A corpus-based behavioral profile analysis of polysemy and antonymy: the case of the ancient Greek size adjectives mikrós and mégas

2.1 Antonymy: from antonymous words to antonymous senses

Antonymy stands as one of the most fundamental lexico-semantic relations, particularly among adjectives (Fellbaum 1995: 284; Murphy 2003: 169). Cruse (1986: 197) highlights the paradoxical linguistic behavior of antonyms, emphasizing their simultaneous resemblance and divergence: they share similarities in distribution and meaning, yet they also diverge notably along a prominent dimension of meaning. Traditionally, antonymy has mostly been viewed as a relation between words, where oppositeness is treated as a fixed configuration in the lexicon (Charles and Miller 1989; Deese 1964; Justeson and Katz 1991; Miller and Fellbaum 1991; among others; for a critical review, see Murphy and Andrew 1993; Paradis and Willners 2011; Sullivan 2012). Charles and Miller (1989: 242) could not be clearer on this point, asserting that “antonymy is a lexical relation between word forms, not a semantic relation between word meanings” (see also Gross et al. 1989). In this perspective, meanings are perceived as existing solely within a language rather than in human cognition (on the core principles of this tradition, see Geeraerts 2010: 87). However, Cognitive Linguistics has shifted this view, portraying antonyms as dynamic and contextually sensitive, construable within specific contexts (Cruse 1986; Jones et al. 2012; Kostić 2015; Kotzor 2021; Paradis 2011).

This context-sensitive approach redirects focus from the word level to the level of senses. One study that paved the way for this shift in focus is Murphy and Andrew’s (1993) work on the conceptual basis of antonymy, in which they demonstrate that contextual factors, such as the noun phrases the adjectives modify, can alter the interpretation of the adjective (cf. the meanings of fresh in expressions such as fresh shirt and fresh fish; Murphy and Andrew 1993: 309). Their experiments revealed variability in antonymic pairings among participants depending on context, suggesting that antonymy operates at the level of senses rather than words. Justeson and Katz (1991: 6) similarly acknowledge the semantic dimension of opposition, albeit with findings that somewhat straddle between the two theoretical perspectives. Finally, Sullivan (2012) offers the most systematic exploration of antonymy at the level of senses, analyzing the adjectives hard, soft, rough, and smooth through the BP approach. Her study finds that antonymy manifests most meaningfully at the sense level rather than at the word-/lemma-level, and concludes that studying antonymy at the sense level reveals nuanced patterns inaccessible at the word or lemma (Sullivan 2012: 307, 323). In this paper, we subscribe to this view, and our findings further support it.

Note, however, that while this sense-level perspective prevails, it does not entirely preclude the existence of antonymy at the word level. Although meaning is context-dependent, high frequency of co-occurrence of specific antonym pairs can lead to their conventionalization. What begins as a conceptual relationship can, through repeated use, become lexically motivated (Kostić 2015). This frequent co-occurrence strengthens their association in memory, facilitating their conventionalization, which ultimately leads to the entrenchment and automatic activation of the lexical pair (Kotzor 2021).

2.2 The behavioral profile method

The Behavioral Profile (BP) method is a corpus-based approach to lexical semantics, which was developed 20 years ago and continues to gain popularity. It has been applied to the study of polysemy of lexical items in synchrony (e.g., Berez and Gries 2009; Georgakopoulos 2020; Gries 2006; Jansegers et al. 2015; Liu and Dou 2023), to the analysis of near-synonyms (Divjak 2006; Divjak and Gries 2006; Divjak and Gries 2009; Divjak 2010; Huang and Chen 2022; Liu and Espino 2012), and to the analysis of antonyms (Gries and Otani 2010; Sullivan 2012).^[1] Notably, in the last decade, the approach expanded its applicability to the study of the diachronic evolution of lexemes (Jansegers and Gries 2020; Liu 2022). Its basic premise lies in the correlation between distributional similarity and functional/semantic similarity, which is rooted in the distributional theory of semantics proposed by scholars such as Firth (1957) and Harris (1954). It posits that distinctions or similarities among words and senses manifest in their contextual preferences, including various features spanning from lexical collocations to syntactic colligations, and many other aspects (e.g., pragmatic properties; Divjak 2010; Gries and Divjak 2010). The BP method enables the analysis of larger sets of words or senses, thereby facilitating a more detailed and extensive exploration of the data. Importantly, in the analysis of adjectives, it allows for the inclusion not only of the base forms of the adjectives but also of their inflected forms (see Gries and Otani 2010).

