A cross-linguistic study of lexical and derived antonymy

2.1 Lexical typology and research on motivation

Lexical typology has by now firmly established itself as an important field of typological inquiry. Probably most of its significant achievements concern cross-linguistic variation in how languages categorize particular semantic domains (e.g., color, space, temperature, motion, body, etc.) by means of lexical expressions, but there is also a growing cross-linguistic research on lexical motivation, such as polysemy patterns, co-lexification, or semantic shifts associated with particular lexical expressions. Following Koch and Marzo (2007: 263), “[a] lexical item L₁ is motivated with respect to a lexical item L₂, if there is a cognitively relevant relation between the concept C₁ expressed by L₁ and the concept C₂ expressed by L₂ and if this cognitive relation is paralleled by a perceptible formal relation between the signifiers of L₁ and L₂”. Lexical motivation can apply to single lexical expressions, but also to groups and even whole classes of lexical expressions. To illustrate the former, piglet (‘a young exemplar of a pig’) is motivated by pig due to a clear cognitive relation between the concepts expressed by the two paralleled by the presence of the suffix -let, and the verb to cup (‘to form one’s hand or hands into the shape of a cup’) is cognitively related to the noun cup paralleled by the change of word-class affiliation. Moving to the more general patterns of lexical motivation, the Swedish compound päronträd ‘pear tree’ exemplifies a strategy used in denominations for several species of trees by compounding the name of the fruit they bear with the word for ‘tree’ – e.g., also äppleträd ‘apple tree’, plommonträd ‘plum tree’, etc. Typical derivational categories, such as agent, patient, instrument and locative nominals from verbs, evaluation (augmentative and diminutive), various verbal categories (causative and anti-causative, inchoative, frequentative, etc.) and word-class changing categories, or transpositions in Spencer’s (2014) terms (action nominals, denominal adjectives, etc.) all provide examples of lexical motivation applying to whole classes of lexical expressions.

There is in fact a huge complex of issues related to the question of how the lexicon, or at least some of its subparts, is organized in terms of structurally more basic versus more complex lexical expressions and what systematic motivational patterns there are (see Koptjevskaja-Tamm et al. [2015]; Koptjevskaja-Tamm and Veselinova [2020] for overviews). The relevant research is, however, strikingly uneven. Only a few word-formation strategies, in most cases well known from the more familiar languages, like the ones mentioned above, have been described for many individual languages and language families, but systematic cross-linguistic research on word-formation strategies and their functions has so far been surprisingly modest (cf. Müller et al. [2015]; Štekauer et al. [2012] for important steps in this direction).

An excellent example of lexico-typological exploration along these lines is provided by the systematic cross-linguistic research on form and meaning relationships in pairs of intransitive and transitive verbs, such as the causative synu ‘break (intr.)’ → syn-dyru ‘break (tr.)’ in Kazakh, or the anti-causative lomat’ ‘break (tr.)’ → lomat’-sja ‘break (intr.)’ in Russian. The research tradition goes back to Nedjalkov (1969) and Haspelmath (1993), but was further developed methodologically and extended to larger samples in Nichols et al. (2004). The outcome of Nichols et al.’s (2004) study is a typology of 80 genealogically and areally diverse languages based on their treatment of 18 pairs of what the authors view as semantically basic and almost universal intransitive verbs (‘sit’, ‘fear’, ‘laugh’, ‘fall’, etc.) and their transitive counterparts. The verb pairs have been selected with an eye on various parameters, among others, those known or supposed to influence derivational processes, and the languages in the sample (extended to 200 in Nichols [2018]) turn out to be relatively consistent in whether they treat intransitives as basic and transitives as derived (transitivizing languages), whether they derive intransitives by means of anti-causative morphology (detransitivizing languages), whether both intransitives and transitives are encoded by the same labile verb (neutral languages), or whether both intransitives and transitives have the same status (indeterminate languages).

In this study we want to explore to what extent another central lexical or lexico-semantic relation – antonymy – exhibits systematic formal expression. The next subsections provide a further background for this pursuit.

2.2 Earlier studies on antonymy

Antonymy has been a popular topic in lexicology, semantic theories, and logic; there is also a growing psycholinguistic, neurolinguistic, and corpus-based research literature focusing on antonymy (cf. Jones et al. [2012]; Kotzor [2021] for an overview). Antonymy evokes the notion of opposition, or oppositeness, with the traditional distinction between contradictory (or, complementary) versus contrary opposites going back to Aristotle: the former are “either or”-relations, which exhaustively partition a particular domain into two subdomains (dead vs. alive, or even vs. odd), whereas the latter express opposite poles on a scale that also includes a gradient middle ground between them (big vs. small, or good vs. bad) (Horn 1989: 5–45). Researchers differ in how narrowly or broadly they define antonymy: in much of the literature antonymy is understood narrowly, as basically restricted to gradable (scalar), or contrary opposites, while for others antonymy includes not only contradictory opposites, but reversives (fall vs. rise, or dress vs. undress), and conversives (child vs. parent, or teacher vs. student) (cf. Lehrer and Lehrer 1982).

The crucial notion of oppositeness, underlying antonymic relations, involves two components – logical incompatibility and maximal semantic similarity (see Jones et al. [2012: 3] for a useful account). A stone cannot be both big and small simultaneously, so big and small are logically incompatible. However, logical incompatibility is not sufficient for defining oppositeness – for instance, big and generous are not semantically opposite only because a big stone cannot be simultaneously described as generous. For being opposites, the two expressions need to be semantically quite similar, or – seen from the other side – minimally different from each other. For instance, they should be able to be used in the same context, among other things, to apply to the same entity. In our example, generous and big are simply too different from each other semantically and generous cannot apply to stones at all, either big or small.

