Not Known: Anonymous, Unknown or Non-known? A pilot test on the interpretation of negated absolute adjectives in Romanian

Absolute adjectives (open/closed), in contrast with relative adjectives (tall/short), are said to behave symmetrically, the negation of one form entailing the assertion of the other. However, Paradis and Willners (2006) show that absolute adjectives behave rather asymmetrically when negated. This paper investigates the profiles of absolute adjectives in relation to negation in Romanian. Based on the assumption that the weaker sentential negation nu A ‘not A’ may be used as a substitute for a stronger option, a judgement test was designed, in which the interpretation of nu A ‘not A’ (not healthy) is tested against the affixal negation neA ‘unA’ (unhealthy) and non-A (non-healthy) and the lexical antonym B (ill). This may prove useful as the negative affixes and polar antonyms have different encodings, pointing to different places on the scale. While nu ‘not’ usually denotes the entire negative part of the scale, the antonym B denotes the absolute end of the scale and ne‘un-’ usually gives rise to scalar negative compounds, denoting an intermediate position on the scale. The results confirm the asymmetric behaviour, the profiles of the adjectives showing different degrees of scalarity, displayed on a continuum ranging from having no scaling potential to having a strong or weak relative-like interpretation. Accordingly, three main categories of adjectives have emerged: absolute, relative-like and adjectives that can be either absolute or relative. In sum, the results show variation among the absolute adjectives as a class but also inside the mentioned subclasses. Furthermore, the adjectives do not seem to cluster in pairs but rather independently. Although only a few negative compounds are lexicalized, the participants have supplied the non-lexicalised compounds in order to fill in the gap between nu A ‘not A’ and the polar opposite B.


Introduction
It is claimed that absolute adjectives have truth-conditions that make reference to fixed (either maximal or minimal) standards of comparison (Kennedy and McNally 2005;Kennedy 2007). A consequence of this is that, in the case of a pair of antonyms, the negation of one entails the assertion of the other: the door is not open entails the door is closed. Although absolute adjectives are expected to behave similarly, it has been shown that the entailments do not always hold, four asymmetric patterns being identified in Swedish (Paradis and Willners 2006): (a) dead = not alive; alive = not dead, (b) wrong = not right; right ≠ not wrong, (c) bound ≠ not free; free = not bound, (d) empty ≠ not full; full ≠ not empty. The asymmetries are attributed to the scaling potential the absolute adjectives acquire contextually.
This paper aims at investigating the profiles of negated absolute adjectives in Romanian on the basis of experimental protocol. In order to see whether absolute adjectives have some scalar, relative-like features, a judgement test was designed in which the interpretation of the absolute adjective used in relation to sentential negation nu A 'not A' (Ea nu este sanatoasa 'She is not healthy') was tested against the affixal negation neA 'unA' (nesanatoasa 'unhealthy'), the affixal negation non-A (non-sanatoasa 'non-healthy') and the polar opposite B (bolnav 'ill'). Contrasting the interpretation of the negated absolute adjective nu A 'not A' with its morphological and non-morphological equivalents may prove useful, as the negative affixes and polar antonyms have different encodings, pointing to different places on the scale and different types of semantic opposition. The polar opposite B denotes the absolute end of the scale (De Clercq 2013: 32-36) whereas the negative affix ne-'un-' usually gives rise to contrary negation and scalar negative compounds, denoting an intermediate position on the scale. The polar opposite B and the affixal negation non-A preserve the entailment relations, giving rise to contradictory negation.
For the present study, the prerequisite was that the adjectives under investigation had the polar opposite B lexicalised, while this was not a necessary feature for the negative compound neA 'unA'. It should be noted that in the absence of the polar opposite B, the negative affix ne-'un-' can have a contradictory meaning, leading to antonym pairs, such as clar -neclar 'clear -unclear', potrivit -nepotrivit 'fit -unfit', sigur -nesigur 'certain -uncertain'. Our hypothesis is that when the polar opposite B is lexicalised, the negative compound neA 'unA' is not strengthened to the meaning of the polar opposite. In other words, when both sănătos 'healthy' and bolnav 'ill' are lexically available, the negative compound neA 'unA' (nesănătos 'unhealthy') does not express the same meaning as the polar opposite B 'ill', i.e. it does not denote the absolute end of the scale, but an intermediate position between nu sănătos 'not healthy' and bolnav 'ill'. This hypothesis is supported by some preliminary testing in which the participants (N = 13) were asked to indicate the position neA 'unA' occupies on a scale from 1 to 10, where 1 was the positive form of the adjective and 10 was the lexicalised antonym. The adjectives tested were: nevinovat 'unguilty ', incorect 'incorrect', neatent 'unattentive', nesănătos 'unhealthy', neadevărat 'untrue', necunoscut 'unknown', nemort 'undead', negreşit 'unwrong', nedistrat 'undistracted', nebolnav 'unill'. The results of the preliminary testing showed that, regardless whether the stem to which the affix attached was positive or negative, none of the neA 'unA' negative compounds were considered to express the same meaning as the polar antonym B. Instead, their position varied between 5 and 10 on the scale, depending on the adjective.