Since a crucial aspect of the method involves identifying a word’s contextual preferences, the procedure inherently entails extracting a high number of contextual features from the corpus sentences wherein the word appears. It is precisely these tokens’ morphosyntactic, semantic, and pragmatic features (collectively referred to as ID Tags, following Atkins’ 1987 terminology) and their distributional frequencies that play a crucial role in uncovering a word’s behavioral profile. The more fine-grained and detailed the study of the distributional characteristics of lexical items, the more precise the results. This approach to meaning is particularly useful, especially considering that our sample comes from a dead language, and intuitions are not available (cf. Georgakopoulos et al. 2020).

The specific four steps of the process as applied to antonyms are outlined as follows (Divjak and Gries 2009: 61; Gries and Divjak 2010: 338; Gries and Otani 2010: 128):

1. Retrieval of all instances of the lemmas of the antonyms from a corpus in the form of a concordance (at least at the level of sentence/utterance where the word occurs).

2. Thorough manual annotation of many properties of the use of the word forms, i.e., annotation of the ID tags.

3. Generation of a co-occurrence table indicating how often each ID tag occurs with a particular sense; the series of co-occurrence percentages (i.e., the BP vectors) shown in this table is what is called the behavioral profile.

4. Evaluation of the table through hierarchical agglomerative cluster analysis (HAC), which is the main exploratory technique used in BP studies. HAC represents the semantic relationships between the senses based on their distributional (dis)similarity, visualized through a dendrogram.

Through this process, it is anticipated that the correlation between distributional similarity and functional similarity will emerge, with elements demonstrating similarities appearing more tightly linked to each other than to other elements. As is evident from the fourth step of the process, the BP approach offers the possibility of a more complex statistical analysis of the data, not limited to observed frequencies and percentages (Gries 2012). This quantitative nature of the analysis enhances the accuracy and validity of the results.

3.1 Corpus data

The current study focuses on adjectival meanings in Classical Greek (500-336 BCE), using texts from five Classical authors, i.e., Aristotle, Demosthenes, Euripides, Herodotus, and Plato. This selection ensures a diverse representation of genres including philosophy, orations, tragedy, and history. The data were retrieved from PhiloLogic4 (https://artflsrv03.uchicago.edu/philologic4/Greek/; last accessed December 2022), a full-text search, retrieval, and analysis tool from the Project for American and French Research on the Treasury of the French Language (ARTFL), with texts available from the Perseus Digital Library (Perseus Project, http://www.perseus.tufts.edu/hopper/).

3.2 Methods of corpus query

In line with the protocol of the BP approach, we initially extracted 4,487 instances of the two adjectives (that is the full set of examples from our tailor-made corpus), comprising 1,187 tokens of the lemma mikrós and 3,300 of mégas. These instances included both the valid and the non-valid tokens. Tokens were marked as valid only if they had an attributive or predicative syntactic use, since in both cases they modify nouns explicitly, and this results in a more homogeneous, comparable and less complex data group. Instances where the element did not seem to modify a noun, such as when it functioned as a complement of a prepositional phrase or was used as a noun or an adverb, were considered invalid. For example, smikrón often appeared in adverbial phrases with the prepositions epí and katá, forming the expressions epí smikrón and katá smikrón, both meaning ‘by a little’. These cases were classified as invalid.

To maintain consistency between the two adjectives, 500 tokens were randomly selected from the valid ones for the study. The tokens were randomly sorted using the “=rand()” formula in MS Excel 2016. Similar to other BP studies (e.g., Gries and Otani 2010; Sullivan 2012), we incorporated a portion from each lemma in both the comparative and superlative forms (2 % of each lemma in each degree of comparison), namely mikróteros and meízōn as well as mikrótatos and mégistos, respectively. It is worth noting that Sullivan (2012) used a smaller percentage of superlative forms (1.1 %) compared to comparative forms (2.2 %). In the present study, however, we extracted an equal percentage of comparative and superlative forms (2 %).