In structuralist approaches to the lexicon (cf. Lehrer and Lehrer 1982) and in much lexicographic work – e.g., in the WordNet project (Fellbaum 1998; Miller et al. 1990) – antonymy is taken to be the most basic means of organizing the adjective lexicon. There is also extensive psycholinguistic (and so far relatively modest neurolinguistic) research on the role of antonymic associations in the mental lexicon and memory (e.g., Bentin 1987; Deese 1966; Herrmann et al. 1986; Gross et al. 1989; Jeon et al. 2009; Roehm et al. 2007; see also Hay 2001) and a growing corpus-based research on discourse functions of antonymy and its typical manifestations in corpora (e.g., Jones et al. 2007; Kostić 2015; Lobanova et al. 2010; Muehleisen and Isono 2009; Willners and Paradis 2010; Wu 2017). There is, thus, accumulated strong evidence for antonymy being both a psychologically real and an important relation.^[2] In view of what has been mentioned so far, antonymy seems to be a good candidate for giving rise to systematic motivational patterns across languages, whereby a recurrent cognitive relation between the two members in at least some antonymic pairs may be paralleled by a perceptible formal relation between the two (see the definition given in Section 2.1).

However, opinions vary as to what kind of relation antonymy actually is. The traditional view, further developed in the structuralist position, speaks about lexical relations or lexico-semantic relations, i.e., (semantic) relations among words. However, one and the same word can have different antonyms, often corresponding to distinct word senses. These, in turn, can be linked to different entities whose properties are characterized by the different antonym pairs (Herrmann et al. 1986). For instance, in English, tall is opposed to short when talking about people, and to low when talking about buildings, and the opposite of white in the context of wine will be red and not black, which is conceived of as its typical antonym when referring to a color spectrum (Jones et al. 2012: 14). Antonymy is thus not so much about words, but rather about words used in a particular sense. Pursuing this line still further, antonymy can be understood as a conceptual relation, i.e., a relation between construals (Croft and Cruse 2004: 169). It has even been suggested that antonym relations are pragmatically driven and are derived in contexts of use, making them much less dependent on the words’ lexical properties (Murphy 2003).

These differences in opinion are indicative of the inherent heterogeneity in which pairs of linguistic expressions can be classified as antonymic, with some being antonyms par excellence and others much less so. What counts as minimal difference/maximal similarity is, of course, to a certain extent pragmatically determined, with some contrasts being highly context dependent and others more or less conventionalized. Much current evidence points to a continuum of goodness of antonymy, in the spirit of Herrmann et al. (1986). According to Herrmann et al. (1986: 134–135), the ability to recognize a pair of words as gradable antonyms depends on two groups of factors: the nature of the dimension of meaning obtaining between them and the relative position of their meanings on it. First, the denotative meanings of the two words should share at least one dimension, with clearer dimensions more likely to be perceived as such. For example, good versus bad, sharing a single dimension of goodness, is a better antonym pair than holy versus evil, which includes at least two (goodness as well as moral correctness). The two words should also share all the relevant denotational dimensions: holy versus evil, sharing both dimensions, is a better antonym pair than holy versus bad, in which only the first word includes the element of moral correctness. The two meanings in better antonym pairs should also be on opposite sides of the dimensional midpoint (hot vs. cold is better than cool vs. cold) and preferably at the same distance from it (hot vs. cold is better than hot vs. cool).

It has also been suggested that languages have “canonical antonyms”, i.e., “a limited core of highly opposable couplings that are strongly entrenched as pairs in memory and conventionalized as pairs in text and discourse, while all other couplings form a scale from more to less strongly related” (Paradis et al. 2009: 381). The dimensions that have been suggested, primarily on the basis of detailed studies of English and Swedish, include speed (slow–fast), luminosity (dark–light), strength (weak–strong), size (small–large), width (narrow–wide), merit (bad–good) and thickness (thin–thick) (Paradis et al. 2009; Willners and Paradis 2010). In other research strands it has been argued that dimensional adjectives are particularly well structured with respect to antonymy, i.e., they come in antonymic pairs, whereas oppositions for other domains frequently involve clusters of adjectives at the oppositive poles, e.g., brave/bold/courageous versus cowardly/timid/fearful (Bierwisch and Lang 1989; Morzycki 2015: 138–140).

Much of the theoretical discussion on gradable antonymy has centered on the nature of scales (e.g., unbounded/open vs. bounded/closed) and on the asymmetries between the members of different antonym pairs. Cruse (1986: 206–214) and Cruse and Togia (1995) distinguish between three major types of antonymic construal with respect to scales: polar, equipollent, and overlapping. Polar antonyms are monoscalar, whereas equipollent and overlapping are biscalar. Polar antonyms are arranged along a simple scale. e.g., short–long along the scale of length. Equipollent antonyms involve two scales with adjacent zero points pointing in opposite directions, e.g., cold–hot arranged on the scales of coldness and hotness. Finally, overlapping antonyms involve two overlapping scales, one major and one minor; for example, good–bad are arranged on the scale of merit, which is the major scale covering the whole range, and the scale of badness, which is the minor scale ranging from a mid-scalar position to the lower end of the major scale.

