, video), event • physical (construction, examination), event • information (exam), event • food (lunch, dinner), event • person (appointment) (see Gotham 2012, 2015, 2016).

3.2. Linear dot-types

Chatzikyriakidis and Luo (2015, 2016, 2018), working within a Generative Lexicon framework, claim that a meaning shift from one sense to another (e.g. physical to information) is a flexible process exhibiting few constraints, relying on a single lexical of representation. The authors use this observation to motivate the claim that with newspaper this meaning shift exhibits a degree of rigidity – because, the authors claim, we need the physical and information senses to not occur with the institution sense. Chatzikyriakidis and Luo claim that structures such as the following exhibit degraded acceptability, and they suggest that the institution sense of newspaper cannot co-occur with other senses:

• (5) That newspaper is owned by a trust and is covered in coffee.

In order to account for this, they invoke resource sensitive dot-types, or linear dot-types, which are single lexical entries. In linear logic (Girard 1987) there are no rules of weakening and contraction, and so an assumption has to be used only once. This proposal leads to the conclusion that if the institutional sense is used it must appear in isolation. However, it seems somewhat psychologically inflexible to claim that once the institution sense is triggered the other senses cannot be accessed, as we will see below.

3.3. Univocity

In contrast to the above, Liebesman and Magidor (2017, 2019) argue that theorising about copredication requires no revisions to type systems or metaphysics (i.e. the abandoning of referential semantics, as in Chomsky 2000; see also Collins 2015, 2017a, 2019; Murphy 2014, Vicente 2019) if it can be argued that book (physical) and book (abstract) are separate lexical entries, hence with distinct referential scope. They claim that nominals like book are not ambiguous but encode a single sense (a form of univocity) simultaneously designating both physical and informational entities, and that it is via contextual restriction that multiple readings are permitted.

For example, an argument they make against the ambiguity of book is that cases like ‘Three colours are on the canvas’ – involving three distinct shades of red – cannot automatically be seen as supporting the ambiguity of the nominal colours. Rather, it is other words in the sentence like three which constrain the meaning of colour, or it is red which is ambiguous; in addition, colour is clearly contextually restricted (2017, 136). Yet, it might simply be that contextual restrictions are active for colour-type scenarios while genuine ambiguity is at play for book-type nominals – in particular given that book encodes senses of clearly distinct types, while colour can only encode different gradations of the same physical type.

Liebesman and Magidor have another aim: to show that copredication-licensing nominals are not the only linguistic entities that can instantiate semantically distinct properties. By doing so, they claim that this proves that the use of copredication to reject referential semantics is illegitimate. To illustrate their point, they discuss properties that can be instantiated by both events and non-events, such as being ‘tolerable’: A certain condescending explanation can be deemed intolerable (an event reading) by virtue of the mental state of the individual listening to it, while an individual can also be judged as intolerable (a non-event reading) because of their penchant for delivering condescending explanations. As such, ‘properties can be instantiated by different objects in different ways’ (Liebesman and Magidor 2017, 143), and they claim that copredications can be derived from this basic metaphysical fact. Yet, the crucial difference between event and non-event readings of intolerable and event and non-event readings of lunch is that the latter involves the simultaneous instantiation of these readings; one cannot combine both readings of intolerable or attribute them to a single entity (for further discussion, see Hogeweg and Vicente 2020; Viebahn 2020).

3.4. Activation packages

Differing from the claim in Liebesman and Magidor (2017) concerning the non-ambiguity of copredication-licensing nominals, by focusing their attention on the nominal school and its range of sense combinations Ortega-Andrés and Vicente (2019) claim that school constitutes a ‘multiply polysemous word’; i.e. a nominal with a large number of meanings. They generalise from this the broader conclusion that certain institutional copredications involve ‘robust activation packages’ (2019, 1), which are lexical items used to trigger certain conceptual representations. Their explanation for copredication acceptability dynamics is that polysemous nominals constitute activation packages of senses (rather than dot-types) but that ‘senses which are more closely related have higher rates of co-activation’ (Ortega-Andrés and Vicente 2019, 3). As such, their central claim is that copredications like ‘The newspaper fired its editor and fell off the table’ are putatively anomalous because the senses involved are not sufficiently related, and that the institutional sense ‘fails to activate’ the physical sense (2019, 13).

Yet, the claim that the institutional sense of newspaper fails to activate the physical sense cannot account for cases of acceptable copredications like (6a) and (6b), since even though relative clauses and modificational structures seem to boost acceptability (an issue returned to below), this alone cannot account for why the physical sense seems to be activated:

• (6) a. The newspaper that offered me a job can be found on the table.

  b. The town’s most famous newspaper is on the desk.

It seems that the physical sense is not facilitated to the same degree in institution-physical shifts than in information-physical shifts. I will elaborate on this possibility below.

3.5. Neighbouring senses

Copestake and Briscoe’s (1992, 1995) account of copredication constitutes a notable extension of the Generative Lexicon model. These authors discuss forms of polysemy relating ‘two or more senses’ (1995, 15) and forms of polysemy involving a single sense of a lexical entry being underspecified for its meaning and contextually specified. Considering book-type copredications, the authors treat these as involving the selection of an ‘appropriate aspect of the meaning of the complement’ (Copestake and Briscoe 1995, 32). As such, rather than claiming that these types of nominals are ambiguous (exhibiting multiple conventional readings), the authors claim that they are vague between event and object interpretations (Chao 1959).

The authors note that there are certain cases of copredication which are not clear-cut as to their polysemy classification. Newspaper-type copredications are a case in point. Copestake and Briscoe (1995, 54) note that ‘[i]t seems plausible to suggest that newspapers are regarded as (named) institutions in themselves’. This provides a much broader range of sense combinations, and so they claim that certain newspaper-type copredications can readily be classified as single lexical entries having their meaning underspecified (and contextually derived), while others seem to involve relating two distinct senses much more readily without the need for contextual support. Hence, they assume two lexical entries for newspaper: a physical representation and an institutional representation. They motivate this distinction based on a range of putatively degraded cases arising from coordinating both putative structures, and also based on the fact that both the institution and physical senses are counted separately (although see Gotham 2015):

• (7) Three newspapers have been attacked by the opposition and publicly burned by demonstrators.