In the second step, we manually annotated the 1,000 tokens for various morphological, syntactical, and semantic properties (ID tags). These ID tags were customized specifically for the analysis of ancient Greek, as some of them are expected to be language-specific. For example, in a language like ancient Greek, where nouns are morphologically marked for gender, one might include an ID tag for gender, whereas in English, where adjectives are not grammatically gendered, such an ID tag would not be applicable. The annotation scheme contained 18 ID tags, with varying numbers of levels (which again may differ depending on the language under consideration). The semantic ID tags have the highest number of levels, while other ID tags, usually syntactic ones, have a minimum of two levels. Table 1 shows the ID tags each token was annotated for, their type, and their levels.

Although the method is rather objective – given that the criteria applied in the different steps adhere to specific, replicable protocols – the classification of data can sometimes become subjective, as acknowledged by Divjak and Gries themselves (Divjak and Gries 2009: 277). As the morphosyntactic tagging of the tokens was a relatively simple and straightforward process, this limitation primarily pertains to word sense tagging. To mitigate this limitation, we adopt a practice employed in a few corpus-linguistic studies, wherein the data coding is conducted by multiple annotators (see Diessel 2008; Glynn 2010; 2016; Zeschel 2010).^[2] Specifically, the data were coded independently by the two authors of the current paper and a PhD student specializing in Ancient Greek Linguistics. The first coder coded the entire dataset. Subsequently, the second coder (the co-author) and the third coder coded a randomly selected 10 % of the data (100 tokens), focusing solely on the most challenging semantic ID tag, namely the sense of the adjective. Following this, inter-coder reliability was calculated using the Fleiss’ κ (Fleiss 1971), as more than two coders were involved. For the mikrós dataset, κ = 0.77 (z = 24.1, p < 0.001), indicated substantial agreement, while for the mégas dataset, κ = 0.82 (z = 26.90, p < 0.001), indicated almost perfect agreement.

Regarding the levels of the sense of the modified noun, we followed Sullivan’s method of using the classification system developed by the Corpora e Lessici dell’Italiano Parlato e Scritto (CLIPS) project by Federico Albano Leoni, “General Ontology for Nouns and Verbs” (http://webilc.ilc.cnr.it/clips/Ontology.htm; last accessed December 2022). We tried to use classes that were not too specific, so as to keep the number of senses included reasonable while at the same time avoiding using too general hyperclasses. It is important to note that we consistently coded for the literal sense of the noun, even when the intended meaning was figurative. Consider, for instance, the phrase pros húdōr smikrón (Plato, Theaetetus 201b), which translates to ‘in the short time’, as húdōr (‘water’) stands metonymically for the water clock (klepsúdra), a means of measuring time. In this case, we coded the noun as belonging to the ID tag level artifact, which is the literal meaning of the expression, rather than time, which is the intended meaning of this expression.

Regarding the annotation of the senses of the two adjectives, we relied on the sense distinctions provided in the three most significant ancient Greek-English dictionaries: LSJ (Liddell and Scott 1996), the Brill Dictionary (Montanari 2015), and the Cambridge Greek Lexicon (Faculty of Classics, University of Cambridge & Diggle 2021), as well as on the corpus data (for the full list of meanings included in the three dictionaries, see Appendix/Tables A.1–A.3). Since different dictionaries offer different sense distinctions, we needed to establish a consensus on the pool of senses to rely on for word sense tagging. Take, for instance, examples (4) and (5), which the Brill Dictionary would consider as belonging to the same sense (see Appendix/Table A.1/sense 3.3), while the Cambridge Greek Lexicon would treat them as belonging to different senses (see Appendix/Table A.3/senses 7 and 12). Since our strategy aimed at accepting more fine-grained distinctions, we also maintained the distinction between them, namely we distinguish between the senses ‘intensity’ and ‘volume’. It is worth noting that, in addition to our strategy adopted for practical reasons, there are also purely linguistic reasons for keeping these two senses separate. Specifically, the sense of ‘volume’ encompasses additional meanings compared to ‘intensity’, and the two senses appear in different collocations (with ‘volume’ exclusively modifying the noun phōnḗ ‘voice’).