In the current judgement test, the participants were asked to indicate what the meaning of a negated absolute adjective nu A 'not A' (Ea nu este sănătoasă 'She is not healthy') was by choosing among the polar opposite B (bolnav 'ill'), the negative compound neA 'unA' (nesănătos 'unhealthy') or non-A (non-healthy). The prediction is that if the participants choose the neA 'unA' compound as the intended interpretation for nu A 'not A', regardless whether the negative compound is lexicalised or not, then the adjectives have no longer equivalent meanings, i.e. the bi-implication can no longer be applied. Instead the adjectives have to some extent scalar, relative-like features. The present study is a pilot test and uses only Romanian data. In what follows, the main properties of gradable adjectives will be discussed, followed by a discussion on the asymmetric behaviour of some absolute adjectives when used in relation to negation. In section (3), the online judgement test is described, with an emphasis on the main hypotheses, predictions, items and procedure. The paper ends with the discussion of the main results and indicates possible follow-up studies.

Properties of gradable adjectives
Gradable adjectives have been extensively investigated in the literature and various classifications have been suggested (Croft and Cruse 2004;Kennedy 1999;2007 2004; Toledo and Sassoon 2011;Yoon 1996;Unger 1975). First, gradable adjectives can be either 'relative' or 'absolute' adjectives. Expressing increasing or decreasing measures, relative adjectives (tall, wide, intelligent) can be laid out on a scale, denoting a range on the associated scale but lacking maximal or minimal values. They have context-dependent standards of comparison based on the properties of the adjectives and on the features of the context of utterance (Kennedy 2007). In order to have the standard established, comparison classes are needed, as indicated by the for PP in the following example: a tall boy for his age. They generate open scales, their standard lying at some point along the scale.
In contrast, absolute adjectives have context-independent, endpoint-oriented standards of comparison, i.e. their interpretation is based on a conventionally fixed standard of comparison. The standards of comparison come in two subtypes, leading to a distinction between maximum and minimum standard adjectives (Kamoen et al. 2011). The minimum standard requires their arguments to have some minimal, non-zero degree of the property in question (wet, open, dirty), while the maximum standard requires their arguments to have a maximal degree of the property (dry, closed, clean). As a result of their boundedness, i.e. when their standard is located at their scale's absolute maximum or minimum point, they are said to form partially closed scales: for instance, dirty generates a lower closed scale, while clean generates an upper closed scale, as a result of the standards they relate to. 1 There are also absolute adjectives that generate fully closed scales, i.e. they are compatible with either minimum or maximum standard interpretations in the positive form (opaque/transparent, full/empty). They display interpretive variability and can take on maximum standard interpretations in some contexts and minimum standard interpretations in others (Kennedy 2007).
In an antonymic pair of absolute adjectives, because they share the same scale and differ only by the ordering relation, one term is usually a maximum standard adjective and the other one is a minimal standard adjective. Since a minimal positive degree corresponds to a maximal negative degree on the same scale, the entailment relations lead to symmetric patterns, i.e. the negation of one term entailing the assertion of the other (1a-c). In contrast, relative adjectives lead to asymmetries, as shown in (1d)  Second, gradable adjectives can also be classified based on their logical properties (see Horn 1989). The theory of opposition discusses the logical relations displayed in the square of opposition. Contradictoriness and contrariety are two of the four logical relations. Contrariety characterizes the propositions or concepts that cannot be true at the same time but can be simultaneously false. A middle ground is created, allowing for a 'neither-nor' interpretation. In contrast, the members of a contradictory pair cannot be true or false simultaneously and do not allow a middle ground. Absolute and relative adjectives give rise to different types of opposition when negated: absolute adjectives give rise to contradictory negation while relative adjectives give rise to contrary negation. Third, gradable adjectives are classified based on criteria such as gradability, oppositeness and boundedness (Paradis 2001). Opposed to non-gradable adjectives (daily, classical), gradable adjectives are further divided into 'bounded' vs. 'unbounded' adjectives, on the one hand, and 'scalar' vs. 'non-scalar', on the other. Unbounded adjectives can be laid out on scale but they never reach the end-point, do not express any boundary, which makes the expression neither wide nor narrow perfectly natural. Bounded adjectives are absolute and divide some conceptual domain in two distinct parts (Paradis andWillners 2006: 1052). In consequence, the entailment relations hold for both members of the adjectival pair: "she is dead" entails that "she is not alive" and "she is alive" entails that "she is not dead". The adjectives are complementary and can be conceptualised as either dead or alive. Scalarity is a feature of unbounded adjectives, as they express a range on a scale. Bounded adjectives are considered non-scalar, as there are no degrees implied. To illustrate, dead and true are examples of gradable, non-scalar and bounded adjectives while long, good and nasty are examples of gradable, scalar and unbounded adjectives. 2 Although there are theoretical differences between 'relative vs. absolute' adjectives and 'bounded vs. unbounded' adjectives, the two classifications share similar features which complement one another. Accordingly, the following correlations can be made: relative -unbounded adjectives and absolute -bounded adjectives. Therefore, relative adjectives are inherently scalar and unbounded, whereas absolute adjectives are inherently bounded and non-scalar. Claiming that absolute adjectives are non-scalar is not in contradiction with the typology of scale structures suggested in Kennedy and McNally (2005) and Kennedy (2007), as non-scalar here is used in the sense that there are not any intermediate degrees involved along the scale. In other words, absolute adjectives generate closed scales which have a minimum or a maximum degree but there are no intermediate degrees involved along the scale. In order to make the article reader friendly, only the labels relative and absolute adjectives will be used throughout the paper.