As regards asymmetries, research on antonymy frequently evokes the notion of markedness, with one member of an antonymic pair as marked and the other as unmarked, with certain semantic and syntactic properties typically ascribed to or expected from the unmarked member. It is useful to distinguish between morphological markedness, with morphologically marked words containing additional morphs as compared to the unmarked ones (impossible vs. possible), and semantic markedness (see Ingram et al. [2016]; Lehrer [1985] for overviews). Some of the recurrently mentioned indicators of semantic markedness include the neutrality of unmarked gradable antonyms in questions such as “How long/big/old etc. is X” compared to “How short/little/young etc. is X”, as well as the differences in frequency and evaluation (valence) of the two members of antonymic pairs. It may be expected that the two types of markedness are related to each other in that morphological markedness would be a symptom of semantic markedness – the idea at heart of the notions of cross-linguistic (Greenberg 1966) or typological (Croft 2003) markedness, which couples asymmetries in morpho-syntactic behavior of related categories to their semantic relations or asymmetries. However, the notion of semantic markedness turns out to be fairly elusive when applied to a larger class of contrasting pairs, even within one and the same language, with different tests and properties not necessarily pointing in the same direction (Lehrer 1985). In fact, as shown by Cruse and Togia (1995), most of the tests work best or exclusively for monoscalar antonyms. Semantic markedness has recently been evoked as a potential explanation for preferred antonym sequences (e.g., long preceding short, or alive preceding dead) in various languages – at least in English (Ingram et al. 2016; Jones 2002), Serbian (Kostić 2015) and Mandarin Chinese (Wu 2017) – with different interpretations of the results. The research on markedness-related phenomena in antonymy seems to suggest that markedness is difficult to apply across the board and that it might be easier to break it down into several different factors, especially in cross-linguistic research. It appears that valence (emotional evaluation) and magnitude, when applicable, can explain a large portion of asymmetries in the behavior and processing of antonymy pairs, in particular combined with frequency differences (which often follow from valence).

Notably, there is also an established tradition to talk about polarity and to ambiguously apply the terms “positive” versus “negative” antonyms for at least two different contrasts – one having to do with evaluation (valence) and the other with magnitude. Thus, from the point of view of valence, ‘difficult’ and ‘dirty’ are negative, while ‘easy’ and ‘clean’ are positive. From the point of view of magnitude (as, for example, evident from the neutrality in the above mentioned questions), ‘short’ and ‘new’ are negative, while ‘tall’ and ‘old’ are positive. The two distinctions do not necessarily point in the same direction: it is not quite clear which one of the ‘dirty’ versus ‘clean’ antonyms is positive from the point of view of magnitude, ‘difficult’ is, probably, evaluated more negatively than ‘easy’, while the evaluation of ‘old’ versus ‘new’ as more positive or more negative is fairly context- and culture-dependent. The problematic and ambiguous use of such terms as “(un)marked”, “positive”, and “negative” has been noted and commented on in various connections (see also Morzycki 2015: 124). However, of special interest for our purposes is the claim that the semantic distinction between the two members in many antonymy pairs is that of polarity and involves the presence or absence of a negative element, particularly evident in dimensional adjectives (Bierwisch and Lang 1989; Cruse and Togia 1995: 123), but not confined to them (see Cable [2018]; Heim [2019]; Kennedy [2001]; Morzycki [2015: 124–134] for different suggestions on how to account for these facts).

In summary, two points discussed in this section are particularly important for our study:

1. Given that there is a general cognitive relation of “oppositeness” between the members in (at least prototypical) antonymic pairs, we may expect that this relation is mirrored by a direct formal relation, such as derivation of one from the other, between the two members.

2. Given that the idea of polarity/negation in a broad sense is often evoked to explain the semantic difference between the members in at least some antonymic pairs, we may expect that such formal markers will often have the semantics of negation – in a broad sense – paralleling im-possible in English or ne-vozmožnyj (‘neg-possible’) in Russian.^[3]

2.3 Earlier studies on affixal negation in antonymy

Given that adjectives with the negative affixes, akin to the English impossible and its Russian equivalent nevozmožnyj, are widely attested in Germanic, Slavic, and Romance, the phenomenon has attracted considerable attention in linguistic research. Some of the leitmotifs in these studies are the restrictions on the application of the negative affixes, on the one hand, and the meaning/interpretation of adjectives with negative affixes, on the other (see Horn [1989: 273–308] for an overview of the field, which to a large extent still holds). To illustrate the first type of query, why are unwise and unhealthy possible, but not *unstupid and *unsick? For a language like English, with multiple negative adjectival affixes (un-, in-, dis-, non-), the additional challenge is to provide a principled analysis of the choice among them, cf. unhappy, inaccurate, dishonest, nonverbal (see Lieber [2004: 111–125, 154–177] for a theoretical account). To illustrate the second type of query, do unhappy and impossible mean the same as not happy and not possible, respectively? Traditionally, the main idea is that affixal negation often creates gradable rather than complementary opposites, but the situation is far more complicated (Horn 1989: 273–308; see also, e.g., Colston 1999; Farshchi et al. 2021).

