The argument for antonym pairs as constructions involves canonical antonym pairs (Murphy 2003b) [10] — that is, pairs of words in binary semantic opposition that are associated by convention as well as by semantic relatedness. English examples of canonical antonym pairs include black/white, fall/rise, alive/dead, off/on, private/public and so forth. In this category, I also include words that may not be antonymous according to some logical definitions of “antonym”, but which are nevertheless incompatible terms that are conventionally paired, like cat/dog and niece/nephew. The notion of canonical antonymy contrasts with simple semantic opposition in which the meanings are incompatible, but the words are not necessarily conventionally paired – for example, cold/scorching, blue/orange, and descend/rise.

There is plenty of evidence that canonical antonym pairs are closely linked in our minds. The members of these pairs tend to elicit each other (and not other semantic opposites) in psychological tests such as free word association (Deese 1965; Charles & Miller 1989), and people are faster at recognizing them as antonyms than non-canonical opposites (Herrmann et al. 1979; Charles et al. 1994). Charles, Reed and Derryberry (1994) also found that canonical (or in their terms, direct) antonym recognition is not affected by the semantic distance between members of the pair, whereas semantic distance in non-canonical opposites delays reaction times for those pairs. In semantic priming tests, canonical antonyms prime each other more strongly than non-canonical opposites (Becker 1980). These types of evidence have been used to argue that the relation between the words is accessible without recourse to semantic processing (e.g., Gross et al. 1989; Charles et al. 1994). If, as I am about to argue, such pairs constitute complex lexical items in their own right, such effects can be accounted for.

At the textual level, members of canonical pairs tend to co-occur within sentences at higher than expected rates (Justeson & Katz 1991, 1992). Jones (2002) estimates that one in about every fifty sentences of journalistic English contains an antonym pair, and current work on spoken English (Jones 2006), and child and child-directed English speech (Jones & Murphy 2005) reveals similar rates of co-occurrence. The canonical pairings are preferred to other semantically possible pairings (Willners 2001). For example, the non-canonical pair large/little is rare in text, but big/little and large/small are extremely frequent (Charles & Miller 1989). Furthermore, these studies show that knowing antonym pairs is not just a matter of knowing set phrases in which they occur, like the long and the short of it or neither here nor there. Instead, we can use any pair in a range of different constructions, as illustrated in (13), and any contrastive construction can contain a wide range of antonym pairs, as shown in (14) and (15):

It may be useful to briefly compare the data in (13) and (15) to set phrases that conform to particular contrastive constructions, such as from rags to riches (‘from poor to rich’) or the (American English) set phrase, from soup to nuts, meaning ‘exhaustively, from start to finish’. While the from X to Y phrase in (16) superficially has the same structure as those in (15), the set phrases are less schematic.

While from soup to nuts is a set phrase, from rich to poor (13d) instantiates the integration of the from X to Y construction with the lexical pair rich/poor. This is evident in the fact that the contrast between soup and nuts is only clear with reference to that particular phrase. The pair soup/nuts is used in other contrastive constructions only as a type of metalinguistic humour device, as indicated by the examples in (17):

Similarly, rags and riches co-occur frequently in the phrase from rags to riches, but phrases such as both rags and riches and neither rags nor riches are relatively rare (from rags to riches is found nearly a thousand times more often than these in a web search), and phrases like both riches and rags and neither riches nor rags do not occur at all on the web. Compare this to pairs like rich/poor, which are found across the range of contrastive constructions that allow nouns and adjectives (as in (13)), as well as in more abstract contexts, such as word-association tests and thesauri, and in cases of Ancillary Antonymy, as in (18).

The fact that rich/poor can trigger Ancillary Antonymy, in some cases creating a contrast where it is usually not found (as for stupid/ignorant in (18c)), indicates that the co-occurrence of the pair is not merely a result of language users having internalised rich and poor in a number of set phrases. In these cases, the phrases are not set, but the opposition between rich and poor is.