Lastly, Copestake and Briscoe’s theory of copredication acceptability is also based on whether the two senses are ‘neighbours’ in a continuum from abstract to concrete. To take some examples from their list of 12 sense examples for newspaper, they give the following as interpretations which seem to shift from one sense to another (1995, 54): ‘That newspaper is owned by a trust’ > ‘That newspaper carries long articles about the internal struggles of the Labour Party’ > ‘That newspaper is covered with coffee’. Using this method, with the authors claim that the association of ‘distant’ senses is often unacceptable due to it associating for them two distinct lexical entries. However, the authors do not define how ‘distant’ (in terms of how many senses away from each other; 1, 2, 3 … ?) senses need to be before they yield semantically deviant readings.

3.6. Pragmatics accounts

So far, we have seen that formal semantic approaches to copredication focus on the contribution of lexical representations during composition. In contrast, pragmatic approaches focus on speaker intentions and conversational principles (Nunberg 1978, 2004; Recanati 2010; Ward 2004). Crucially, these are not necessarily rival camps. Semanticists have readily acknowledged the importance that pragmatics can play in improving copredication acceptability: Gotham (2012:, 14) notes that ‘[i]t certainly seems that some kind of pragmatic story needs to be told about … [anomalous] copredication’ (see also Pustejovsky 1991, 410).

Taking the most commonly studied framework, in the pragmatics literature on meaning shifts the following factors have been typically assumed to permit shifts from a property to an individual: The salience of the property identifying the individual; some form of functional correspondence between the property and individual; and contextual support (Jackendoff 1997; Murphy 2016; Nunberg 1995).

In Nunberg (1995), the original conception of predicate meaning transfer is stated: ‘‘Transfers of meaning’ are linguistic mechanisms that make it possible to use the same expression to refer to disjoint sorts of things’. In later work, Brandtner (2008, 2011) maintains that it is always the first predicate that fixes the meaning of the nominal, and that later predicates shift the meaning. Brandtner assumes homonym-like ambiguity for book, such that different homophones of book denote distinct meanings. Brandtner (2008, 30) claims that ‘this mechanism [of predicate meaning transfer] does not act on the assumption that the nominalization has two readings at the same time’. For instance, in ‘Roth is Jewish and widely read’, Brandtner (2008) argues that we are accessing two distinct properties of the notion person, but that since we do not have to shift the nominalization due to its singular meaning, only the first predicate’s meaning is attributed to the nominal representation. Upon encountering the second predicate, the meaning of the nominal is shifted (such that we access the second lexical representation), yielding: ‘Roth₁ is Jewish and [Roth₂] widely read’. Lastly, Nunberg (1995, 112, 2004) states that meaning transfer is only permitted if there is a salient relationship between predicates and arguments, as in ‘I am parked out back’, where a salient relationship of ownership exists between the speaker and the car (see also Elbourne 2005; Rothschild 2011 for a related perspective from dynamic semantics).⁶

Brandtner (2009, 2011) proposes that the semantic salience and coherence between predicates and arguments in copredications heavily factor in their acceptability. She argues that the establishment of copredication cannot be reduced to semantic principles, claiming that event and result readings can only be coordinated when a salient relation obtains between them. For instance, Brandtner claims that the copredications yielded by German deverbal –ung nominalization cannot be addressed by compositional operations alone, but additionally require pragmatic processes. In the case of Übersetzung (‘translation’), the selectional restrictions of two conflicting interpretations (event and result) simultaneously apply to one token of the nominal:

• (8) Die [langwierige]ev Übersetzung [verkaufte sich millionenfach]re.

  The tedious translation sold themselves million-fold

  ‘The tedious translation sold million-fold.’

Certain cases in which the predicates are unexpected (in a context in which ‘easy’ translations are not expected to sell many copies, and ‘tedious’ translations are expected to sell well) are argued by Brandtner to make copredication licensing more difficult:

• (9) Die [einfache]ev Übersetzung [verkaufte sich millionenfach]re.

  The easy translation sold themselves million-fold

  ‘The easy translation sold million-fold.’

Norrick (1981, 115) also claims that slight alterations to copredications involving temporal markers impact their acceptability:

• (10) a. #Judy’s dissertation is thought provoking and yellowed with age.

  b. Judy’s dissertation is still thought provoking though yellowed with age.

As such, these accounts help us understand the ways in which copredication licensing interacts with context, and I will expand on these insights into causal and salient predicate relations below.

4. Evaluating copredication

Thus far, I have delimited the definition of copredication and presented a selective overview of semantic, psychological and pragmatic accounts. This has allowed me to discuss a number of intriguing and unexplained features of copredication (e.g. patterns in anomalous interpretations). In this section, I will discuss the two main themes that emerged from the above review of theoretical accounts: predicate ordering and predicate coherence. The motivation behind this is that these will be the two main factors claimed here to be responsible for copredication acceptability dynamics. They have only recently been empirically investigated in relation to copredication (Murphy 2021), and are yet to be accounted for within present theories.

4.1. Enumerating senses in copredication

A topic I have yet to address concerns the number of senses involved in possible copredications. I noted in the previous section that only the institution, event and physical senses are needed to derive interpretations of institutional copredications. Vicente (2018:, 955) has noted that, ‘[i]n principle, there is no limitation as to the number of aspects that constitute a dot object’ (see also Falkum and Vicente 2015), while Kinoshita, Mineshima, and Bekki (2018, 156) note that copredications involve ‘two or more’ predicates. Recall that Copestake and Briscoe’s (1995) model is compatible with n-senses being assumed for nominals involved in copredication. Ortega-Andrés and Vicente (2019, 3) have also observed how copredication ‘can involve a larger number of senses’ than two, citing the following, which invokes institutional, physical, and nation-related cultural meanings:

• (11) Brazil is a large Portuguese-speaking republic that is very high in inequality rankings but always first in the FIFA ranking.