An interesting case comes from sign languages, which despite generally being under-researched languages have garnered some attention in typological studies of negation, and by extension antonymy. As languages in a visual modality with multiple simultaneous articulators, sign language morphology is not always as linear or overtly (de)compositional as spoken languages, making the definition of a morpheme somewhat complex. Research on negation across sign languages has shown that while some languages make use of manual, linear negators, others favor negation expressed simultaneously through non-manual markers (e.g., headshakes, facial expressions) or changes to the (affirmative/neutral) stem (e.g., changing or reversing the path or direction of the movement, substituting the handshape) (see Zeshan 2006). With regard to antonyms more specifically, some of these patterns have been shown to be present in antonym pairs. For example, some antonymic word pairs in Chinese Sign Language have been shown to have mirrored/reversed articulation or use opposite dimensions of movement (e.g., horizontal vs. vertical) (Yang and Fischer 2002). A larger study across many sign languages showed a general trend of positive members of antonym pairs being more likely to have upward movement, compared to their negative counterparts, arguably mirroring the spatial metaphor good is up (Börstell and Lepic 2020). In cases of non-linear alternations, it is often more difficult to argue that one form is necessarily the more “basic” form with the other being derived from it, which is why we, for this study, focus on spoken languages.

The primary inspiration for the current investigation is Zimmer’s (1964) dissertation, which, in contrast to most of the other studies, is an early attempt to approach the phenomenon of affixal negation in adjectives cross-linguistically and to formulate cross-linguistic generalizations. It is primarily devoted to the distribution of adjectives with the negative affixes in English, German, French, and Russian, with the data coming from a systematic search in several dictionaries and texts. In addition to these four well-described European Indo-European languages, the study also contains somewhat less systematic observations from a few other non-Indo-European languages (Jordanian Arabic, Mandarin Chinese, Finnish, Hungarian, Ilokano, Japanese, Kabardian, Tamil, Thai, and Yoruba). Zimmer is chiefly interested in the direction of derivation in those cases where one of the antonyms is derived from the other one, i.e., whether and to what extent it is possible to formulate any “universals” or generalizations regulating this formally asymmetric relation between the members of antonym pairs.

Zimmer starts from the two different versions of what he calls “the derivational universal” – largely inspired by the earlier research on German, Swedish, and English by researchers such as von Jhering, Wundt, von Ginnehen, Noreen, and Jespersen – meant to explain the asymmetry of unwise versus *unstupid and unkind versus *unmean:

1. Negative affixes are used primarily with adjectival stems that have a “positive” value on evaluative scales such as good – bad, desirable – undesirable

2. Negative affixes are not used with adjectival stems that have a “negative” value on evaluative scales such as good – bad, desirable – undesirable (Zimmer 1964: 15)

It is primarily the second, more careful formulation whose validity Zimmer seeks to test in his study. The challenge, however, is to determine what counts as the positive versus negative value of an adjective. Zimmer’s solution is to restrict the use the evaluative terms to such clear cases as good versus bad, healthy versus sick, honest versus treacherous, where “positive”, “in a more or less neutral context (such as They are…) would be understood as expressing a favorable judgment, or as describing a state generally considered desirable” (Zimmer 1964: 16). Other adjectives, such as smooth, rough, regular, or sporadic, should be considered as “neutral”. As Zimmer notes, this intuitive solution is not entirely satisfactory, but should probably work quite well for the more obvious positive and negative terms, at least as long as one is dealing with languages and cultures one is familiar with.

Zimmer’s results show that the “derivational hypothesis” in any of the two formulations does not hold in the languages under investigation, even though there seems to be some truth to it (already English provides counter-examples to both, such as unselfish). In his sample, it is Russian that stands out in being much more tolerant in its derivation of negative adjectives than the Germanic languages and French. However, as Zimmer notes, the Russian negative prefix ne- is homonymous with the negative particle ne, which has a very broad distribution. And this, as he “is tempted to speculate”, may account for its seemingly unrestricted application in that “various ‘odd’ combinations into which the prefix enters might be explained as developments from the homonymous phrases involving the negative particle” (Zimmer 1964: 66). Zimmer suggests therefore the following, more cautious derivational hypothesis in terms of preferences rather than absolute restrictions, which will play an important role in our study:

We could perhaps say that for any given language negative affixes that are distinct from the particle(s) used in sentence negation are likely to have a greater affinity for evaluatively positive adjective stems than for evaluatively negative ones. What this means in practice is that for any language with such negative affixes we would at least expect that the number of “negative” adjectives among their derivatives would exceed the number of “positive” adjectives among their derivatives. (Zimmer 1964: 82)

But even Russian is not completely free in deriving adjectives by means of negative prefixes. Zimmer suggests a local restriction operating in Russian, whereby in antonym pairs designating what he calls “dimensions of length” a ne-adjective can only be derived from the word denoting the larger pole, e.g., vysokij ‘high’ – nevysokij ‘neg.high’ – nizkij ‘low’ – *nenizkij ‘neg.low’ (Zimmer 1964: 64). This observation will also be of interest for our study.

Likewise important for our purposes is Zimmer’s observations that certain terms recur frequently in lists of words with negative affixes in different languages (e.g., words for ‘uncommon’, ‘ignorant’, ‘dishonest’, ‘unjust’). These in turn underlie his suggestions for a future quest for possible generalizations governing the distribution of lexical versus derived antonym pairs across the different concepts.

Another problem for further investigation would be the degree to which there is cross-linguistic similarity in the concepts that are designated by simplex terms, and the degree to which antonym pairs of the schema ‘x vs. un-x’ can be matched in different languages having negative affixes […] The questions to be investigated would be of the following kind: Is it generally true that words for ‘just’ have no simplex antonyms? Are there languages in which ‘common’ is customarily designated by an expression meaning ‘not rare’, or ‘regular’ by an expression meaning ‘not random, not haphazard’? Such questions of lexical universals (whether they be “factual universals” or significant preponderances of certain lexical features) are of considerable interest and can moreover be investigated with a fair degree of ease. (Zimmer 1964: 90)

Our study, seeking to explore which types of property words are typically targeted by derivational versus lexical antonymy, and why, is largely inspired by Zimmer’s observations, hypotheses, and suggestions for further research, which also underlie some of the more specific hypotheses formulated and tested in this study.