Asymmetries
Recent studies have shown that absolute adjectives have an asymmetric behaviour when used in relation to negation, four combinatorial patterns being identified for Swedish (Paradis and Willners 2006). 3 These results indicate that some of the adjectives investigated tend to show scaling potential, 4 posing a challenge to the hypothesis in Kennedy and McNally (2005) and Kennedy (2007), according to which absolute adjectives behave uniformly when used in relation to negation: (2) (a) dead = not alive alive = not dead (b) wrong = not right right ≠ not wrong (c) bound ≠ not free free = not bound (d) empty ≠ not full full ≠ not empty interpretation of negated absolute adjectives in Romanian Art. 7, page 5 of 19 The first pattern highlights absolute adjectives that have an inherently bounded structure and preserve the entailment relation not A = B and not B = A. 5 The two expressions are said to express the exact same meaning, the negator being interpreted as a logical operator which expresses the absolute opposite p versus ¬ p (Paradis andWillners 2006: 1073). In contrast, the fourth pattern shows that neither of the adjectives preserve the entailment relation. Although full and empty have bounded end-points, these adjectives also show scaling potential, i.e. displaying a scale structure. Both patterns in (2b) and (2c) display asymmetries but only at the level of one member. The second pattern shows a symmetric relation between wrong = not right but there is an asymmetry between right ≠ not wrong. In contrast, the third pattern shows a symmetric relation between free = not bound but an asymmetric one between bound ≠ not free. Looking at the adjectives, wrong is a maximum standard adjective while right is a minimum one, and bound is a maximum standard adjective and free is the minimum one, respectively. The two patterns seem to behave exactly the opposite way. Paradis and Willners (2006) discuss the asymmetries between the adjectives in terms of scaling potential, i.e. right and bound are interpreted as being scalar in comparison with the typical bounded adjectives (dead and alive in this case). The authors highlight the configurational complexity of bounded adjectives and point towards their contextual readings, concluding that contextual requirements play a major role in the process of meaning construction.
There have been more theoretical attempts to explain the asymmetric behaviour of gradable adjectives in the literature but they do not seem to explain all the patterns identified in Paradis and Willners (2006). First, the asymmetrical patterns (2b) and (2c) pose problems to the prediction regarding the symmetric behaviour of minimum and maximum standard adjectives in relation to negation (Kennedy and McNally 2005): contrary to the initial prediction, the bi-implication relation holds neither for the minimum item in (2b) nor for the maximum one in (2c). The fact that one item has a maximum and the other one a minimum standard does not seem to be enough to have the entailment relation preserved. Furthermore, the standard of the adjective does not seem to be the factor triggering the asymmetries as in some cases the maximum standard adjectives display scaling potential while in others the minimum standard adjective. Second, pattern (2c) may be the expected outcome if we take into consideration Horn's postulation about the behaviour of unmarked terms with negation: "The negation of a favourable (unmarked) term typically conveys the affirmation of an unfavourable (marked) term, but not necessarily vice versa." (Horn 2014). However, this applies neither to the second pattern where the unmarked term gives rise to the asymmetric pattern nor to the third pattern where the marked term gives rise to a symmetric pattern. To sum up, none of these accounts seem to explain entirely when and why these asymmetries appear.

Preliminaries
When the bi-implication not A = B is preserved the adjectives are interpreted as expressing the same meaning: for instance, a person who is not dead is alive. If the bi-implication can no longer be applied, then the adjectives are said to have some scalar features. As shown in Paradis and Willners (2006), some absolute adjectives may be interpreted as a hybrid between the typical absolute adjectives and the relative ones: they have a 'neithernor' interpretation which indicates that their denotation can be laid out on a scale. interpretation of negated absolute adjectives in Romanian Art. 7, page 6 of 19 While the negative operator not usually denotes the entire negative part of the scale, which may give rise to ambiguous readings between an absolute and a relative interpretation, negative affixes like un-, in-, non-have more precise encodings, denoting different positions on the scale (De Clercq 2013;Horn 1989: 280-3). When the polar opposite B is lexically available, un-and in-denote the outer negative end of the scale, invoking a middle ground. 6 In contrast, non-forms non-scalar compounds, usually creating binary opposites and giving rise to contradictory negation. Additionally, the affixes have different functions: non is used to classify, indicating that the property is absent, while un-and in-are used to characterize, usually invoking a scale between two opposed predicates (De Clercq 2013). In order to investigate the interpretation of a negated absolute adjective nu A 'not A', a survey has been conducted, in which the interpretation of nu A 'not A' is tested against possible non-morphological and morphological equivalents, such as: the polar opposite B, i.e. the lexicalised antonym of A, the affixal negation neA 'unA' and the affixal negation non-A (see Appendix 1, where all these theoretical possibilities are illustrated).
Our assumption is that the weaker sentential negation nu A may be used as a substitute for the stronger option, i.e. the negative prefixes or the lexical antonym. In other words, sentential negation nu A 'not A' may be equivalent in meaning with the polar antonym B or the un-/in-or non-negative compounds, as shown below: (3a) and (3c) suggest that nu A 'not A' is to be interpreted as contradictory negation, (3b) suggests that nu A 'not A' is to be interpreted as contrary negation. The difference between (3a) and (3c) is that the lexical antonym B points to the absolute end of the scale (open -closed) while the non-A compound gives rise to binary opposites (black -nonblack). Although a productive affix in English and in French, non-is not very productive in Romanian, especially when the lexical antonyms B of the absolute adjective is lexicalised.