Comparing canonical antonym pairings to morphologically derived antonyms provides further evidence that the lexical (i.e., not morphologically derived) antonym pairs are directly represented in our language production faculties. While morphological opposites are readily available in English, we tend to prefer lexical opposites in contrastive constructions. For example, Jones (2002) searched for the word natural in various contrastive constructions. Artificial and man-made were the most frequent opposites found, while the morphologically derived antonym unnatural did not occur at all in the majority of the contrastive constructions investigated by Jones, and occurred at much lower rates than artificial and man-made in the constructions in which it did occur. So, while we have the ability to make predictable morphological opposites for many words, we often prefer to use morphologically unrelated words as opposites in contrastive constructions. The regularity with which we use the same words as antonyms in the face of the range of other available semantic opposites indicates a learned preference for pairing particular words.

Regular co-occurrence in contrastive constructions can cause word pairs to be added to the antonym canon, and their use as antonyms then extends beyond the original contexts in which they were found. For example, Jones (2002) looked at occurrences of style in contrastive constructions and found that it was frequently contrasted with substance. However, a shift in application of this opposite pair was evident in the nine years covered in his newspaper corpus (1988–1996). While in the earlier years it had been mostly applied in political contexts (particularly referring to Tony Blair and New Labour), by the end of the corpus period (and to the present), it was used much more broadly, indicating that the relation between the two words has been lexicalised at a more abstract (i.e., less context-bound) level. Similarly, the fact that rags and riches do co-occur in contrastive constructions other than from X to Y indicates that experiential reinforcement of that pair may have brought them to the periphery of the antonym canon.

The pairing of polysemous words can provide another piece of evidence that particular words are paired as opposites. Antonymy is usually classed as a sense relation (e.g., Lyons 1977), meaning that it relates word senses rather than the words themselves, which have a range of senses. Thus, the opposite of hot is cold when it is used as a temperature term, but the opposite is mild when hot is used to mean ‘spicy’. Nevertheless, when a member of a canonical antonym pair acquires a new sense, the opposition can be carried into the new semantic field, thus indicating that we perceive the words as related even apart from their usually opposed senses. So for example, black/white are opposed as achromatic colour extremes, but when one is used to describe some other state, the other is still available as an opposite, as in (19).

In each of these cases the earlier of the two usages is leftmost. So, for example, black was in regular use as a racial term before white was (Flexner 1976). While few of the items referred to in (19) are literally black or white, because one has been designated black, we can productively label its opposite white on the grounds that the black/white antonym pair is well known to us. In some of the cases in (19), it may also be the case that a metaphorical transfer has occurred. For instance, there are deeper semantic reasons to call a certain kind of testing white-box testing (19d) since the metaphor of a monochromatic colour scale has been imported to the testing schema. In this case, we might naturally invent the term grey-box testing if a test with partially visible internal processes were invented. However, the phenomenon does not reduce to metaphorical relatedness, since examples (19a) and (19c) are not metaphorical. Coffee with only a little milk isn’t grey coffee, for example.

The proposal that pairs like rich/poor and black/white are represented in the lexicon (e.g., Gross et al. 1989) is not new, but what is rather different is to propose treating antonym pairs as syntagmatically related words. [11] Approaches like lexical field theory (Lehrer 1974) or lexical networks like WordNet (Fellbaum 1998) treat antonyms as lexically linked to each other in a paradigmatic relation. These models represent paradigmatic relations (not just antonymy, but also synonymy, hyponymy and others) in order to partly represent the meanings of the words involved. Because they focus on the paradigmatic and semantic aspects of the relations, the syntagmatic aspects of antonymy in particular are neglected.