2.4 Hypotheses

The hypotheses presented in this section derive from the theoretical discussion in the preceding sections. First, relating to the notion of (semantic) markedness we propose the hypothesis that semantically marked property concepts will be more likely expressed through negative derivation than their unmarked counterparts. We admit that, given the problems associated with the notion of semantic markedness discussed in Section 2.2, this hypothesis may be seen as only weakly justified. Nevertheless, since the notion has been so prominently present in the literature, we do find value in attempting to find a way to address it empirically and propose an operationalization of this notion for the purposes of our study in Section 3.2. The next two hypotheses relate to Zimmer (1964), and rest on a firmer theoretical grounding also in the light of the discussion in Section 2.2. They narrow down the comparison to the two major contrasts behind the notion of semantic markedness: valence (evaluation) and magnitude. We will hypothesize that evaluatively negative property concepts and property concepts denoting smaller magnitude will be more likely expressed through negative derivation than evaluatively positive ones and ones expressing larger magnitude, respectively. The next hypothesis is likewise inspired by Zimmer’s (1964) suggestions for a future study on possible generalizations governing the distribution of lexical versus derived antonym pairs across the different concepts, for the purposes of this study operationalized as a comparison of property concepts from the different semantic types in Dixon’s (1977) sense. We hypothesize that oppositions involving property concepts from core semantic types will be more likely expressed without negative derivation – i.e., with plain lexical forms – than those involving oppositions from the non-core semantic types (cf. Section 3.2 for the details on the semantic types). And finally, based on the principle of economy, we will hypothesize that there will be a trade-off between lexical and derivational expression of the property concepts involved in the antonym pairs. This hypothesis is not necessarily based on an expectation that lexicons should follow the principle of economy; surely there are various semantic and pragmatic factors at play as well, and in the case of triads like happy–sad/unhappy the two alternative antonyms of happy are not synonyms. It is, however, interesting to investigate in a broad quantitative perspective to what extent languages follow the principle of economy, and in order to do this, we have formulated the idea as a testable hypothesis. The hypotheses are numbered and formulated as follows.

With the hypotheses now explicitly formulated, we will move on to presenting the material and methodology we use in order to test them. Section 3 will introduce the necessary technical terms we use in operationalizing the hypotheses, most importantly the distinction between neg-constructed and plain lexical forms as the short-hand for words containing the derivational element with the meaning of negation and those that lack such an element (see Section 3.3).

3.1 Language sample

The research reported in this paper is based on questionnaire data from 55 spoken languages across 23 language families (see Figure 1 and Table 1). We have aimed at including languages from all continents and from a wide variety of families and branches in the spirit of variety sampling (see Miestamo et al. 2016). However, as is very often the case in questionnaire-based work, the availability of experts to fill in the questionnaire has led to certain biases in the sample of languages surveyed. Languages of western Eurasia are well represented in the sample, but the other continents much less so. We are aware of the problems that the underrepresentation of these continents and language families will pose to the generalizability of our results. At the same time, the dense coverage of Eurasian, especially Indo-European (n = 13) and Uralic (n = 10), languages has the advantage that it will serve as a basis for addressing diachronic issues and more local contact phenomena in future work. For the purposes of the present paper, we can say that the sample gives a good coverage for western Eurasia and the presence of a fair number of languages from other continents provides preliminary impressions on the degree of the generalizability of the results beyond Eurasia. Note, however, that our statistical models (Section 4) do address the bias, thereby increasing the degree of generalizability. In Table 1, we list the sampled languages by macroarea and family, with the family classification taken from Glottolog 4.7 (Hammarström et al. 2022). In Figure 1, the languages are shown on a world map, illustrating the geographical distribution of our sampled languages. Additionally, the languages are listed in alphabetical order in Appendix A with a listing of the experts who helped us by filling out the questionnaire and answering our subsequent questions. Their contribution is gratefully acknowledged.

Figure 1:

The language sample by geographical distribution and genealogical family classification. 1. Amharic, 2. Akan/Twi, 3. Eton, 4. Gurenɛ, 5. isiNdebele, 6. Orungu, 7. Swahili, 8. Wolof, 9. Umpithamu, 10. Warlpiri, 11. Hebrew, 12. Basque, 13. Bulgarian, 14. Dutch (Flemish), 15. German, 16. Italian, 17. Khowar, 18. Lithuanian, 19. Punjabi, 20. Romanian, 21. Russian, 22. Sinhala, 23. Slovak, 24. Spanish, 25. Swedish, 26. Japanese, 27. Georgian (Modern), 28. Korean, 29. Khalkha Mongol, 30. Aghul, 31. Cantonese, 32. Dejongke, 33. Sakha, 34. Turkish, 35. Erzya, 36. Estonian, 37. Finnish, 38. Hungarian, 39. Komi-Zyrian, 40. Mari, 41. Nganasan, 42. North Saami, 43. Udmurt, 44. Veps, 45. West Greenlandic, 46. Yucatec Maya, 47. Choctaw, 48. Indonesian, 49. Kankanaey, 50. Kilmeri, 51. Mian, 52. Bunaq, 53. Mapudungun, 54. Hup, 55. Shipibo-Konibo.