‘intensity’

‘volume’

Additionally, we aimed to maintain consistency between the lists of senses for the two adjectives for the sake of comparability. For instance, we divided the ‘small in amount’ sense of mikrós as found in the Cambridge Greek Lexicon, into two separate senses: ‘amount quantity’ (see, ex. [6]) and ‘value’ (see, ex. [7]), reflecting a similar distinction found in the adjective mégas.

‘x.amount quantity’

‘x.value’

Finally, for comparability purposes, we used the same sense labels for both adjectives, indicating their difference by adding the symbol ‘x’ before the sense label for the adjective mikrós. In the aforementioned example, we used the labels ‘amount’ and ‘value’ for mégas and the labels ‘x.amount’ and ‘x.value’ for mikrós. This strategy also facilitated a clearer portrayal of the antonymous senses in a joint diagram. The final list of senses for both adjectives is provided in Table 2, along with a more precise description of each meaning, presented as a paraphrase.

Note that only a subset of the 15 senses was included in the analysis. We considered only those senses that met the threshold of 10 tokens in our corpus. Table 3 presents the frequency each sense is attested. The cells shaded in grey indicate the senses that were excluded from further analysis for not meeting the threshold.

The final two steps of the BP method involve generating a co-occurrence table and conducting a statistical evaluation of the data. These steps were carried out following the protocol outlined in Levshina (2015: chapter 15). Table 4 is an extract from the initial spreadsheet used for data collection and annotation, with each row representing an occurrence of the adjective and each column corresponding to a linguistic feature.

Table 5 presents the absolute frequencies for the levels of the ID tags negation and syntactic use for both adjectives. They indicate the number of times each level is realized for each adjective, in the current data. The sum of the absolute frequencies needs to be equal to the total number of tokens, namely 500 for each adjective.

In order to be able to compare them, the absolute frequencies need to be converted to relevant frequencies, as seen in Table 6. Table 6 is the output that was produced by the script from our input data (Table 4). Table 6 makes up the behavioral profile of the words. This translates to each word being identified by the total of relative frequencies of co-occurrence elements in a clause/sentence.

The final step, namely the evaluation of the behavioral profiles with statistical techniques, was performed using the R statistical software package (R Core Team 2022). Our results will not be assessed based on frequencies and percentages of co-occurrence but rather on hierarchical agglomerative cluster analysis (HAC). This statistical technique yields dendrograms that merge data points at different levels in the vertical axis based on their (dis)similarities; the lower they are merged the more similar they are. HAC divides the data into clusters based on similarity. Data items that are in the same cluster are more similar to each other than to data items from different clusters. Thus, HAC results in the production of numerous clusters, which enables the comparison of the data (Divjak and Gries 2006: 36). The graphs will be computed using the Canberra measure of distance and the amalgamation using Ward’s method for cluster analysis, both of which have been used in several BP studies (see Divjak and Gries 2006; Divjak and Gries 2009; Divjak 2010; Levshina 2015). The plotted graphs are presented in the following section.

4.1 The hierarchical agglomerative cluster results

This section presents the output of the BP script as a cluster dendrogram. First, we ran the script for each adjective separately, resulting in the dendrograms shown in Figures 1 and 2.^[3] Then, we ran the script using a .txt file that included the datasets for both adjectives, producing a joint dendrogram shown in Figure 3. Each dendrogram illustrates how the different senses are grouped together. Regarding the remaining notations on the dendrograms, the red figures show the Approximately Unbiased (au) p-value, whereas the green figures show the Bootstrap Probability (bp) value (the former is considered more accurate than the latter). Note that the alpha was set at 0.10 (see Levshina 2015).

Figure 1:

Hierarchical cluster analysis of 9 senses of mégas.

Figure 2:

Hierarchical cluster analysis of 11 senses of mikrós.