The evidence presented so far seems to necessitate this n-sense framework, while acknowledging that certain complex polysemous nominals licensing copredications will only host two senses (e.g. lunch), depending on the nominal in question and its apparent conceptual address.⁷

Nevertheless, while we can agree with Vicente (2018) that ‘there is no limitation as to the number of aspects that constitute a dot object’, the semantics literature contains some claims which make this principled conclusion less likely to be arrived at by the outside observer. For instance, in Pustejovsky and Jezek (2008:, 186) it is claimed that dot-types ‘have a symmetric internal structure consisting of two types clustered together by the type construction • (‘dot’), which reifies the two elements into a new type’. On the other hand, more recently Pustejovsky and Batiukova (2019, 176) have claimed that location • human group • organization dot-types (e.g. university, city) can be motivated. Further, although Asher’s (2011, 2015) work strictly adopts the idea of ‘dual aspect nouns’ for complex polysemy, he discusses ‘nouns that involve a location, a physical object, and an institutional aspect like banks or cities’ (2011, 319).

How can we relate these observations to the current literature? The semantic equivalent of Ortega-Andrés and Vicente’s (2019) ‘activation packages’ which could also accommodate n-sense copredication might take the form of a cluster of aspects/senses which can be related by dot-object conjunction. Call this an elliptic-type, composed of •-relatable senses, with no stipulated restrictions as to the number of possible sense conjunctions (so-called due to the nature of the unbounded array of senses being composed, ‘ … ’). In terms of their structure, elliptic-types would be similar to what Arapinis and Vieu (2015) refer to as ‘complex categories’, adding additional senses to a standard dot-type. As such, we can assume that they might be implemented via a form of meaning modification over a single lexical entry as in Asher (2015), such that newspaper(information) shifts to newspaper(information • institution), rather than newspaper₁ shifting to newspaper₂. Without pursuing the formal issues here, we can at least surmise that the structure of elliptic-types can be implemented via a revised Modern Type Theory or a Pietroskian conjunctivist semantics (Pietroski 2018).

These ideas are compatible with the idea that individual lexical items are instructions (‘recipes’) to build complex concepts (Pietroski 2018), with a simple nominal like lunch being able to call upon/address various subsenses – with these concepts in turn functioning as ‘generative pointers’ or ‘unstructured symbols that point to memory locations where cognitively useful bodies of information are stored and can be deployed to resolve polysemy’ (Quilty-Dunn 2020, 158).

Furthermore, these assumptions pertaining to n-sense copredications allow us to accommodate structures of the following kind without the need to radically modify type-theoretic accounts, invoke linear dot-types, or propose multiple lexical entries. Instead, an elliptic-type could generate a cluster of senses related to a single lexical entry:

• (12) The well-written newspaper that I held this morning has been sued by the government.

This supports the claim that newspaper copredications are not simply metaphors extending over institutional readings – as argued by Bahramian, Nematollahi, and Sabry (2017) – since semantically distinct senses can be attributed to newspaper without relative loss of acceptability compared to supposedly non-metaphorical cases of genuine polysemy.

4.2. Predicate ordering

This section will explore in greater detail the apparent effects of predicate ordering on copredication acceptability, evaluating certain claims presented in Section 3.

To begin, consider how Gotham (2015:, 140) notes that (13a) is infelicitous. Yet notice that reversing the predicate order of Gotham’s example seems to alter acceptability in (13b):

• (13) a. #The bank is FTSE-100 listed and used to be a police station.

  b. The bank used to be a police station and is FTSE-100 listed.

As this section will discuss, effects of predicate ordering on copredication acceptability have been noted in the literature but not systematized; that is to say, little is known about what we are re-ordering.

Consider how translation can refer to a process or a physical text (Bisetto and Melloni 2007), yet while (14a) is well-formed, (14b), which reverses the predicate order, seems to differ in acceptability, as claimed by Brandtner (2009):

• (14) a. The translation lies on the table and was difficult.

  b. #The translation was difficult and lies on the table.

Brandtner (2009) claims that this example suggests that ‘structural aspects’ impact copredication acceptability. But which aspects of ‘structure’, exactly?

As with translation, other derived nominals can permit copredication (Asher and Denis 2005, 15):

• (15) a. The reproduction (of the painting) took place in that workshop and [was/is] eight feet tall.

  b. #The reproduction (of the painting) [was/is] eight feet tall and took place in that workshop.

This phenomenon is not currently well understood. For instance, Kallmeyer and Osswald (2017) note that ‘there seem to be cases where a predication over one aspect blocks the other meaning aspects for further access’ (a claim similar to Chatzikyriakidis and Luo’s linear dot-type explanation), but no specific explanation has been presented.

Relatedly, coercion seems to operate unidirectionally in certain cases: (16a) is a case of coercing a physical referent into an eventive reading, but one cannot easily shift an eventive referent into a physical reading (Liebesman and Magidor 2017, 150–151):

• (16) a. That dog bone lasted two minutes.

  b. #The lunch meeting was delicious, but lasted hours.

Note that the opposite acceptability pattern obtains when an event is combined with informational content. While event-to-physical coercion seems anomalous and physical-to-event is acceptable, the examples below suggest that event-to-information is anomalous and information-to-event is acceptable; the example in (17a) is taken from Asher (2011, 132), but as discussed here, sense reversal is rarely conducted and so (17b) is not considered by Asher:

• (17) a. Mary’s testimony occurred yesterday and [was/is] riddled with factual errors and inconsistencies.

  b. #Mary’s testimony [is/was] riddled with factual errors and inconsistencies and occurred yesterday.

These examples seem to further illustrate that re-ordering senses (along an axis of semantic complexity, as will be argued below) can impact semantic acceptability.

The sense order effects documented here may also bear some principled relation to the observation that quantificational NPs do not usually conjoin with names to form complex phrases, as Liebesman and Magidor (2017, 141) observe:

• (18) a. #Somebody and Emma are barking.

  b. Oli and Emma are barking.

Coordinating outside categories, as in (18a), seems much less acceptable than coordinating within categories, as in (18b). Yet notice that sense reversal of (18a) seems to improve acceptability:

• (19) Emma and somebody (else) are barking.

It is as if, with copredication, that once the more concrete or simple sense (Emma, a definite and semantically specific token) is introduced, the more abstract or complex sense (somebody, an indeterminate token) can more readily be conjoined.⁸

Lastly, consider how cities are composed of town (i.e. government/polity or people) and physical (specifically, building/physical artefact, being intentionally constructed) senses.⁹ Asher (2011, 63) notes that the order in which these senses are combined impacts acceptability:¹⁰

• (20) a. The city has 500,000 inhabitants and outlawed smoking in bars last year.

  b. #The city outlawed smoking in bars last year and has 500,000 inhabitants.