3.2 Questionnaire

As mentioned in Section 3.1, our study is largely inspired by Nichols et al.’s (2004) cross-linguistic research asking to what extent the members in pairs of intransitive versus transitive verb pairs are involved in derivational relations with each other, based on translational equivalents for the set of 18 pairs of concepts. For our data collection we used a list of 41 property concept pairs that stand in an antonymic relation to one another, accompanied in the questionnaire by one or several nouns that they typically modify, e.g., ‘long versus short stick’, ‘deep versus shallow river/lake’. The questionnaire is addressed to experts in particular languages who are asked to translate the expressions given in the metalanguage to the language under study, to gloss the data and, if possible, to answer a few further questions to help us in our analysis of the data. (The original questionnaire sent out is in English, but in some cases, language experts semi-informally translated the questionnaire into other languages for eliciting responses from their language consultants.) First, when several different alternatives are given for a particular concept in a language, we are interested in the potential semantic and pragmatic differences among them. We are also interested in the morphosyntactic features of the property concepts in the language, i.e., in their word-class affiliation (whether there is a separate class of adjectives in the language or whether such concepts are lexicalized as verbs and nouns), and in the predicating and modifying constructions used for property concepts in general and for those in the list in particular. In the questionnaire, we also ask several questions pertaining to negation (standard and ascriptive) and to the markers used in the expression of antonymy in case there are such examples in the data. Thus, the data collected could be used to address various additional aspects of antonymy. However, in this study, we focus our attention on the formal distinction between lexical and derived antonymy.

Our method of data collection follows the long tradition of collecting cross-linguistic data by means of concept sets, or word lists, for different purposes, including research on various morphosyntactic and word-formation phenomena (e.g., Beavers et al. 2017, 2021; Haspelmath 1993; van Lier 2016; Ye 2021, among others) – see also List et al. (2016) for examples of different types of concept lists. Remarkably, there is hardly any discussion of the foundational matters in the ample research based on concept lists, and the same concept lists are recycled across multiple studies without further notice. The assumption underlying such endeavors, in most cases tacit, is that the concepts in the lists are easily understandable and easily translatable across languages, which is, of course, a very crude simplification of linguistic reality. On the contrary, as repeatedly confirmed by the accumulated evidence of the growing field of lexical typology, there are very few meanings that can easily translate among languages. In fact, the so-called translational equivalents in two languages hardly ever mean the same, among other things, but almost always differ in their denotational range and range of uses, and in general, in their place in the lexical system of the language and their relations to other lexical items – cf. Kibrik (2012) for convincing examples of the incongruity among the English list of presumably basic verbal concepts and its correspondences in Russian and Koyukon.

However, the extent to which the (exact) semantics and (precise) semantic identity of the items in the lists matters depends on the research goals. For instance, the 18 verb pairs in Nichols et al. (2004) have been chosen on pragmatic grounds, as representing certain combinations of general parameters, corresponding to frequently encoded situations and having approximate translational equivalents in many languages. Stricter requirement on semantic comparability would create obstacles for achieving the principal objective of the study – cf. Koptjevskaja-Tamm (2008: 35–37) for a discussion of concept lists in lexical typology.

At the same time, it is necessary to avoid obvious pitfalls in the method, stemming from the frequent ambiguity or multifunctionality of many forms. In the extreme case one and the same form in the concept list corresponds to two different homonymic words (e.g., light – ‘not dark’ vs. ‘not heavy’) or has two very different meanings (e.g., dull – ‘blunt, not sharp’ vs. ‘lacking interest or excitement’), and the translation may relate to the wrong one, i.e., not to the one envisaged in the original concept list (cf. List et al. 2016). In other cases, the differences between different senses, readings, or uses of one and the same form are significantly subtler (and therefore not as easily appreciable), but may receive different translations in other languages. For instance, sharp in English corresponds to tranchant and aigu/pointu in French: the first one applies to knives, swords, saws, and other instruments with a blade (a “functional edge”), while the second applies to needles, arrows, and other “instruments with a functional endpoint” (Kyuseva et al. 2022). German and Italian also have special ‘sharp’ adjectives for needles and arrows (spitz and appuntito), as opposed to the more general adjectives scharf and affilato. If ‘sharp’ is included in a list of concepts with no clear indication as to which of its uses is meant, its translations into some languages may end up being incongruent with each other (see Section 2.2 for the discussion of antonymy as applying to words used in particular senses).

One way of improving the applicability of a concept list as a tool for collecting data and its value as tertium comparationis in cross-linguistic research consists in supplying concept labels with short definitions, following the line adopted in Concepticon (List et al. 2016, 2023). Continuing with our example, ‘sharp’ is defined in Concepticon 3.1.0 as ‘having the ability to cut easily’ (List et al. 2023: https://concepticon.clld.org/parameters/1396), which would not apply to sharp arrows and needles.

Our methodology makes use of the frequently noted close connection between the interpretation of a property concept word and the entities it applies to, as seen in the ‘sharp’ example: knives and saws have an edge and can cut (easily), while sharp needles and arrows will be able to pierce something, but hardly ever to cut anything.^[4] Following this idea, we have provided the antonym pairs in the list with one or several nouns that they typically modify in a particular reading/use or, in several cases, in two or more particular readings/uses, e.g., ‘sharp versus dull knife/arrow/spear/needle’, which may result in several different translations. By adding nouns from a specific category, we hope to have achieved a sufficient degree of semantic congruity among the translational equivalents in our sample.

