SEMANTIC RELATIONS AMONG ADJECTIVES IN POLISH WORDNET 2.0: A NEW RELATION SET, DISCUSSION AND EVALUATION

Adjectives in wordnets are often neglected: there are many fewer of them than nouns, and relations among them are sometimes not as varied as those among nouns or verbs. Polish WordNet 1.0 was no exception. Version 2.0 aims to correct that. We present an overview of a much larger set of lexical-semantic relations which connect adjectives to the other parts of the network. Our choice of relations has been motivated by linguistic considerations, especially the concerns of the Polish lexical semantics, and by pragmatic reasons. The discussion includes detailed substitution tests, meant to ensure consistency among wordnet editors.


Introduction
A wordnet is an electronic thesaurus in the form of a graph.Its nodes are lemmas -word/sense pairs -called lexical units (LUs); its edges are lexico-semantic relations.A set of constitutive relations determines the shape of the wordnet.The choice of such relations is a fundamental design decision; they are very likely to include some forms of synonymy, antonymy, hypernymy, hyponymy, holonymy and meronymy, but beyond that the lexico-semantic makeup of the language is a central consideration.Relations are central in wordnet construction: they help distinguish the senses of a lemma by linking its word with its "semantic neighbours".Groups of LUs which share all constitutive relations with the rest of the graph (they are essentially near-synonyms) are called synset.We say that a relation is a synset relation if it can be shared among groups of lexical units; lexical unit relations have more individual patterns.
The "mother of all wordnets" is Princeton WordNet (PWN), designed at Princeton University in the late 1980s on the psycholinguistic and computational prin-ciples.After its emergence, research groups around the world have set to work on wordnets for their own languages.Early, and influential, systems include Eu-roWordNet (EWN) and GermaNet (GN).All three have been an inspiration for Polish WordNet (plWordNet)1 ; the first phase of the project, [2005][2006][2007][2008][2009], resulted in version 1.0 with 26990 lexical units -18611 nouns, 4498 verbs and 3881 adjectives.
The choice of semantic relations for plWordNet 1.0 was guided by wordnetbuilding tradition, theory of lexical semantics and lexicographic practice (Derwojedowa et al. 2008).Most relations were taken over from PWN and EWN, not the least for compatibility. 2 The set included antonymy, hypernymy, meronymy, conversion, two derivational relations -more regular relatedness (transpositional, syntactic derivation) and less regular pertainymy -and fuzzynymy which captures irregular, infrequent semantic phenomena (Piasecki et al. 2009).
Adjectives in plWordNet 1.0 have been not so much neglected as left for future work.This is in keeping with how creators of new wordnets apportion attention to parts of speech (invariably, nouns are by far the most numerous).Among the 3881 LUs in 1160 synsets there were only 155 instances of hyponymy, 1618 of antonymy, 1226 of relatedness, 295 of pertainymy and 423 of fuzzynymy.Table 1 shows the "relation density" of plWordNet 1.0: the average number of synset relations per synset and LU relations per LU.The scarcity of synset relations clearly suggests that we need many more adjectival relations than the minimal set in plWordNet 1.0.This has been the latest concern in the ongoing plWordNet 2.0 project, and it is the focus of this paper.We will present a much-improved set of lexico-semantic relations among adjectives, accompanied by substitution tests and examples.
There have been several rather divergent takes on adjectives in PWN, EWN and GN.It is clearly not easy to describe adjectives in general, and it is distinctly hard to come up with a language-independent relation set.Table 2 sums up the matching relations.Clearly, analogies are imperfect and each of the wordnets has relations not found elsewhere.