Asher (2011, 64) notes that this indicates that copredications are subject to ‘discourse effects’. However, the precise discourse factors are not well understood. It should also be noted that ‘category mistakes for the most part must be understood relative to a background, contextually supplied set of types’ (Asher 2011, 9), which may confound certain contrasts and can explain the difference below:

• (21) a. #The number two could have been red.

  b. If numbers were physical objects, then the number two could have been red.

Nevertheless, the factor of sense order may be inadequate to explain the full range of copredication acceptability dynamics. Of crucial importance is the observation that Asher’s example seems to be improved by adding some form of clear, motivated relationship between the predicates, as in ‘The city outlawed smoking in bars last year and has 500,000 smokers’ or ‘The city outlawed smoking in bars last year but this will affect few of its 500,000 inhabitants’.¹¹ This idea is in line with the notion of coherence relations in discourse interpretation (Hobbs 1979; Kehler 2002; Sanders, Spooren, and Noordman 1992), through which explicit semantic features are held in common between each predicate (i.e. the concept smoking is referenced in both). This is the issue I will turn to next.

4.3. Coherence relations

4.3.1. Introducing coherence

Though initially applied to text comprehension across a series of utterances, coherence relations have more recently been shown to be relevant for intra-sentential relations. Coherence relations are ‘the cornerstone of comprehension’ (Graesser, McNamara, and Louwerse 2003, 82). De Beaugrande and Dressler (1981: vi) claim that part of coherence is found in a ‘continuity of senses’. Vakulenko et al. (2018) adopt a more specific definition, where coherence is the ability to ‘perceive meaningful relations’ between concepts.

As Kehler (2002) emphasises, the importance of coherence in language has very often been sidelined by research focusing on syntactic and semantic factors. Yet the seemingly unavoidable importance comprehenders place on coherence can be seen in the following contrast (Kehler 2002), where comprehenders seek some form of connection between distinct sentences:

• (22) a. John took a train from Paris to Istanbul. He has family there.

  b. #John took a train from Paris to Istanbul. He likes spinach.

Kehler (2002) goes on to argue that coherence relations can be used to explore a range of linguistic phenomena, including VP-ellipsis, gapping, and pronominal and tense interpretation. In this section, I will argue that coherence relations also impact the acceptability of copredications.

Recall how Brandtner (2009) and Norrick (1981) noted that creating a more salient relation (e.g. a salient temporal relation via an adverbial) between predicates appears to improve the felicity of copredication:

• (23) a. #The newspaper was founded in 1878 and is typed in Frankfurt.

  b. The newspaper was founded in 1878 and is still typed in Frankfurt.

Relatedly, Schumacher (2013, 2) notes that the most standard copredication test, involving coordinate structures, is potentially problematic since other linguistic factors can interact with it: ‘[O]ther criteria that are independent of the meaning shift may be responsible for the failed copredication or coordination such as coherence, morphosyntax or the combination of incompatible predicates’ (emphasis added). However, no robust definitions of coherence have been given, and my goal here is to draw a clearer conception of coherence in order to establish more concretely how pragmatic factors interact with linguistic factors in the establishment of copredications.

4.3.2. A coherence model for copredication

Coherence relations are ‘fundamentally conceptual’ (Kamalski, Sanders, and Lentz 2008, 324) in nature, but can be made explicit via certain linguistic devices and markers (Knott and Sanders 1998). There are two main types of coherence relations that I will argue are applicable to copredication.

The first type is causal connection (e.g. problem-solution, cause–consequence). In the literature, this has been referred to as relational coherence (Spooren and Sanders 2008). These relations are readily applicable to the predicates involved in copredication, and have in fact been used widely in the literature, often without the factor of coherence being discussed explicitly (e.g. compare ‘The newspaper was sued [and was]/[for being] offensive’, with the case of ‘for being’ seeming more acceptable and easier to infer a causal relation from). The most common surface cues of coherence relations are cue phrases (e.g. ‘by comparison’) and causal connectives (Degand 1998; Hoek 2018; Mann and Thompson 1992; Sanders 1997), realised often via conjunctions like ‘because’ and ‘so’ or adverbs like ‘therefore’. However, causal relations can readily be inferred without these cue connectives, and can be inferred from simple conjunction, as in ‘John was fired and was unhappy’. Causal connections can readily be apparent in copredications: cause–consequence in ‘The school had unruly students and lots of graffiti’; problem-solution in ‘The city was polluted and was investing in green energy’.¹² We can define causal connections along pre-existing axes defined in the literature: forms of semantic, external or subject matter coherence (defined as part of the relational domain; Degand 1998; Kamalski, Sanders, and Lentz 2008). In addition, constructions such as ‘The newspaper went bust yesterday and was very poorly written’ violate Chatzikyriakidis and Luo’s linear dot-type model, yet a clear (possible) causal relation, or at least a thematically related second predicate, licenses the copredication.

The second type of coherence relation relevant to our concerns does not have a general name. I will call it extensional overlap. With respect to copredication, extensional overlap involves two predicates denoting a shared feature/concept. For instance, consider a case we have already discussed: ‘#The city outlawed smoking in bars and has 500,000 inhabitants’. This exhibits no extensional overlap, since the organisation and the population are not the same entities (as assumed in Lefeuvre-Halftermeyer, Moot, and Retore 2019; Vicente 2015, 2017, 2018 and much other work exploring copredication sense types). In comparison, consider ‘The city outlawed smoking in bars and has 500,000 smokers’. Here, the organisation and the sub-population who smoke are also not the same entities – yet the acceptability is improved. Under the account I am pursuing here, the construction exhibits clear extensional overlap in the shared salient feature/concept of smoking. This account may also be able to be framed in terms of situation semantics (Elbourne 2005), whereby it is by virtue of an event hosting a particular shared feature/concept that permits acceptability increases. Extensional overlaps can be realised via simple conjunctions like ‘and’ or ‘but’, however since ‘and’ is the most commonly discussed type in the literature I will be focusing on this much more than other conjunctions.¹³ We can define extensional overlaps as a form of pragmatic (Sanders, Spooren, and Noordman 1992; Van Dijk 1977), internal (Martin 1992) or presentational (Mann and Thompson 1988) coherence (defined as part of the discourse domain). Although I will not interrogate the notion much further here (partly because it is ultimately an empirical notion), the question of what exactly constitutes the type of features/concepts used in extensional overlaps is an entirely open one.