As explained in the introduction to the questionnaire, we are not interested in adnominal modification as such. Rather, the reason to collect the data in the form of noun phrases with property concept modifiers is to avoid confusion with clausal negation. During the project we have become increasingly aware of the fact that not all languages allow or prefer adnominal modification to the same degree. We have therefore kept open the possibility to give the property concepts in predicative position as well. But in those cases we have to be sure that we are dealing with derivational negation on the word level and not with clausal negation. In other words, the property expression may be given in a predicative position as long as negation is attached to the term on the lexical/phrasal level as in (1a), but the clause may not be negative as in (1b).

When the property concepts are given in predicative position, it is especially important that the respondents specify how the negative expressions in the data relate to clausal negation. It is perhaps worth pointing out that clausal negation is a morphosyntactic construction whose function is to negate a clause, and it typically instantiates sentential negation (as defined by Klima [1964]; cf. Jespersen [1917] for “nexal negation”), but sentential negation can be expressed by other means as well (for a discussion, see Miestamo 2005: 3–5, 39–42).

The list of antonym pairs <Antonym 1, Antonym 2> follows the broad semantic classification of property concepts suggested in Dixon’s (1977) seminal work and further developed in Dixon and Aikhenvald (2006). Dixon (1977) differentiates among core semantic types, found in languages with both large and small adjective classes, peripheral semantic types, associated with medium and large adjectival classes, and, finally other types, associated with large adjective classes in some languages. We have aimed at a relatively balanced representation of Dixon’s three classes and have also made sure that all canonical antonyms (cf. Paradis et al. 2009: 381) are included. We have also tried to ensure a good representation of the two main types of opposition distinguishing the two members in an antonym pair: valence (positive vs. negative)^[5] and magnitude (more vs. less).^[6] We aimed at being as consistent as possible in the choice of the order between the members of an antonym pair, according to the main principles:

1. When the two antonyms differ in their valence, Antonym 1 is the more positive one (e.g., ‘good’ vs. ‘bad’, ‘rich’ vs. ‘poor’); this will enable us to test Hypothesis 2.

2. When the two antonyms differ in their magnitude, Antonym 1 designates the greater degree (e.g., ‘big’ vs. ‘small’, ‘heavy’ vs. ‘light’); this will enable us to test Hypothesis 3.

The two principles together covered the lion’s share of the antonym pairs in the questionnaire. In very few cases, when these principles gave contradictory results (e.g., ‘light’ vs. ‘heavy’) or could not apply straightforwardly (e.g., ‘black’ vs. ‘white’), we simply resorted to our intuition as speakers of several European languages, partly supported by the studies on preferred antonym sequencing in English, Serbian, and Mandarin Chinese, such as Jones (2002), Ingram et al. (2016), Kostić (2015), and Wu (2017).^[7] On the whole, Antonym 1 and Antonym 2 more or less correspond to what is usually understood as the semantically unmarked and marked members, respectively, and this dichotomy is thus directly relevant to testing Hypothesis 1 as well. It is of course conceivable that semantic markedness relations between antonyms are not always translatable among languages, but we hope that our methodology is sufficiently consistent for the purposes of the study and will briefly discuss this issue in Section 5.

Our original list comprised the following semantic types of property concepts, which are targeted by Hypothesis 4:

1. Core semantic types:

   dimension (1–6)

   age (7)

   value (8–9, 34, 36)

   color (10)

2. Peripheral semantic types:

   physical property (11–18, 25, 27, 28)

   human propensity (19–24, 26)

   speed (29)

3. Other semantic types:

   difficulty (30)

   similarity (31–32)

   qualification (33–39)

   position (40–41)

The questionnaire sent out to the language experts included a total of 41 antonym pairs <Antonym 1, Antonym 2>. After manually going through the data, we discovered that three of the antonym pairs (‘light’ vs. ‘dim’, ‘like/similar’ vs. ‘unlike/different’, ‘same’ vs. ‘different/other’) received uncertain or misinterpreted responses for multiple languages and were consequently excluded from further analysis. Furthermore, two of the antonym pairs (‘possible’ vs. ‘impossible’ and ‘probable’ vs. ‘improbable’) were in many responses treated as identical, i.e., there was no difference in the lexicalization of the concepts along the lines of distinction made in English. We resolved this by collapsing these two antonym pairs into a single one (‘possible/probable’ vs. ‘impossible/improbable’). The original questionnaire as it was sent out to the language experts is available in the OSF repository (https://osf.io/8kuzh/). Table 2 shows the resulting 37 antonym pairs from our questionnaire data that we used in our analysis, split into antonym (members) 1 and 2, and categorized into groups based on class (core vs. peripheral vs. other) and semantic type. When applicable, the antonym members are provided with their values for valence (positive vs. negative) and magnitude (more vs. less) – one antonym pair, ‘rich/wealthy’ versus ‘poor’, is included in both valence and magnitude subsets.

3.3 Data processing

As mentioned in Section 3.2, the kind of data we have collected with the questionnaire can be used for addressing many aspects of antonymy. In this paper, we focus on the formal expression of antonymy, paying attention to whether the members of antonym pairs are expressed with neg-constructed lexical forms versus forms that are not neg-constructed, which we will call plain.^[8] In fact, for any antonym member, a language may have both forms available for a single concept, or one or none of the forms. We then ask how the plain and neg-constructed forms are distributed between the members of the antonym pairs, across antonym pairs and across languages – that is, for which concepts a word form (or, formation strategy) exists, but the two are not mutually exclusive by default as both can co-exist (e.g., sad and unhappy represent plain and neg-constructed forms, respectively, for the same concept in English).