The following predictions have been tested: • If nu A 'not A' is found to express the same meaning as the polar opposite B, then both of them are absolute adjectives with a conventionally fixed standard. They have an 'either-or' interpretation and they bi-implicate each other when used in relation to negation. • If nu A 'not A' is found to express the same meaning as neA 'unA', then the adjective has a relative-like interpretation and can be displayed on a scale. • If nu A 'not A' is found to express the same meaning as non-A, then the adjective is absolute, leading to an 'either-or' reading and a binary opposition.

Material and procedure
The software used for the online judgement test is SurveyMonkey, 7 an online survey development cloud-based software. A web link was generated and made public on social media for three weeks. 8 The survey was anonymous and no information regarding the partici-6 In comparison with De Clercq (2013) who, following Zimmer (1964), considers the prefixes un-and in-to be allomorphs, i.e. they undergo morphological change due to the adjectival stem to which they attach, Horn (1989: 282-3) distinguishes between them: unlike in-which tends to be assigned a contrary connotation, often involving an opposition to some expected or established norm, un-is situated between the in-and nonforms, depending on how productively or freely the prefix combines with a given base. We remind the reader that for consistency reasons we use the following correlations: contradictory/absolute and contrary/relative. 7 https://www.surveymonkey.com/dashboard/. 8 The study targeted mainly non-linguits and, therefore, the link was available on different parenting groups on social media.
pants' name, age, sex or occupation was demanded. The language used for the questions and answers was Romanian. 66 native speakers of Romanian participated. In the instructions presented before the survey it was clearly stated that the participants needed to be native speakers. In addition, the survey was available on different groups on social media were the only language used was Romanian.
The following question was addressed to the participants: Which of the following answers express in your opinion the same meaning as the following negative utterance: "Subj. is not A"?. All the questions had the same form: the subject (subj.), the verb 'to be' in the present tense (este 'is'), the negative operator (nu 'not'), and the absolute adjective (A). The subject was represented by either nouns (the animal, the answer, the glass, the author, the premise, the writer, the student) or the pronoun he. No other contextual information was provided. The participants could see only a question at a time and they could not continue unless an answer was provided. Given the nature of the study, no distractors were used.
The participants were asked to indicate whether the meaning of the nu A 'not A' structure was represented by the following three options: (a.) the corresponding polar opposite B, (b.) the affixal negation neA 'unA', (c.) the affixal negation non-A. The fourth option is represented by "other; please specify", as shown in Figure 1. The four possible answers were randomised when presented to the participants. If the first three options are meant to be measured quantitatively, the fourth one is designed qualitatively, requesting a written answer in order to collect different possible meanings for the negative utterance. To illustrate, one of the questions is: Which of the following answers express in your opinion the same meaning as the following negative utterance: The animal is not alive?. The following options were presented: it is dead, it is unalive, it is non-alive and other, please specify. Table 1 shows the adjectives that have been tested. The first four pairs (dead/alive, wrong/correct, bound/free, empty/full) are taken from Paradis and Willners' study and five more pairs have been added for the sake of diversity, 18 adjectives being tested in total. In choosing the adjectives, the following conditions had to be met: first, the adjectives should be absolute, i.e. they should have a definite end-point and have a non-scalar meaning and, second, both the adjective and the polar opposite should be equally lexicalised.
If the adjective in the positive form and the polar opposite are both lexically available, the negative compounds present different constraints. For instance, Horn (1989: 274-6) points out the tendency of evaluatively positive or neutral stems to form negatively affixed adjectives. This highlights the availability of only these stems (or of unmarked contradictory adjectives) for negative prefixation: unclean and unsafe are possible but undirty and undangerous are not. De Clercq and Vanden Wyngaerd (2019) discuss a syntactic constraint, according to which two immediately consecutive negative features are impossible in the functional sequence. Following these accounts, only the items in the second column in Table 1 should form negative compounds with ne-'un'. Other factors that play an important role in the derivation of the negative compounds are the degree of lexicalization and the frequency of use. After forming the compounds in order to verify their frequency and degree of lexicalisation (see Appendix 2 where all the adjectives together with their compounds are displayed), we have checked them in the dictionary (https://dexonline.ro/ 9 ), in a corpus of contemporary Romanian language CoRoLa: Corpusul de referință pentru limba română contemporană (http://corola.racai.ro/) and online (https://www.google.com/). As expected, their frequency of use and degree of lexicalisation vary, as follows. First, if the compounds appear in the dictionary, then they are attested forms, fully lexicalised and frequently used. Incorect 'incorrect ', nesănătos 'unhealthy', nevinovat 'unguilty', neadevărat 'untrue', necunoscut 'unknown', neatent 'unattentive' are in this situation. There are three adjectives that are attested as old or archaic words: neliber 'unfree', neînchis 'unbound', nemort 'undead'. There are also two negative adjectives that have polysemantic meanings: negreşit 'undoubtedly' when used as an adverb and nesănătos 'unhealthy' when used with inanimate subjects. However, these should not pose any problems of interpretation, as the questions used in the survey explicitly demand for the adjectival form.