Our decision on the set of lexico-semantic relations among adjectives in plWord-Net 2.0 was again based on three central criteria (Derwojedowa et al. 2008): solutions in influential wordnets (mainly PWN and EWN, sometimes GN), lexicosemantic reality of the Polish language, and established lexicographic custom.For every relation we seek motivation in the literature.Sections 2-7 present the main groups of relations, along with a relevant discussion.The lexicographic criterion was implemented via the sampling of the Universal Dictionary of Polish Language (UDP) (Dubisz 2004).We chose 100 adjectives by randomly drawing a volume number and page number, and taking the first adjective on the page (if there was none, we took the first adjective from the next page). 4he selected adjectives were described with the appropriate substitution tests and equipped with relations to other lexical units.Table 3 shows the whole set.The statistics of the relations in the UDP sample appear at the end of the paper.In plWordNet 2.0, pointers to adjectives go from verbs (state, cause and process relations, see Maziarz et al. 2011b) and from nouns (cross-categorial synonymy, formations ending with -ość, see Maziarz et al. 2011).In this paper, we are only interested in pointers from adjectives.Table 3, unlike Table 2, shows a few relations not described in this paper.Derivational relations in Table 2 are divided into groups which match the similar phenomena in other wordnets.Table 3 groups them under the label derivational relations.

Hyponymy among adjectives
"In natural language, hyponymy is loosely defined as the 'kind of' relation": a horse is a type of animal and a martagon is a kind of lily (Murphy 2003: 217-218).The notion of hyponymy among adjectives, however, is not so clear.According to lexicologists and semanticists (Cruse 1997: 141-3) hyponymic taxonomies for adjectives are shallow.There exist proper hyponym pairs, such as red -maroon «of a dark reddish purple colour» (Croft, Cruse 2004: 142) or large -huge «extremely large» (Atkins, Rundell 2008), but for many adjectives a superordinate is quite hard to find.
Many adjectives simply have nouns as superordinates.Round , square, oval are kinds of shape, sweet, bitter , sour or salty are kinds of taste, happy, sad , angry are kinds of emotions (Lyons 1977;Murphy 2003: 218;Murphy 2009), while hot, warm, cool , cold relate to temperature (Maienborn, von Heusinger 2011: 460).Not all colour adjectives have an adjectival hypernym in English (Lyons 1977: 298). 5 The same is true for Polish.There is, for example, no hypernym for czerwony «będący koloru krwi lub dojrzałego pomidora» 'red «the colour of blood or ripe tomato»' UDP, ODP.No such word exists; the adjective kolorowy '± colourful' has a specific meaning, so it rather is the noun kolor 'colour' which serves as a hypernym of czerwony (Lyons 1977: 298).For the adjective głodny 'hungry', the UDP gives the definition «odczuwający głód, pragnący pożywienia» '«feeling hunger, craving food»', while ODP defines it using a verb phrase: «Komuś, kto jest głodny, bardzo chce się jeść» '«Someone who is hungry very much wants to eat»'.No obvious adjective candidate for hypernym comes to mind, 6 but a noun can be easily proposed, for example potrzeba '(physical) need'.Adjectives, then, may be naturally defined by nouns.Lyons calls such a relation quasi-hypernymy (1977).
Polish WordNet 2.0 will include both adjectival hyponymy and quasi-hyponymy; the latter relation is named value of the attribute, a borrowing from Princeton WordNet's attribute relation ("WordNet contains pointers between descriptive adjectives and the nouns by which appropriate attributes are lexicalized", see Miller 1998: 48). 5 [T]here is no superordinate term in English of which all the colour-words are co-hyponyms.(Logicians frequently cite as an example of analytic implication If it is red, then it is coloured.But this implication does not in fact generally hold for all colour-terms in normal English usage.The adjective coloured is in contrast with white in certain contexts -in sorting out the laundry, in the classification of people according to their race, etc. -and with transparent in others: e.g.There was some colored liquid in the bottle -one might also wander whether coloured is in contrast with white, as well as with transparent, in contexts of this kind.)(Lyons 1995: 456).
Hyponymy between adjectives is absent in PWN (Miller 1998) and in EWN, but it was introduced into the GN structure (Hamp, Feldwag 1997).Adjectival hyponymy also appears in RussNet (Azarova, Yavorskaya 2010: 217-8).Murphy (2003: 222) proposes the following test for adjectival hyponymy: • To be excellent is to be good to a certain degree.