To determine what counts as a plain versus neg-constructed form, we used the following criteria. We count as neg-constructed forms in which the added derivational element has the semantics of negation in a broad sense, and as plain forms that are not derived with an element that has the meaning of negation in a broad sense. What we mean by negative in a broad sense includes, in addition to pure negation, cases where the negative force is only partial, such as ‘little’ or ‘few’, cf. the French expression peu profond for ‘shallow’ discussed below. Clear instances of plain are found in examples such as big–small and fast–slow where both terms are morphologically simple. Clear instances of neg-constructed can be found in examples such as happy–unhappy and wise–unwise where the latter term is formed by adding an element which clearly has negation in its semantics. Thus, the first term in both of these is referred to as plain, whereas the second is referred to as neg-constructed. The latter is a more technical term for what we referred to as expression via negative derivation in formulating our hypotheses in Section 2.4, and conversely, plain then corresponds to expression without negative derivation.

Furthermore, to count as neg-constructed, a term does not have to share the stem with its antonym or to be derived from it. For example, the Turkish term imkansız ‘impossible’ derived from imkan ‘possibility, facility’ with the caritive/privative suffix -sIz, counts as neg-constructed, even though its opposite olası ‘possible’ is not related to it at all.

Noteworthy, a plain form does not have to be simplex, but may be derived, as long as the derivational marker is not negative in its semantics. An example of a derived expression that would count as plain for our purposes would be cowardly, which contains the derivational marker -ly, which is not negative in its semantics and does not turn the meaning of its base to its opposite. One of its opposites, fearless, will, however, count as neg-constructed due to the negative semantics (caritive/privative) of -less. In Estonian, both õnnelik ‘happy’ versus õnnetu ‘unhappy’ are derived from õnn ‘happiness’ by means of the general adjective suffix -lik and the caritive/privative suffix -tu. For our purposes, only õnnetu ‘unhappy’ qualifies as neg-constructed due to the negative semantics of the derivational suffix, whereas õnnelik ‘happy’ counts as plain. Examples like this are relatively common in our database.

An example of a less clear case of constructed expression is provided by French peu profond ‘shallow’ (lit. ‘little deep’), where the added element peu ‘little’ is not straightforwardly negative in its semantics. For the purposes of this study, our broad definition of negative semantics includes meanings like ‘little’ that have only partial negative force. Note also that in this example, the added element is not morphologically bound to the base; morphological boundness is not a requirement for a term to count as constructed. What is crucial, however, is the status of the whole combination as a lexicalized property expression.

This brings us to cases like ‘sad’ in Hup hãwɨg hi-hũʔ- [heart/spirit fact-finish] ‘(to have) one’s heart/spirit be ending’, or ‘generous’ in Akan-Twi, yam ye [stomach be.good] ‘(someone’s) stomach is good’. Since we are interested in property words, such idiomatic phrasal expressions are disregarded at an earlier stage of data processing. Examples like these are fairly common in our database, in particular for concepts describing human propensities.

The process of coding the questionnaire data from the language experts into a uniform machine-readable format entailed giving categorical values to each possible word form type per antonym pair. This gives us four cells to fill for every table row representing an antonym pair: one for plain and one for neg-constructed for each member of the pair. In our data coding, we have used the values “yes” (word form exists for the antonym pair member), “no” (word form does not exist for the antonym pair member) and “-” (not applicable to the antonym pair member for the language in question, used if, e.g., the language only has a phrasal expression for the meaning or data is simply missing). The data were coded (i.e., interpreted and transferred from questionnaire responses to our annotated database) by the authors with the help of research assistants. All coded entries were checked and double-checked by the authors, and any inconsistencies or unclear cases were discussed and resolved after mutual consultation and agreement.

Table 3 shows how these different cases presented above are coded in our database.

For some languages, there are more than one antonym sets available per antonym pair, for instance if the language has several synonyms that all fall within the concept of the antonym pair as described in the questionnaire. The total number of data points from our questionnaire is 9,032. Since we in this study are interested only in what is possible per antonym pair and language, we conflate all such multiple data points by the presence of any “yes” value: if a language has a plain word form for one antonym set but not for another set, both belonging to the same antonym pair, the value for that language and antonym pair will be “yes” – thus, we give each language the same weight, namely one data point per potential word form per antonym pair. For example, English has clever/wise–stupid–∅–unwise for the antonym pair ‘clever/wise–stupid/unwise’, which would be coded as having a “yes” value for Ant1 plain (clever and wise), Ant2 plain (stupid), and Ant2 neg-constructed (unwise), but “no” for Ant1 neg-constructed (as there is no *unstupid form for the meaning ‘clever/wise’). This coding resulted in a data set of 8,021 data points, representing one value for each of the possible word forms (plain and neg-constructed) for each of the antonym members (1 and 2), for each language (n = 55) and antonym pair (n = 37). In this final data set used for the subsequent analyses, we have also excluded any “-” values as they represent absence of data. The tidying, statistical analysis, and visualization of the data were done with the statistical language R 4.3.2 (R Core Team 2023): for the data wrangling and analyses, we used {tidyverse} (Wickham et al. 2019), {readxl} (Wickham and Bryan 2023), {lme4} (Bates et al. 2015), {broom.mixed} (Bolker and Robinson 2022), and {emmeans} (Lenth 2022); and for data visualization we used {tidyverse} (Wickham et al. 2019), {scales} (Wickham and Seidel 2022), {maps} (Becker et al. 2021),{ggnewscale} (Campitelli 2022), {patchwork} (Pedersen 2022), {ggrepel} (Slowikowski 2023), {ggbeeswarm}(Clarke et al. 2023), and {sjPlot} (Lüdecke 2023).