Examining this notion further, consider the distinction in Degand (1998) between connectives exhibiting completely overlapping, partially overlapping or non-overlapping meanings.¹⁴ This was originally used by Degand (1998) to denote the degree of semantic overlap between distinct connectives (‘and’, ‘but’, …) and the corresponding degree of semantic acceptability generated by the sentences they are embedded in when substituted for one another (i.e. take sentence X with connective Y, substitute the connective for connective Z and compare the derived meanings of both sentences). Applying this distinction to connectives in copredication, this generates the following three possibilities for extensional overlaps:

• (24) The government outlawed smoking in bars and has

  a. 500,000 smokers [overlap]

  b. 500,000 inhabitants [partial overlap]

  c. five skyscrapers [non-overlap]

There is a complete overlap in the concept smoking for (24a). With respect to the partial overlap in (24b), this exists insofar as some inhabitants are assumed to be smokers. Lastly, there is no extensional overlap between the government outlawing smoking and the existence of five skyscrapers in (24c).¹⁵ The reason why these distinctions are important for our purposes is that copredication seems to increase in acceptability as we move from non-overlapping through to overlapping. These ideas seem supported by Haber and Poesio’s (2020:, 114) experimental work, which showed that ‘sense similarity appears to be a major contributor in determining co-predication acceptability’.

Vicente (2019) notes that ‘we are very far from having … an account’ for why the acceptability contrast in (25) exists. Assuming that copredications require a coherent spatiotemporal frame between predicates, this contrast seems puzzling since one would expect (25b) to be just as unacceptable as (25a).

• (25) a. #The school caught fire when it was on excursion.

  b. The school in New York caught fire when it was celebrating 4th of July in Chicago.

Following the account presented here, the clear coherence relation between the predicates in (25b) (namely, extensional overlap, with the overlapping feature being location-related information) licenses the copredication. To repeat this point in a slightly different manner: If another school physically located in Houston organised a trip to New York, it would be odd to say (26), where the people/group are visiting New York and the physical building remains in Houston – chiefly because no coherence relations are present.

• (26) The school in New York caught fire.

In sum, the model of coherence relations I claim to be relevant for copredication can be found in Table 1.

Table 1. Model of coherence relations taken to be relevant for copredication licensing.

	RELATIONAL COHERENCE	DISCOURSE COHERENCE
RELATION TYPE	CAUSAL CONNECTION	EXTENSIONAL OVERLAP
COPREDICATION EXAMPLE	‘The newspaper was sued because it was offensive.’	‘The newspaper purchased more ink and had five colourful pages.’

4.4. A Copredication hierarchy

In this article, I have reviewed a relevant selection of copredication theories, as well as some literature from neighbouring domains in linguistics, and proposed that predicate ordering and coherence relations play a major role in modulating copredication acceptability. As mentioned, there is currently little in the way of a systematic method for classifying types of copredication. As such, I will here attempt to provide such a classification.

Copredication appears to have a high range of diversity and semantic complexity (see Hoffman, Lambon Ralph, and Rogers 2013 for related discussion). To formalise this, I will stipulate five types of copredication, of increasing semantic complexity. Certain simpler senses seem to be monadic, i.e. cannot be decomposed, like location/physical.¹⁶ This contrasts with institution, which has conjoined people and polity senses:

Copredication Hierarchy

Type I: monadic • monadic

Type II: monadic • conjoined

Type III: conjoined • conjoined

Type IV: monadic • monadic • conjoined

Type V: monadic • conjoined • conjoined

This yields a simple, and purely conjunctivist account of the varieties of copredication: Type I copredications involve book-type nominals containing two simple physical and informational senses. physical object • aperture copredications like ‘John painted the door green and walked through it’ are also of Type I, which refer to the object but also the space associated with it (e.g. window, fireplace, room; Pustejovsky and Boguraev 1993).¹⁷ Type II copredications include lecture and film, which combine the simple information and complex event senses (‘The boring lecture lasted nearly an hour’).¹⁸ Type II copredications also involve lunch and construction cases, with the event sense being decomposable into sequentially structured subevents (such as states, processes and transitions; Pustejovsky 1991, 420). event • institution copredications, such as ‘The church starts at 10:30am and has a professional choir’, are of Type III, as are city-type copredications. Type IV copredications involve a dot-type being conjoined to an additional monadic sense, and are of the newspaper (information, physical, institution) and translation (information, physical, event) varieties.

This type system is inherently dynamic. For instance, school and church can instantiate a higher type copredication (Type V), as in ‘The beautiful church starts at 10:30am and has a professional choir’, selecting for a monadic physical sense, while lecture-type nominals can instantiate Type IV copredications and be ‘loud’ or ‘quiet’, hence hosting a physical quality as well as the information • event composite.

The Copredication Hierarchy relies on the notion of semantic complexity. Since there is currently no widely agreed upon definition or framework for the notion of semantic complexity, it will here be defined in terms of the number of cognitive modules, or ‘core knowledge systems’ (e.g. number sense, agents and their actions, basic geometric reasoning; Carey 2009), a given sense draws upon. Spelke (2016), Strictland (2017) and others have proposed that various innate core knowledge systems create new concepts during development through novel compositions of these elementary modes of structuring conceptual knowledge. This includes knowledge about place (Spelke 2016), number (Hyde and Spelke 2011), social relations (Spelke and Kinzler 2007), geometrical properties (Dillon, Huang, and Spelke 2013) and the behaviour of physical objects (Baillargeon 2001). For instance, the physical sense of newspaper only draws on basic geometric reasoning (e.g. flat, smooth surface), while the institution sense of newspaper draws on social reasoning (Fiske and Taylor 2013), network dynamics and intentionality (e.g. agents working to a common end). Adopting standard assumptions in the literature (Asher 2011; Buitelaar 2000; McCaughren 2009), we can assume a hierarchy of complexity for the four most common sense types, which also adheres to this general perspective (‘<’ = ‘less semantically complex than’):

Complexity Hierarchy

physical < information < event < institution

To illustrate further, while both senses in book are monadic, they differ in semantic complexity, with information potentially ranging across a number of domains. The abstract content of a book is a more complex semantic construct than its concrete features, for reasons explored in much psycholinguistic work. For instance, it is known that abstract concepts such as idea are more difficult to process than concrete concepts such as table; a well-known phenomenon termed the concreteness effect which applies across a variety of tasks such as memory retrieval, comprehension and lexical decision (Barber et al. 2013; Jessen et al. 2000). Wiemer-Hastings and Xu (2005) tested the hypothesis that abstract concepts are more complex than concrete concepts because they contain a larger variety of features/concepts (similar to my assumption that complexity scales with cognitive module variety). They discovered that abstract concepts had a greater number of properties relating to subjective experiences than concrete concepts, and also had more relational properties than concrete concepts. Given that polysemous senses are typically assumed to be conceptual representations of some form (Pustejovsky and Batiukova 2019; Quilty-Dunn 2020), this seems an appropriate measure of complexity.