Second, if they appear in written or oral corpora and not in dictionaries, it means that they are used by native speakers but they are rather new, recently formed items. Third, if the compounds only appear online, it means that people use them, but their frequency is reduced in comparison with the ones present in corpora. To illustrate, in addition to the ones having a dictionary entry, neviu 'unalive' and neliber 'unfree' have been found in the consulted corpus and neanonim 'unanonymous', neplin 'unfull', neînchis 'unbound' and nebolnav 'unill' appeared mostly online. Only a reduced number of occurrences (mainly online occurrences) have been found for nemort 'undead', negol 'unempty', nefals 'unfalse', nedistrat 'undistracted'. The occurrences found online are part of literature extracts, names of songs or technical texts. Regarding the compounds with the non-affix, they do not appear in the dictionary and are not very productive. The following have been found online: non-bolnav 'non-ill', non-sănătos 'non-healthy', non-adevărat 'non-true', non-cunoscut 'non-known', non-anonim 'non-anonymous'. On balance, taking a look at all the negative compounds, it appears that 11 compounds from eight adjectives (six neA compounds and five non-A compounds) are used, as follows: • three adjectives can form both neA and non-A compounds: sănătos 'healthy' (nesănătos, non-sănătos), adevărat 'true' (neadevărat, non-adevărat) and cunoscut 'known' (necunoscut, non-cunoscut) • three adjectives form neA compounds: corect 'correct' (incorect), vinovat 'guilty' (nevinovat), atent 'attentive' (neatent) • two adjectives form non-A compounds: bolnav 'ill' (non-bolnav), anonim 'anonymous' (non-anonim) interpretation of negated absolute adjectives in Romanian Art. 7, page 9 of 19 Taking a comparative look at the negative compounds with ne-'un-', these findings reveal that the lexical and syntactic constrains are to some extent overriden by the way in which they are used. According to the former, all the adjectives denoting the absence of some negative or undesirable property can form negative compounds, but it appears that only incorect 'incorrect', neatent 'unattentive', nesănătos 'unhealthy', neadevărat 'untrue' and necunoscut 'unknown' are lexicalised. In contrast, nevinovat 'unguilty', which has a marked stem, is also a lexicalised negative compound.

Results and discussion
Choosing more than one option was possible in the survey. Choosing only one option indicates that there is meaning correspondence between the negative utterance and one of the possible answers. If the participants choose more options, it indicates that they consider that the meaning of nu A 'not A' is found in the combination of these answers. If the option 'other' is chosen, it indicates that they consider the intended meaning of nu A 'not A' to be different from the given possibilities. Unlike previous studies that propose participants a scale, this survey is designed as a 'yes or no' question, where each answer stands for a different interpretation. In this way, the scaling potential is measured indirectly.
The study is designed so that for every question six answer types are available -100, 010, 001, 110, 101, 011, where 1 stands for the chosen answer and 0 for the answers that have not been chosen: • 100: only the polar opposite B has been ticked • 010: only neA 'unA' has been ticked • 001: only non-A has been ticked • 110: both the polar opposite B and neA 'unA' have been ticked • 101: both the polar opposite B and non-A have been ticked • 011: both neA 'unA' and non-A have been ticked Although there were 66 participants in total for each question, the number of answers varies depending on the question, as the following answers have been disregarded due to inconsistency: the answers where all the options have been chosen and the answers where the participants have chosen 'other, please specify' and one of the three options. Table 2 summarizes the counts for all the answers each adjective has received, as well as the number of the discarded answers. While there are some peaks, none of the questions has received only one type of answer.
In order to interpret the significance of the different distributions of the answer types and to identify the profiles of the adjectives, the Principal Component Analysis (PCA) has been carried out and the Rstudio software and the FactoMineR and Factoshiny packages have been used. PCA is used to study the similarities between individuals by taking into account all the variables. It identifies the individuals' profiles by summarizing a data table where the individuals are described by (continuous) quantitative variables (Cornillon et al. 2012: 209). PCA is a factor analysis in the sense that the principal axes resulted are linear combinations of the initial variables, which are in turn ordered and independent. The strength and importance of PCA is given by this independence of the new variables which were highly correlated in the beginning. In other words, PCA makes it possible to reduce the large number of explanatory variables to three or four components and to summarize the mass of information present in the dataset in a visual way.  Applied to the present study, this analysis helps us identify clusters of adjectives and create their profiles. PCA is applied to the information displayed in Table 2, where the rows are the variables (i.e. the possible answer) and the adjectives represent the population. The intersection of a column with a row indicates the number of participants that gave that particular answer to that particular question. For example, if we look at the intersection between sănătos 'healthy' and the column 101 we have the number 1 which means that one person has given the answer 101 to this question. PCA reduces the number of dimensions (the six possible answers in our case) to two or three. Having more concise data to examine, similarity patterns can be more easily found between the adjectives.
The last two answer types can be easily dismissed as the numbers are very small: <2% for 011 and <3% for 101. This means that the participants considered that the intended meaning for the negative utterance nu A 'not A' is represented neither by the answer type 101 (the combination between the polar opposite B and the affixal negation non-A) nor by the answer type 011 (the combination between the affixal negation neA 'unA' and the affixal negation non-A). Therefore, the profiles of the adjectives investigated are a result of the different combinations of the four answer types: 100, 110, 010, 001.
Dim1 and Dim2 in the Variables factor map illustrated in Graphic 1a explain 96.6% of the variance of the information, which means that if all the information is projected to these two components 96.6% of the data is preserved. The variables (i.e. the answers types) are represented by the arrows. The angle between two arrows or between an arrow and an axis gives the correlation of the answer with the axis or dimension: the smaller the angle the bigger the correlation (positive or negative). Orthogonality means no correlation whatsoever (for instance, the answer types 100 and 001 are orthogonal, i.e. they have no correlation, and therefore they are independent).