CxG’s sister theory, Frame Semantics (Fillmore 1978 and elsewhere) is often likened to field approaches (Kittay & Lehrer 1992), but it differs in that

“words or word senses are not related to each other directly, word to word, but only by way of their links to common background frames and indications of the manner in which their meanings highlight particular elements of such frames” (Fillmore & Atkins 1992: 77)

Yet in the case of canonical antonyms the word-word relation needs representation as well.

Construction Grammar provides the potential to bridge the gap between the syntagmatic and the paradigmatic. Thus, in addition to the representation of the semantic, paradigmatic relation of opposition in conceptual structure, I propose that certain pairings of opposite words are reinforced through linguistic experience and that such learned associations must be represented in the mental lexicon (or “constructicon”) in order to explain the particular distribution of these word pairs in discourse.

Treating antonym pairs as constructions provides an explicit means for describing and explaining the interaction of antonym pairs with the particular types of construction in which they appear. However, as discussed below, the antonym construction is not a contiguous phrase, and so we might consider it to be a paradigmatic construction, that is, a complex lexical construction that unifies with other constructions, resulting in syntagmatic usage. So, just as there is no strict division between the lexical and the grammatical in CxG, the syntagmatic and the paradigmatic constitute a continuum, rather than two distinct linguistic categories.

I begin the discussion of antonym constructions with a general antonym construction – a schematic version that summarises the common properties of the lexicalised versions, and that can also be used in the sanctioning of new antonym pairings. The proposal here is that canonical antonym pairs constitute lexical “items” made up of two lexical items that are syntactically similar and semantically related, but which are not in any phrasal relation to each other. In this way, this construction is unlike any that have been proposed before, but it is possible that other similar “paradigmatic” constructions could exist, for example among larger sets of contrasting categories (e.g., animal/vegetable/mineral) or hyponyms and their hyperonyms (apple < fruit). However, there is less evidence for canonical relations within paradigmatic relations other than antonymy and less opportunity for such relations to be experientially reinforced (see the discussion in Murphy 2003b, chapters 2 and 5).

The Antonym Construction requires the following properties (I-VI).

That is, although this construction has daughters, they are not in any specified syntactic relation to each other, therefore, it is not a phrase, but a complex lexical item.

That is, the syntactic categories of the two daughters must be unifiable.

A further feature specification is needed to indicate that this [lex +] construction is not a single lexical item. This is necessary because the lexical and categorical specifications alone do not distinguish single nouns, verbs, or adjectives from pairs such as these, and only the former can serve as daughters in constructions requiring a single noun, verb or adjective. The default specification for this unit feature is 1, but it is specified as 2 in pair-wise paradigmatic constructions such as the antonym construction. So, rich is [lex +, cat Adj, unit 1], while rich-poor is [lex +, cat Adj, unit 2]. A paradigmatic relation construction is thus one that is [lex +], but which has daughters. Ideally, the specification of the unit feature should result from the calculation of the number of daughters in a [lex +] construction. In the case of [lex ‑] constructions, the unit specification should automatically be 1.

An alternative analysis, which would not posit a differentiation between paradigmatic (unit >1) and syntagmatic (unit 1) constructions, would be to treat antonyms in the same way as other discontinuous but interdependent word pairs, such as both...and. In both cases, the words act as a pair, but are not (on their own) a phrase. Nevertheless, the relation between both and and can be represented in a phrasal construction: both X and Y. Following this example, one could try to model a phrasal construction that is extremely underspecified except for the co-occurrence of the two antonyms – something like (X) rich (Y) poor (Z). The material around the antonyms in this case is underspecified to the extent that we can say nothing about its category, lexical/phrasal or discourse status, or even whether it occurs at all (since in some constructions nothing comes before rich and in others nothing comes between rich and poor, and so forth). This approach has not been pursued here because it seems to force the antonyms to be ordered with respect to one another. As discussed in section 3.4, mandatory ordering of antonyms within an antonym construction is not desirable.

This can be represented as a feature incompatible in the semantic specification of the Antonym Construction, which takes the two daughters as its specification. This means that when two words appear in this construction, they are necessarily interpreted as being semantically incompatible.