4.5. Incremental semantic complexity

As a way of proposing an alternative perspective for the acceptability contrasts presented in this article, I will now suggest that the effects of sense order may be modulated by the semantic complexity of polysemous senses.

The only current work which has empirically tested the effects of predicate ordering on copredication acceptability is Murphy’s (2019, 2021) experiments, which found that copredicated sentences exhibiting a Concrete-Abstract predicate order (Murphy 2019), and also a Simple–Complex order (Murphy 2021), were significantly more acceptable than sentences exhibiting an Abstract-Concrete and Complex–Simple order (see also Murphy 2017). This seems to match the general acceptability contrasts reviewed throughout this article. While further empirical research is needed to refine our understanding of copredication acceptability, the common theme of the material reviewed here appears to motivate the claim that the parser is sensitive to a specific bias, called Incremental Semantic Complexity in Murphy (2021):

Incremental Semantic Complexity (ISC)

Seek to process linguistic representations in incremental stages of semantic complexity.

The acceptability contrasts reviewed here suggest that the parser appears to prefer semantically more complex objects to be accessed after less complex ones.

ISC may indeed be a general principle given that a number of other, unrelated constructions seem to exhibit parsing preferences such placing the more complex object first results in reduced acceptability. Manipulating Number is a case in point; (27b) below is acceptable in a particular context, but seems somewhat more difficult to parse than (27a), which does not need any contextual assistance:

• (27) a. I saw a picture of us.

  b. We saw a picture of me.

It may also be worth following the lead of Tversky and Gati’s (1978) work on similarity relations. These authors pointed out that to say that a toy train is like a real train is much more acceptable than to say that a real train is like a toy train, because most features of the toy train are contained within the real train, but the toy train only has a small number of features exhibited by the real train. This effect may influence the acceptability of some of the copredications discussed. For instance, to think about the populace sense of city invites related intuitions about its institutional structure, yet when prompted about the institution sense of city it seems less likely that any intuitions about its populace would be triggered. The same observations apply to book-type nominals, whereby speaking of the information sense of a particular book does not trigger its physical sense as easily as when the order is reversed. This may yield an acceptability disadvantage to Abstract-Concrete or Complex–Simple sense orders. Notice, however, that this Tversky and Gati-inspired account is approximately identical to a complexity-based one: The notions of similarity which Tversky and Gati rely on crucially correlate with semantic complexity, and so there may be no principled way to differentiate the two accounts.

4.5.1. Complexity effects

In most approaches to abstract concepts in the semantics literature, there appears to be an agreement that increasing ‘abstractness’ is defined in terms of boundedness to features of the denoted situation, such as physical features or the degree of boundedness to space, time and agents (Asher 2000; Consten and Knees 2008; Maienborn 2003; Schumacher, Consten, and Knees 2010). Propositions, for instance, are typically considered to be the most abstract structures since they are not strictly tied to one particular world or state of affairs (unlike facts).

There seems to be a linear relation of increasing abstractness exhibited by the physical, information, event and institution senses investigated in this article. The information sense of polysemous nominals is considerably bound by physical features, though certainly not entirely. The event sense seems less tied to physical features, though it is usually characterised in some way by these physical features. Lastly, the institution sense is very typically not bounded by physical features (banks can relocate and act independently of their headquarters), and so this seems to be the most abstract of the senses permitting copredication (intuitions already defended in Arapinis 2013). Assuming a particular abstractness scale (Event < Process < State < Fact < Proposition; cf. Asher 1993; Maienborn 2003), Schumacher, Consten, and Knees (2010, 842) note that ‘[s]ince events are defined as spatio-temporal entities with specified beginnings and endings as well as certain agents, they have a high degree of boundedness and are the ‘least abstract’ complex entities’ (the authors do not discuss informational aspects, however).

In further work which relates to ISC and its core predictions, Consten, Knees, and Schwarz-Friesel (2007) conducted corpus analyses of ontological shifts across discourses and their apparent constraints. For example, they looked at sentences when states are spoken of and are subsequently referred to using anaphoric phrases (namely, complex anaphors) denoting events. They provide evidence that ontological changes in discourse are restricted by what they call the Abstractness Constraint. This states that the complex anaphor cannot shift the type of the referent to a discourse entity that is less abstract than the type assigned by the antecedent:

Abstractness Constraint

*x ← y if x > y (if x is higher on the abstractness scale than y, where x represents the ontological type assigned by the antecedent and y that of the anaphor).

For example, a process anaphor cannot pick up a proposition antecedent. As such, an event anaphor (‘this event’) cannot be used to refer to a more abstract referent (p ← *e). For example (where [E] = event, [P] = process, [S] = state, [F] = fact, [PP] = proposition):

Bethesda is developing the next Elder Scrolls title _[P]. This _[P]/This process _[P]/ This state _[S] will presumably last for five years. This fact _[F] is well-known to fans. Until recently journalists were not allowed to verify this possibility _[PP]/*this event _[E].

This constraint appears to arise because ontological features that are not specified by the antecedent are unable to be reconstructed by the anaphor. Along with the Abstractness Constraint seeming to match a range of felicity judgements, Consten, Knees, and Schwarz-Friesel (2007) examined a list of 60 complex anaphors taken from the TIGerKorpus (a German newspaper corpus) and found no violations of the constraint (see also Consten, Knees, and Schwarz-Friesel 2009). Supporting these intuitions of acceptability, Schumacher, Consten, and Knees (2010) show that violations of the Abstractness Constraint elicit a pronounced N400 (a negative-going waveform peaking approximately 400 ms post-stimulus associated with memory retrieval and surprisal) relative to maintaining the ontological type of the antecedent via a complex anaphor (for electrophysiological basis of higher-order semantics, see Benítez-Burraco and Murphy 2019; Murphy 2018, 2020).