Dim 1 and Dim 2 indicate the different contributions of the answer patterns to the profiles of the adjectives. Dim1 is positively correlated (value 0.99, p-value < 0.01) with the answer 100 and negatively correlated (value -0.96, p-value < 0.01) with the answer 010. Therefore, the positive part of the axis Dim1 is to be interpreted as the answer 100 and the negative part of the axis as the answer 010. The orientation of the arrows points out towards a clear delimitation between these two answer types: an adjective which has many answers 100 has very few 010 and an adjective that has many answers 010 has very few answers 100. This leads linguistically to a delimitation between an 'absolute' and a 'relative-like' region.
Dim1 is also positively correlated (value 0.68, p-value < 0.01) with the answer 101 and negatively correlated (value -0.74, p-value < 0.01) with the answer 110. The contribution of 101 to Dim1 is very small (<0.2%), so it can safely be ignored. The answer type 110 is situated in the relative-like region and is partially opposed to the answer type 100. This means that the answer type 110 is to be interpreted linguistically as having a 'weak relative-like' reading because of the absolute features in its profile, in comparison with the answer type 010 which has 'strong relative-like' reading, as highlighted in Graphic 1b.
Dim2 is positively correlated with 110 (value 0.67, p-value < 0.01) and negatively correlated with 001 (value -0.75, p-value < 0.01). This means that Dim2 opposes some absolute features as indicated by the answer 001 (i.e. affixal negation non-A) and some relative-like features, as indicated by the answer type 110 (i.e. the weak relative-like interpretation). Building on these distinctions, in order to make the discussion more reader friendly, the numeric answer types will be replaced with some more meaningful notions, as follows: 100 with ABS, 010 with strong REL, 110 with weak REL and 001 with reinforced ABS. Concerning the interpretation of the adjectives vs. possible answer, the following remarks can be made: if an adjective is plotted close to the direction of an arrow (for instance, the answer type strong REL and vinovat 'guilty') that means that the adjective has received many of these answers. In other words, the answer type strong REL shapes the profile for vinovat 'guilty'. The following graphic shows the adjective map: The projection of the adjectives on the axis Dim1 gives the information about the answer types ABS and strong REL. The more the adjective is distant from the centre (either to the left or to the right), the larger is the difference between the answer types ABS and strong REL. The closer it is to the centre, the closer are the counts for the answer types ABS and strong REL. The projection of the adjectives as a whole on Dim1 is much bigger than the projection of the adjectives on Dim2 which means that Dim1 captures most of the information, i.e. the answer types ABS and strong REL give the core of the profiles for most of the adjectives. This is the reason why the cloud of adjectives is elongated along Dim1 and rather narrow on Dim2. However, there are some adjectives which are represented on all three dimensions but with different distributions (for instance, gol 'empty' and adevărat 'true') and their interpretation will require a deeper analysis to describe their profile (see Appendix 3 for details).
The adjectives positioned to the right of the axis Dim1 will be defined by the answer type ABS, i.e. they will have an absolute profile in the form of the polar opposite B. The following adjectives are in this situation: mort 'dead', viu 'alive', bolnav 'ill', distrat 'distracted', liber 'free', anonim 'anonymous', fals 'false', închis 'bound'. The more the adjectives are placed to the extremity ABS of the axis, the bigger the difference between ABS and strong REL will be. This means that mort 'dead' has received many ABS answers and very few or no strong REL answers at all. Given the fact that the adjectives are elongated on the right part of the axis Dim1, a further distinction among mort 'dead', viu 'alive', bolnav 'ill', distrat 'distracted', liber 'free', anonim 'anonymous', fals 'false', închis 'bound' is to be made: if mort 'dead' can be interpreted as an absolute adjective with no scaling potential, închis 'bound' shows some scaling potential, as it is placed nearer to the relative region and its profile is defined by a combination of ABS and strong REL answer types. Therefore, mort 'dead', viu 'alive', bolnav 'ill', distrat 'distracted', liber 'free', anonim 'anonymous', fals 'false', închis 'bound' can be further classified according to their scaling potential, ranging from no scaling potential to little scaling potential, as follows: • mort 'dead' is the only adjective that has an absolute reading by excellence and does not seem to show any scaling potential. • viu 'alive' and bolnav 'ill' display similar patterns and indicate a strong absolute reading with very little scaling potential. • distrat 'distracted', anonim 'anonymous' and liber 'free', and fals 'false' show little scaling potential. The slightly different position they occupy on the axis Dim1 is given by the contribution the other answers make to their profiles. • închis 'bound' is an absolute adjective that displays some scaling potential.
The adjectives positioned to the left of the axis Dim 1 will be defined by the answer types strong REL and weak REL, respectively. The following adjectives are in this situation: vinovat 'guilty', atent 'attentive', corect 'correct', cunoscut 'known'. If vinovat 'guilty' can be interpreted as a strong relative-like adjective, the profile of atent 'attentive' is a combination between the two (hence its position between the strong REL and weak REL arrows in Graphic 2). The fact that atent 'attentive', corect 'correct' and cunoscut 'known' have the answer type ABS in their profiles situates them higher in the weak relative-like region.