In Figure 1 for mégas, two main branches are evident. The first branch (left) contains the senses: (1) amount_quantity, (2) value. The second branch (right) includes a greater number of senses: (3) intensity, (4) importance, (5) power_ability (first sub-branch); (6) size_man_made (7) dimension, (8) power_control, (9) title_significance (second sub-branch). The two clusters make sense semantically because they distinguish between different types of ‘magnitude’: The first cluster captures a more concrete and measurable aspect of magnitude, focusing on entities that can be clearly measured or valued numerically. For instance, in example (8), mégas modifies a noun from the semantic class artifact, in the subcategory of money, which is an entity that can be both counted and valued in numerical terms.

‘value’

The second cluster represents abstract and qualitative attributes, reflecting a wider range of meanings related to power, significance, and impact that are not strictly quantifiable but are central to social or conceptual contexts (see examples 9–13). The division between the two clusters effectively organizes the senses based on how they express ‘greatness’, whether it is grounded in concrete, countable terms or abstract, qualitative attributes.

The minimum distance between the items is observed between ‘power_control’ and ‘title_significance’ (min(megas.dist): 33.97, as calculated using the dist() function in R. Consequently, these senses are merged lower on the tree. Consider examples (9) and (10), which illustrate these senses in specific contexts.

‘power_control’

‘title_signif’

The semantic connection between the two senses of mégas involves a progression from describing outstanding characteristics of power and influence (sense: ‘power_control’) to formally recognizing and honoring these characteristics with a title (sense: ‘title_significance’). Their close association can also be explained by their distributional similarity. To identify the features that distinguish this cluster from all other senses, we examined the absolute differences in the proportions of each value for every variable (for details and code, see Levshina 2015: 313–315).

To compare the features that differentiate the senses ‘power_control’ and ‘title_signif’ from all other senses, we first extracted the rows from the megas.bp data frame corresponding to these two senses (c1 in the code). The remaining rows, corresponding to all other senses, were then extracted into a separate data frame (c2 in the code).

We calculated the column means for each subset (c1.bp for ‘power_control’ and ‘title_signif’, and c2.bp for the other senses). By computing the differences between these column means, we were able to quantify how each feature contributes to distinguishing these two senses from the others. Finally, the differences were sorted in descending order to identify the most distinguishing features.

The analysis revealed that the top differentiating feature is the absence of an inanimate modified noun for the adjective, with an average difference of 44 %. The second key parameter is the ID tag sem_modif_noun.institution (see the noun pólis in both [9] and [10]), with an average difference of 35 %, followed by sem_modif_noun.human, with an average difference of 29 %.

Turning to the second cluster of senses, ‘power_ability’ and ‘importance’ are grouped together because they exhibit similar behavioral profiles. Specifically, both tend to modify nouns associated with the event type ACT, a characteristic they also share with the sense ‘intensity’, which explains their close association in the dendrogram. This is illustrated in examples (11)–(13), which all feature nouns of this category.

‘power_ability’

‘importance’

‘intensity’

However, it should be noted that a single feature is not sufficient for two senses to be considered similar. Instead, the entire behavioral profile of each sense must be taken into account, rather than isolated features. With this caveat in mind, consider Table 7, which lists the top four differentiators for each branch in the dendrogram (each representing two or more senses) as shown in the HAC results in Figure 1.

Turning to mikrós, Figure 2 presents the plot resulting from the hierarchical cluster analysis of 11 senses of the adjective. In this Figure, two main branches are also evident. The first branch (left) includes the senses: (1) x.power_control, (2) x.size_man_made (first sub-branch); (3) x.age, (4) x.size_human (second sub-branch). The second branch (right) includes a greater number of senses: (5) x.dimension, (6) x.importance (first sub-branch); (7) x.duration, (8) x.amount_quantity, (9) x.value (second sub-branch); (10) x.intensity, (11) x.power_ability (third sub-branch). The senses in the first cluster revolve around physical and intrinsic attributes of living beings and objects. They describe qualities that are inherent to entities, such as age, size, and strength (see examples 16–17). In contrast, the second major cluster focuses on abstract attributes and quantities, combining measurable quantities with abstract properties, states, or conditions. Consider, for instance, example (14), which focuses on a measurable quantity, such as the size of an area, and example (15), where mikrós modifies the noun prâgma (‘matter’), representing not only a concrete, measurable entity but also an abstract concept, such as the perceived significance or importance of a matter.