This idea of abstractness utilised by the Abstractness Constraint may be related to the notion of semantic complexity in ISC. The general ban on shifting from a certain abstract type to a less abstract type (i.e. using an event complex anaphor to refer back to a state) seems to mirror the general degradation in acceptability found with shifting from an abstract, or complex polysemous sense to a concrete, or less complex polysemous sense. Both constraints, being applicable at the discourse level, might be caused by the same underlying processing bias.

4.5.2. From priming to copredication

This section will explore how Incremental Semantic Complexity (ISC) could be implemented via a priming effect for complex senses, such that the activation of Simple Sense₁ primes Complex Sense₂, and the activation of a complex sense does not seem to prime (as strongly) the simpler sense(s).

I have claimed that ISC can account for copredication acceptability dynamics and anomalous interpretations noted in the literature, and should be favoured as an account for the stipulation of novel semantic types or the postulation of multiple lexical representations for complex polysemous nominals. It is already known that polysemous senses prime one another (Klepousniotou, Titone, and Romero 2008), and so I am effectively adding to this general finding the more specific proposal of (i) how this priming is realised between semantically distinct senses, and (ii) what scale it is based on (i.e. complexity). This proposal contrasts somewhat with Ortega-Andrés and Vicente’s (2019) claim that all senses are activated as an ‘activation package’ upon encountering the nominal, since the acceptability contrasts presented here demand a more refined perspective on sense activation. As such, I depart from Ortega-Andrés and Vicente’s (2019) claim that institutional senses ‘fail to activate’ non-institutional (physical) senses; rather, the more cognitively plausible priming model defended in this article maintains that the institutional sense simply does not co-activate the physical sense as much as the physical sense would prime the institutional sense.

There are arguably three distinct processes involved in this form of priming (Heyman et al. 2015; Rohaut et al. 2016):

1. Semantic activation of the prime.

2. Spreading activation from prime to target.

3. Activation of the target.

The core differences between Simple–Complex and Complex–Simple predicate orderings may reside in (2), with Complex–Simple orderings involving a reduced activation spread from the Complex to the Simple sense. Priming research in psychology increasingly involves an ever-expanding list of ‘prime-able’ representations (Bargh 2006). Indeed, there are other ways to implement ISC; it may not be that Simple–Complex ordering facilitates/increases acceptability, but rather that Complex–Simple orderings deliver some form of processing cost. Further empirical work is needed to address this.

4.5.3. Additional structural factors

The present account might shed some light on something previous work has been unable to explain; namely, the peculiar facts about relative clauses (‘The newspaper John is reading is being sued by Mary’) and modificational structures (‘The most provocative newspaper of the year has been sued by the government’). These structures often increase the acceptability of copredication in comparison to coordinated structures (Chatzikyriakidis and Luo 2015). It is possible that these structures boost the coherence of the predicates, in particular with modificational structures since these provide additional semantic information and context. Relatedly, the same explanation can be extended to the observation above that the inclusion of temporal adverbials like still and already can transform an unacceptable copredication into an acceptable one (Brandtner 2009), and that the inclusion of some narrative-based explanation for the copredication can drastically improve acceptability. The inclusion of temporal adverbials likely serves to increase the coherence of the predicates in cases in which the adverbial occurs in one predicate and the other predicate refers to an event or process, thus serving to increase discourse coherence.

Nevertheless, the acceptability improvements yielded by relative clauses and modificational structures remain somewhat puzzling. While an explanation involving predicate ordering and coherence seems reasonable, this account is admittedly stronger for modificational structures than for relative clauses (for instance, why should relative clauses aid to strengthen coherence relations?). Ortega-Andrés and Vicente (2019, 12) suggest that acceptability differences yielded by relative clauses in copredication might come about because the relative clause ‘may have the effect of changing the expectations/predictions’ about upcoming senses.

There may be other explanations for these effects concerning relative clauses and copredication, grounded in assumptions from the coherence relations literature. Patterson and Kehler (2013) maintain that certain factors can influence the overt linguistic marking of coherence relations, such as their syntactic position. They compose a model which marks whether the coherence relation is embedded inside another relation (Rel2), contains another relation (Rel1) or shares a segment (or for copredications, a predicate) with another relation (S = segment):

Containing (Rel1)  + Embedded (Rel2) Relation

[S₁] ← Rel1 → [S₂ [S₁ ← Rel2 → S₂]]

Shared Segment

[S₁] ← Rel1 → [S₂ S₁] ← Rel2 → [S₂]

This allows us to take into account the syntactic dependency in a given coherence relation. As such, copredications involving relative clauses exhibit embedded coherence relations. These ideas echo Taboada’s (2009) syntactic division between paratactic (coordinate) and hypotactic (subordinate) coherence relations. One question which recent research on coherence relations (Hoek 2018; Patterson and Kehler 2013) has not addressed is whether syntactically embedded coherence relations are easier to process, or are more acceptable, than other types of coherence relations. This research direction would bear on the present speculations about copredications involving relative clauses.

5. Conclusion

I have argued that copredication acceptability is modulated by predicate ordering and coherence relations, where the former exhibits a Simple–Complex predicate order preference (in terms of semantic complexity), and the latter exhibits sensitivity to the presence of a causal relation or featural overlap between predicates. Summarising current inquiry, Haber and Poesio (2020:, 115) claim that ‘the prevailing understanding of co-predication is that it is rendered felicitous if the different sense interpretations are activated simultaneously and can be shifted between without additional processing costs’. The present article has argued for a modification to this consensus, such that while the degree of coherence and co-activation between predicates is important, a separate factor of predicate ordering can also be isolated, yielding directionality predictions for acceptability dynamics. The theoretical and empirical evidence reviewed in this article can be used to support a generalisation about copredication, which we will term the Copredication Licensing Effect (Murphy 2021):

Copredication Licensing Effect (CLE)

Predicate ordering is a major factor in copredication acceptability, and high levels of coherence between predicates increases copredication acceptability.