There are also some adjectives positioned either in the absolute or in the relative-like region but farther away from the axis Dim 1. Sănătos 'healthy' and greşit 'wrong' are in this situation. They are on the absolute side, but sănătos 'healthy' is placed higher than the rest of the adjectives, which means that it displays some scaling potential (given by weak REL answer type), while greşit 'wrong' is placed lower in the absolute region. This means that greşit 'wrong' has a reinforced absolute profile, given by the combination of ABS and reinforced ABS answer types in its profile.
In contrast, adevărat 'true', inocent 'innocent', plin 'full' and gol 'empty' are on the relative-like side. Inocent 'innocent' and plin 'full' are found in the graphic very close to the intersection between ABS, strong REL, reinforced ABS and weak REL, which means that the two adjectives can be considered either absolute or relative-like depending on the context in which they are used. There are two adjectives which seem to be represented on all dimensions: adevărat 'true' and gol 'empty' are in this situation. A different analysis is necessary in order to identify their profiles.

Concluding remarks and future directions of research
In light of recent results, absolute adjectives are said to behave asymmetrically, as some adjectives have, in addition to the bounded structure, some scalar features. The present online judgement test was designed in order to investigate the interpretation of absolute adjectives in relation to negation in Romanian. The data collected was analysed using the Principal Component Analysis (the Rstudio software and the FactoMineR and Factoshiny packages). The two graphics -the answer map and the adjective map -summarised and presented visually the profile of 18 adjectives, forming nine pairs: mort-viu 'dead-alive ', liber-închis 'free-bound', plin-gol 'full-empty', greşit-corect 'wrong-right', bolnav-sănătos 'ill-healthy', vinovat-inocent 'guilty-innocent', adevărat-fals 'true-false', cunoscut-anonim 'known-anonymous', atent-distrat 'attentivedistracted'. Four answer types contributed to the profiles of the adjectives: ABS, weak REL, strong REL, reinforced ABS. The two graphics showed that the adjectives clustered differently based on their profiles. While Dim1 was defined by the answer type ABS on the positive part and by the answer types strong REL and weak REL in the negative part, Dim2 was defined by the answer type weak REL on the positive part and the answer type reinforced ABS on the negative part. On the answer map, it can be seen that Dim1 opposes the two possible answers ABS and strong REL, which means that they divide the graphic in two sides according to their profile: an 'absolute' side (where the absolute interpretation comes from choosing the polar opposite B) and a 'relative-like' side (where the relative-like interpretation comes from choosing neA 'unA'). Furthermore, the relative-like side can be divided into 'strong relative-like' and 'weak relative-like', because on the left side of the graphic the answer type weak REL is generated (which is the combination between the polar opposite B and neA 'unA'). Therefore, the answer type strong REL is to be interpreted linguistically as having a 'strong relative-like' interpretation while the answer type weak REL is to be interpreted linguistically as having a 'weak relative-like' reading given the fact that it has some absolute features in its profile.
The adjective graphic shows that most of the adjectives are positioned along the axis Dim1, which means that their profiles are determined mostly by the answer types ABS and strong REL. The resulted correlations for each axis indicate that there is no clear-cut difference between an absolute and a relative-like interpretation but rather a continuum between them. The two graphics suggest a classification of the adjectives in question in three main categories: absolute adjectives, relative-like adjectives and adjectives that can go in either category: (1) absolute adjectives, which suggest scaling features, ranging from displaying no scaling potential to having scaling potential: mort 'dead', viu 'alive', bolnav 'ill', distrat 'distracted', anonim 'anonymous', liber 'free', fals 'false', închis 'bound' and sănătos 'healthy'. Greşit 'wrong' is a reinforced absolute adjective, defined by the absolute polar opposite B and the non-A affixal negation.
(2) relative-like adjectives, where a distinction was made between strong and weak relative-like adjectives. Vinovat 'guilty' belongs to the first category while atent 'attentive', corect 'correct' and cunoscut 'known' are relative with absolute features (given by the answer pattern ABS in their profiles).
(3) adjectives that can take either an absolute or a relative-like reading: inocent 'innocent' and plin 'full' belong to this category. For adevărat 'true' and gol 'empty' the analysis was inconclusive. Their profiles are to be further investigated using different methods and statistical analyses.
In sum, the results highlight variation among the absolute adjectives as a class, variation inside the subclasses that have been identified and variation at the level of individual adjectives. First, absolute adjectives appear to be more hybrid than initially thought. Absolute adjectives were expected to behave uniformly but some of them appeared on the relative-like region, which leads to the following classification: absolute -relative-like -relative. In contrast with relative adjectives, the relative-like adjectives have scalar features in addition to the bounded structures. Second, the results also indicate that there is variation in the profiles of the adjectives at the level of each category. For instance, the adjectives labelled as 'absolute adjectives' are displayed on a continuum, ranging from having no scaling potential to having scaling potential. In other words, the results show different degrees of scalarity which vary form one adjective to the other.
Third, the adjectives seem to have mixed profiles as a result of the different distributions of the absolute and relative-like features in their profiles. The source of this variation cannot be accounted for at this stage. It may be contextual, as a result of the lexical associations generated by the subject and the adjective or it may be due to the multiple contexts of interpretation the utterances generate. For instance, an utterance like 'The animal is not free' does not provide explicit information about which animal the utterance refers to or the intended meaning of 'free'. Consequently, the participants may use scenarios to interpret this utterance which modify the reading of 'free' (The animal is not free enough to run or The animal is not free for a horse of its size etc.).