‘x.dimension’

‘x.importance’

The shortest distance between items in this dendrogram is observed between ‘x.power_control’ and ‘x.size_man_made’ (min(mikros.dist): 37.55). Consider examples (14) and (15), which illustrate these senses in specific contexts.

‘x.power_control’

‘x.size_man_made’

In this case, the semantic linkage between the two senses of mikrós involves the notion of reduction or diminishment in some form. ‘x.power_control’ suggests a lack of power, authority, or significance, while ‘x.size_man_made’ refers to something being small or limited in size. Here, physical size is metaphorically extended to refer to power or significance: a ‘small’ state, for example (see example 16), can metaphorically imply weakness. The underlying metaphor in this case is unimportant is small (on the metaphors important is big and unimportant is small, see Yu, Yu and Lee 2017; Grady 1997). From a distributional perspective, the most important feature distinguishing this cluster from all other senses is the inanimate modified noun (level: yes), with an average difference of 26 %. The next three key factors are gender.feminine (13 %), transitive_V.yes (7 %), and article.no (6.5 %).

In the second cluster of senses, ‘x.amount_quantity’ and ‘x.value’ are grouped together because of their shared behavioral profiles. As noted for mégas, both senses share the semantic feature of being linked to entities that are quantifiable or measurable in numerical terms. From a distributional perspective, they are primarily associated with nouns from the categories of possession and quantity. Consider examples (18)–(19) in which mikrós modifies a noun of the category possession.

‘x.amount_quantity’

‘x.value’

As with mégas, Table 8 presents the four most important features that differentiate the groups of senses, as clustered in Figure 2, from all other senses.

4.2 Clusters based on oppositeness or relatedness of meaning?

We now return to our original question: in a joint network that includes the senses of both antonyms, are the senses organized by antonymy or polysemy? To address this, we first present the joint dendrogram that includes the senses of mikrós and mégas that meet the threshold of 10 observations in our corpus (see Figure 3).^[4]

Figure 3:

Hierarchical cluster analysis of the senses of mikrós and mégas.

The results from the joint dendrogram (Figure 3) are particularly illuminating. In every instance, antonymous senses cluster together in tight groups, a strong indication of their similarity. Consider, for example, the pairs <size_man_made; x.size_man_made>, <power_control; x.power_control>, <dimension; x.dimension>, <intensity; x.intensity>, and <importance; x.importance>, which form tight clusters. In contrast, polysemy appears to have no influence on the organization. The outcome is categorical: even when two senses of a single lemma are grouped together (as in the polysemy network), the corresponding antonymous sense is also present within that group. This is exemplified by the ‘power_ability’ sense, which loosely joins the <x.amount_quantity; x.power_ability> cluster.

This finding is partly consistent with Sullivan’s (2012) study, which shows that a dendrogram of antonymous and synonymous adjectives (hard, soft, smooth, and rough) is primarily organized by antonymy. In contrast, it differs from Gries and Otani’s (2010) finding, which did not identify any evidence of antonymy. Furthermore, while Sullivan’s study – focusing on the relationship between antonymy and synonymy – demonstrates a secondary organization by synonymy, our study – examining the relationship between antonymy and polysemy – reveals that the joint network of both antonyms is exclusively organized by antonymy, with no indication of secondary structuring by polysemy.^[5] This finding further supports the argument that antonymous adjectives should be analyzed at the level of senses rather than words. Future research examining near-synonyms, such as brakhús (‘short’) and makrós (‘long’), will likely provide a more definitive answer to this issue.

That the level of senses is more vital for antonymy than that of words or lemmata is further supported by the fact that antonymous senses of mikrós and mégas often co-occur in the same context (cf. Sullivan 2012: 319). This claim aligns with the majority of existing studies on antonymy, particularly those focusing on substitutability, as previously discussed (Justeson and Katz 1991). Although the current work does not primarily focus on the collocations of antonymous senses, the co-occurrence of the two adjectives in the same context is notably high. Using a tool available from PhiloLogic4, one can determine that the adjectives mikrós and mégas appear together in 390 sentences within our selected corpus.^[6] As shown in examples (20)–(21), the distributional behavior of the antonymous adjectives is rather common, sharing the same syntactic elements such as a common verb, subject, or modified noun.