This can be represented as a pragmatic feature contrast, which takes the two daughters as its specifications.

The schematic version of the Antonym Construction sketched above gives the framework for the particular pairings of canonical antonyms (or antonym constructs) that are recorded in the constructicon. The example of rich-poor is given in (20).

These antonyms can then occur in any construction that allows for contrasting adjectives. A two-dimensional box diagram cannot do justice to the intersection of the discontinuous, lexical antonym construction and the phrasal construction it slots into. (It would be best to represent the boxes intersecting in three dimensions.) However, the diagram in (21) demonstrates how an antonym construction consisting of two adjectives (or nominalisations of those adjectives) can slot into a contrastive construction when the two are unifiable with each other. In this case, we have a construction X and Y alike that requires two adjectives and indicates that the two adjectives are being contrasted. The lexical item rich/poor fulfils this criterion, and so can unify with it.

Of course, there is no requirement that the adjectives in X and Y alike consist of a construction in themselves. In the cases we saw in (4), lexical items like green and dry occur in the X and Y alike construction, and there is no reason to believe that such items constitute a pair. Similarly, rich and poor could be inserted as two lexical items that happen to be contrastive or as a lexical pair that is specified as being contrastive. In other words, the speaker has two options in using a contrastive construction – to fill it in with a ready-made antonym construction or to assemble a contrastive pair for the purpose at hand.

The semantic and pragmatic properties of the Antonym Construction (V and VI) deserve further comment. Both of these contend that the antonym relation is fixed at the pair level, in a top-down fashion. In other words, the Antonym Construction represents canonical antonyms as semantically incompatible, but this is not directly calculated from particular semantic properties of the two daughters. Of course, historically, these pairs have come to be lexicalised as antonyms because they are opposite on semantic grounds (in at least some of their senses/uses), but the antonym construction itself records the two words as an antonym pair without reference to those semantic details. There are several advantages to treating antonymy this way, rather than deriving the opposite relation from the words’ individual semantic properties within the antonym pair construction.

Firstly, the top-down approach to canonical antonym pairs is consistent with the experimental evidence (e.g., Charles et al. 1994) that oppositeness is understood immediately for canonical antonyms and not semantically derived anew, as it is for non-canonical antonyms.

Secondly, antonym relations are not limited to those pairs that are logically in a binary oppositional relation, and anything that can be construed as incompatible can be interpreted as antonymous in a particular context (Murphy 2003b). For example, (5) above showed green in context-specific antonym relations with veteran, roasted, experienced and orange. Because it was in a contrastive construction, green in these cases had to be construed in such a way that it was incompatible with the words with which it was contrasted. So, for example, when contrasted with roasted, green has to be construed as meaning ‘unroasted’ in part because of that contrast. In that case, the occurrence of the two words in a contrastive construction leads to their construal as opposites. However, for canonical antonyms, such construals can happen without the help of a contrastive construction. This can be seen in cases like those in (19), in which the pairing of black/white was shown to affect the interpretation of the words in the pair. So, for example, when we hear the term white market (19b), we understand this is incompatible with black market, since we know that black and white are opposites (because they are listed in the constructicon as instantiations of the Antonym Construction). This leads us to construe the meaning of white in such a way that it is incompatible with the meaning of black in the equivalent context. Thus, white is construed as meaning ‘legal’. This can only occur via a “top-down” antonym process, because the inference that white means ‘legal’ is due to the knowledge that it is the opposite of black, not by extending an established meaning of white.