Contrary to expectations, the adjectives investigated do not seem to function in pairs but rather independently. Apart from mort 'dead' and viu 'alive' which seem to behave similarly, the rest of the adjectives do not display similar profiles with their antonym. More situations have emerged: 1. both members of the pair are in the absolute region but they show differences at the level of the scaling potential: sănătos 'healthy', bolnav 'ill'; închis 'bound', liber 'free'; 2. both members are in the relative-like region but they seem to have different degrees of scalarity: vinovat 'guilty', inocent 'innocent'; 3. the members are in opposite regions: anonim 'anonymous', cunoscut 'known'; corect 'right', greşit 'wrong'; atent 'attentive', distrat 'distracted'. The fact that the members of the pairs do not have similar profiles leads to the asymmetric patterns when they are negated.
In line with the results in Paradis and Willners (2006), no clear patterns can be identified with regard to the distinction between the 'maximum' and the 'minimum' standards adjectives. There seems to be a tendency (but there are also exceptions) as most of the adjectives characterised as the absolute adjectives belong to the minimum category (except mort 'dead', liber 'free' and sănătos 'healthy'), while most of the adjectives characterised as being relative-like seem to have maximum standards (except vinovat 'guilty'). In order to clearly state whether the distinction between maximum vs. minimum can be considered a significant difference and a reliable criterion, further investigations need to be conducted.
Three predictions have been formulated in section 2.1 and some remarks on the degree of lexicalisation and frequency of the negative compounds have been made in section 2.2. The predictions are not borne out for all the adjectives as a class, but they can be discussed based on the three categories of adjectives. The first prediction (If nu A = B, then the adjectives are absolute) is validated by the category of adjectives which were interpreted as having an absolute profile with little or no scaling potential. The relation of bi-implication can be applied to the following adjectives mort 'dead', viu 'alive', where the negation of one term entails the other one. The bi-implication relation can no longer be applied if some scaling potential is found in the profile of the adjectives. In other words, it is the different degree of scaling potential the members of the pairs have that leads to asymmetries.
The second prediction (If nu A = neA, then the adjectives have relative-like features) is also borne out by the adjectives found to have a relative-like profile (vinovat 'guilty', atent 'attentive', corect 'correct', cunoscut 'known'). The first three adjectives were expected to have a relative-like profile due to their high degree of lexicalisation and the frequency of use of the neA compounds.
The third prediction (If nu A = non-A, then the adjectives are absolute) is not borne out by the results. The answer type reinforced ABS was found statistically relevant (as it contributed to Dim2), but the extent to which this answer type shapes the profile of the adjectives is usually very reduced. Only the adjective greşit 'wrong' appears to have a reinforced absolute profile, i.e. its profile is given by the answers type ABS and reinforced ABS. Interestingly, the answer type reinforced ABS does not seem to influence the profile of the adjectives that form non-compounds (non-anonim 'non-anonymous' and non-bolnav 'non-ill'), as these appear to have an absolute interpretation with little scaling potential. The lexicalisation of the polar antonym B (which was the prerequisite for this study) has to some extent overridden the non-A compounds.
Two of the three adjectives that have compounds with both negative prefixes are found to have different profiles: cunoscut 'known' is found in the weak relative-like region while sănătos 'healthy' is found in the absolute region. Due to the fact that they are close to the centre, they share features with the opposite class: cunoscut 'known' is a weak relativelike because of the absolute features in its profile and sănătos 'healthy' is absolute with some relative-like features (responsible for its scaling potential). However, there does not seem to be a too big correlation between the tendency of the adjective to form negative compounds in natural language and the profiles of the adjectives. The results show that the profile of the adjectives are the results of the combination of the answer patterns regardless of the fact that the adjectives in question have lexicalised negative compounds or that the stems are positive or negative. In other words, these results do not support the theoretical claim according to which negative affixes are not used with adjectival stems that have a 'negative' value. To illustrate, taking into account the lexical and syntactic constraints, vinovat 'guilty' is not expected to form a negative compound with ne-'un-', but this appears to be very productive. In fact, the results show that vinovat 'guilty' is the only strong relative-like adjective. One may think that, due to its high degree of lexicalization, vinovat 'guilty' and nevinovat 'unguilty' form an antonymic pair. However, the initial pre-testing shows that nevinovat 'unguilty' is not rated as having the same meaning with inocent 'innocent', occupying a lower position on the scale vinovat 'guilty' -innocent 'innocent'. Another example is represented by neinocent 'uninnocent' which does not have and is not expected to form a negative compound, but it has received many answers of this type. It may be the case that the participants have supplied the non-lexicalised compounds in order to fill in the gap between nu A 'not A' and the polar opposite B.
The present pilot test has revealed some novel findings in shaping the profiles of absolute adjectives. It has shown that the adjectives can be grouped in different categories based on their profiles. However, there is variation within these categories, which needs to be further investigated. Follow-up studies using more data and different statistical methods are necessary in order to further investigate the categories found, in general, and the division between weak and strong relative-like adjectives, in particular. One possibility would be to increase the number of the absolute adjectives under investigation in order to see whether the profiles can be more clearly shaped. Another possibility would be to replicate the test in different languages (for instance, English, French, German) to see whether the asymmetries found in the behaviour of absolute adjectives are a language specific or a cross-linguistic phenomenon.