While the CLE is not inherently incompatible with certain theories reviewed in this article, it does crucially depart from some others. For instance, Brandtner’s (2008, 2009, 2011) claims about meaning shifts at the second predicate make no recourse to ordering factors, even though she readily acknowledges the role of coherence. Emphasising the clear impact that sense order has on copredication acceptability, and grounding this in a parsing bias (Incremental Semantic Complexity) and elliptic-types, potentially renders certain mechanisms in the literature redundant. As such, it is possible that CLE is tapping into more generalisable parsing processes (e.g. priming), and also more general representational features of copredications, than other more domain-specific semantic, pragmatic and psychological theories in the current literature.

Acknowledgements

My thanks go to two reviewers for comments on an earlier draft, and to the editor for helpful guidance. For very helpful discussions about copredication, I also thank John Collins, Steven Frisson, Nathan Klinedinst, Daniel Rothschild and Agustín Vicente. My sincere thanks also go to the Department of Linguistics at UCL for supporting my doctoral research, and various other faculty members and students who have engaged with me in developing the ideas in this article over numerous conversations and reading groups. Lastly, my biggest thanks go to Wing-Yee Chow and Andrew Nevins for their kindness, patience and regular encouragement during my time at UCL. This work was supported by an Economic and Social Research Council scholarship (#1474910).

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by an Economic and Social Research Council (ESRC) scholarship [grant number 1474910].

Notes

1 This is actually not true. It has become something of a clichéd, obligatory statement at the beginning of papers concerning copredication for the author(s) to motivate their investigation by citing the ‘neglected’ and ‘sidelined’ nature of the topic, to such an extent that by now (mid-2021) there are, paradoxically, a considerably large number of theoretical and empirical papers about this sorely neglected issue that all simultaneously cite this motivation.

2 Note that all acceptability contrasts have been checked with multiple native English speakers, and elsewhere extensive acceptability judgement surveys on copredication have been carried out (Murphy 2019, 2021).

3 Via Modern Type Theory, Xue and Luo (2012) assume that the three most common types involved in copredication are event, physical and information (see also Luo 2012a, 2012b).

4 The semantically odd nature of copredication seems to be exposed by the fact that question-answers like ‘What is high in inequality and also very good at football? Brazil’ come across as pun-like, as if the distinct senses are better suited to a joke than a true statement about the world. The notion of violating congruency is crucial here, as it is in many types of comedy.

5 Matt Hancock, Secretary of State for Health and Social Care in Boris Johnson’s UK government, announced on May 14th 2021: ‘Today I was at West Suffolk Hospital, one of the 40 new hospitals we’re going to build’. I typically do not agree with his announcements, but he is here exposing something interesting about the institutional-physical complex hospital: Hancock can be at a place that does not exist, and indeed he can go inside a place representating a larger group of future entities, shifting tense to accord with the appropriate sense shift (he was at a place that he’s going to build).

6 Consider also cases of verbal ellipsis. In ‘Your shoes are too loosely tied; so is your bowtie’, the predicate is not shifted such that the property of having untied laces is attributed to the bowtie. Rather, the property of the shoe being untied is transferred. The predicate is not strictly speaking shifted, but rather it ‘exploits the inner-structure of the object it applies to in order to highlight a piece of it’ (Arapinis 2015, 9).

7 It also seems to speak directly against other recent claims from Chatzikyriakidis and Luo (2018:, 137), where they maintain that ‘only two of the three senses can appear together’ in newspaper copredications, and claim that ‘there are no cases where all three senses are coordinated nor [do] we know of any other common noun that allows this kind of situation’ (Ibid.). Nevertheless, the authors also note that ‘in principle’ a binary dot-object theory ‘could easily be generalized for n senses if there is a need to do so’ (2018, 156).

8 Boye (2007) argues, from the perspective of cognitive linguistics, that ‘specific meanings … [are] conceptually more complex than non-specific meanings’, since, following Haspelmath (1997), holding a specific meaning requires maintaining two ‘mental space referents’; i.e. maintaining the specific instantiation along with the more general category, as opposed to simply maintaining a representation of the general category.

9 The centrality of the physical artefact sense becomes clear when we think of fantastical cities which were not intentionally constructed by a given population (e.g. a population of elves or dwarves), but rather the population settled in a given location which is naturally designed in a complex way, quite independent of the actions of the population. In this case, it would be peculiar to call this form of settlement an elven/dwarven city. It would only become an elven/dwarven city if its inhabitants make some intentional modifications (swirling wooden staircases around looming tree trunks, decorated by glass lanterns, and so forth).

10 Dölling (2020) notes that this case illustrates how copredication depends ‘on the discourse context and the rhetorical connections between the two predications’. For my purposes, ‘rhetorical connections’ seems to pertain to predicate ordering and coherence.

11 Thanks to Robyn Carston for this observation.

12 Although he does not appeal to coherence relations, Collins (2017b:, 690, n. 5) notes that (i) is degraded relative to (ii) because of ‘the absence of any connection’, which can be interpreted as gesturing towards a concept not unrelated to coherence relations:

1. #London voted Labour, and is roughly a hexagon.

2. London voted Labour, but given new boundary changes, which [render] it roughly a hexagon, the Tories will stand a better [chance] next time.

13 Although see Kripke (2017) for evidence that ‘and’, as well as ‘but’, can convey conflicting implicatures, and so their applicability ranges considerably beyond conjunction.

14 Degand (1998) refers to these as ‘totally overlapping’, ‘partially overlapping’ and ‘exclusive’ interpretations; I have adapted this terminology.

15 One might respond that the putative overlap here is possibly due to the two predicates sharing the same (morphologically and semantically related) lexical root. However, replacing ‘has 500,000 smokers’ with the colloquial ‘has many citizens who regularly light up’ seems to trigger the same coherence relation.

16 Assumptions about atomic versus decomposable senses are found in other aspects of semantics, like with telic and atelic events (Stockall, Husband, and Beretta 2010).

17 aperture is somewhat different from location, since it involves a specific Euclidean frame and does not seem reducible to the physical object associated with it.

18 This proposal contrasts with Dölling’s (2020) claim that for lunch, lecture and bank, ‘neither [of the senses] … can be viewed as more basic’.