Finally, representing the particular relationship between the two words involved would entail representing many different kinds of semantic relationship, and there is scant evidence that such relational types are relevant to the use of antonyms in discourse. Types of oppositional relation among words include contradiction, contrariety, reversal, and converseness, which all have different logical properties. Which kinds of opposites a word can have depends in part upon the meaning of the word. For example, converse opposites must have at least two arguments (in order that they can be reversed in the opposite word; e.g., buy / sell, parent / child), while contradictories (complementaries) must denote absolute states or properties. While such differences are relevant to the types of inferences that arise due to opposition (and thus must be represented in some way at the conceptual level), such semantic distinctions are not particularly relevant to pairing antonyms in discourse. It has been experimentally shown that most people do not recognise that pairs like up / down, happy / sad, and dead / alive are semantically different types of antonyms (Chaffin & Herrmann 1984). Furthermore, semantic antonym types have little relevance to the contrastive constructions in which they are used. So, for example, the relations between the Xs and Ys in X and Y alike might be complementary (male and female alike), contrary (young and old alike) or directional (up and down alike). [12] Representing these types of opposition semantically would involve different constructional types for each semantic type of antonym. Since there is no evidence that such information is relevant to the insertion of the pairs into other constructions, representing it at the constructional level would be unmotivated and would miss important generalizations that can be made about how antonym pairs are used in natural language.

The top-down antonym relation is represented in this analysis by features at both the semantic and pragmatic levels. [13] The semantic incompatible feature ensures that words are construed in such a way that their extensions do not overlap. This means that if fruit and apple are ever used as antonyms, they must be construed in such a way that fruit means something like ‘fruits other than apples’ (or in such a way that apple is not a type of fruit). It is impossible in this case for the meaning of f ruit that includes apples to be the antonym of apple (if apple denotes a type of fruit). The pragmatic feature contrast ensures that two items that are presumed to be antonyms are interpreted as being minimally different for the purposes of the context. This feature is needed in order to account for the oppositeness of incompatible items that are not logically in a binary relation. For example, sweet is incompatible with a number of other taste terms, for example salty , bitter , sour, savoury (and possibly others like meaty and spicy), and any of these can be an opposite of sweet, given enough context. At least three of these, savoury, bitter and sour, are (in some dialects/jargons, at least) canonical antonyms for sweet . When we choose one of these over the others to act as the antonym of sweet, it signals that the two states (e.g., sweet / bitter) are, for the context at hand, considered to be more like each other than they are like the other potential antonyms. That is to say, when bitter is the opposite of sweet it is because these two properties are perceived as being part of the same scale (from bitter to sweet) or because they are perceived as properties of the same types of things (e.g., pills, chocolate), whereas in this case sour is perceived as being a different sort of property which is found on a different scale and/or which applies to different kinds of things. So, while the semantic feature incompatible guarantees that the two words in an antonym pair are construed as incompatible, the contrast feature in the antonym construction ensures that they are construed as each other’s opposites.

As discussed in section 2, contrastive constructions such as X and Y ( alike ) and more X than Y carry a contrastive interpretation, so that the pragmatic contrast feature introduced for the Antonym Construction in this section could also be integrated into constructional treatments of the contrastive constructions in which antonyms often occur. So, a contrastive construction requires that two of its daughters are in the contrast relation, and an antonym construction provides two lexical items that are already in the contrast relation. This congruence between the two constructional types explains why they are so often integrated in discourse.

So far, I have posited no syntactic relations among the members of the antonym pair, but one might ask whether the two terms should be ordered with respect to each other. (While one item of the pair necessarily comes first in a box diagram like (20), the comma separating the two items indicates that the items are not ordered with respect to one another.) Certainly, in a neutral context poor/rich sounds more marked than rich/poor, and rich and poor is certainly more frequent than poor and rich (about 20 times more frequent in a web search). Nevertheless, there are a few reasons to be cautious about specifying ordering in the antonym construction or in any particular antonym constructs. Firstly, not all antonymic pairs have an inherent order. For example, there is no strong preference in the British National Corpus for cat(s) and dog(s) (42 occurrences) versus dog(s) and cat(s) (56 occurrences) (Murphy 2003b). Secondly, the strength or stability of the ordering varies even for those that do tend to occur in a particular order. For example, Jones’ (2002) corpus included no cases in which indirectly preceded directly, but while happy usually preceded sad, the order was reversed 30% of the time. Thirdly, for those that do exhibit a preferred order, the order can be overcome by semantic or information-structure demands in context. As we saw in (13) above, rich and poor can occur in either order in a range of contrastive constructions. In some contexts, the ordering of the two terms is semantically determined – that is to say, changing the order of the terms leads to a change in the meaning of the sentence, as in (13e), repeated here as (22):

In cases in which the ordering is not truth-conditionally important, the less marked term tends to go first, but this is contextually variable. While rich usually goes first, as it is less marked in more neutral contexts, poor comes earlier if it is unmarked in context. For example, in (23) (repeated from (13a)), the preceding context is about murdered missionaries who had been helping poor people. So, the fact that the man under discussion was gracious to poor people is already given information, and thus it is the unmarked term in this context.

In such cases in which discourse factors are at work in determining the order of antonyms, the issues regarding ordering can be accounted for in the theory in more general ways, by positing certain ordering constructions with reference to given/new information or markedness in discourse (see discussion in Fried & Östman 2004).

We still need to account for cases in which we do find strong preferences for ordering the pair (when discourse and truth-conditional matters are not a factor). Such cases are often part of a general morphological or semantic pattern. Derived antonyms come second in morphologically asymmetrical pairs, such as correct/incorrect and prove / disprove, and words that refer to more positive traits or traits involving having or being “more” tend to precede their opposites, as in rich / poor, true /f alse, big /l ittle. Temporally related antonyms tend to be chronologically ordered (begin / end). Phonological tendencies/constraints may also play a role in which member of a pair comes first, though they generally come second to semantic constraints (Cooper & Ross 1975; McDonald et al. 1993; Benor & Levy 2006). Gender opposites in particular seem to be affected by syllable length, with longer words typically following shorter ones (niece / nephew , ladies / gentlemen , man / woman , male / female). [14]

The tendency of antonyms to occur in particular orders is thus the result of a number of conspiring factors. First, semantic, morphological and phonological and prosodic properties of the words in the pair affect our preferences for ordering in largely predictable ways. Rather than specifying such ordering preferences in the Antonym Construction or any particular antonym construct, these general patterns should be accounted for by more general statements in the grammar – through particular ordering constructions or principles. Second, some pairs seem to be learnt in particular orders, since they occur in other more complex, lexically specified constructions in which they occur in a specific order. This is clearly true for irreversible binomials (Malkiel 1959) like ladies and gentlemen. In this case, we would expect that the ordering would be stronger in those set phrases than in other cases of co-occurrence, although experience of the words in that particular order may reinforce that order when the two words occur in other structures. So, while the disjunction ladies or gentlemen occurs more often than gentlemen or ladies, the difference in frequency between the two phrases is just a fraction of the difference between ladies and gentlemen and gentlemen and ladies. [15] In other words, the sequence is not as fixed in contexts other than the set phrase; nevertheless experience of the set phrase may strengthen a particular order across contexts. Finally, discourse considerations, such as markedness and information structure, play a role in the sequencing of the items. Since these are context-dependent, they cannot translate directly into ordering of the antonym construct, but may be a part of other constructions that the antonym pair enters into – for example, constructions that order given and new information. Experience of ordered antonym pairs in these contexts contributes to the strength of ordering in particular antonym pairs. Since there is plenty of evidence that ordering is predictable, although complexly so, there’s no motivation to include ordering as part of the general Antonym Construction or as part of many particular antonym constructions. For particular antonym pairs, however, experience of the antonyms in one order is likely to reinforce that sequence, and the strength of the sequencing varies with the amount of reinforcement and the number of factors (semantic, morphological, phonological, and so forth) contributing to that ordering. Thus, indication of the strength of ordering for particular pairs might be included in particular antonym constructions.