Note that such a test is strongly connected with gradable adjectives: the two adjectives occupy a predicative position (be + Adj).Croft and Cruse (2004: 142) propose another test suitable for both gradable and relative adjectives (occurring in the attributive position): • The class of scarlet things is a subset of the class of red things.
We merge both tests for use in Polish WordNet 2.0: Quasi-hyponymy (Lyons 1977) sometimes appears in wordnets.In PWN, adjectives are linked to nouns via the attribute relation if they express the values of an attribute, e.g., qualitysuperb, great, good , mediocre, bad , awful , atrocious (Miller 1998: 50-53).In EWN, the XPOS_HYPONYM relation is used "to relate nouns that head a class of adjectival values", e.g., colourblack , white, blue, green, yellow , red (Vossen 2002: 24), but tests are not easily found in the EWN documentation.The relation also appears in RussNet (Azarova, Yavorskaya 2010: 217-218), for example, the adjective кислый 'acidic' is a noun-like hyponym of the noun вкус 'taste'.RussNet constructs the following tests for the two words: • The fruit is acidic implies that the fruit has (certain) taste.
• The fruit has certain taste does not imply that the fruit is acidic.
Note that this test only works for gradable adjectives.We create a similar test for our value of the attribute relation.
Test II.Value (of the attribute), Adj ¡ N • X to przymiotnik, N , P to (stosowne) rzeczowniki 'X is an adjective, N , P are the (suitable) nouns'.
• Jeżeli ktoś/coś jest X, to ma określony N 'If someone/something is X Adj , then he/she/it has a specific N '.
• Jeżeli ktoś/coś ma określony N, to niekoniecznie jest X 'If someone/something has a specific N , then he/she/it need not be X Adj '.
• Jeżeli ktoś/coś jest X P, to ma określony N 'If someone/something is X Adj P , then he/she/it has a specific N'.
• Jeżeli P ma określony N, to niekoniecznie jest X P 'If P has a specific N, then it need not be X Adj P '.
Jeżeli coś jest czerwone, to ma określony kolor 'If something is red, then it has a specific colour'.
Jeżeli coś ma określony kolor, to niekoniecznie jest czerwone 'If something has a specific colour, then it need not be red'.
Jeżeli coś jest czerwonym kapeluszem, to ma określony kolor 'If someone/something is a red hat, then it has a specific colour'.
Jeżeli kapelusz ma określony kolor, to niekoniecznie jest czerwonym kapeluszem 'If a hat has a specific colour, then it need not be a red hat'.
Complementary antonyms are usually defined by entailment: ∼X ¡ Y and Y ¡ ∼X, where X and Y are lexical units (Lyons 1995b: 401, Kreidler 1998: 104-105;Lyons 1995: 128-129;Lyons 1977: 271-272, 279-280).Examples of complementaries are married / single and dead / alive: someone cannot be married and single at the same time; it is impossible for someone to be neither dead nor alive.
For gradable antonyms only the latter implication holds: Y ¡ ∼X (Lyons 1995b: 466-467).Examples of gradable antonyms are good / bad and red / blue (as well as green and so on): something cannot be both good and bad; it is quite possible for things or beings to be neither red nor blue (green and so on) (Lyons 1995: 461).Antonymy is recognized also within the Sense £ Text paradigm with the stipulation that Anti 1 type, i.e., direct opposites, corresponds to complementary antonymy, and Anti 2 and Anti 3 of contradictory meaning types refer to gradable antonymy (Mel'čuk 1996: 47-9;Apresjan 2000: 269, 273-277;Wanner 1996: 8).
It should be noted that the entailment Y ¡ ∼X holds not only for antonyms but also for other words which share the incompatibility of meaning, for instance, co-hyponyms, such as blue and green, or unrelated words, like overture and bottle, justice and cat (Lyons 1981: 154-155;Murphy 2009: 25;Murphy 2003: 249-250;cf. Griffiths 2006: 52-54).Antonyms have a lot in common; an antonym negates only part of the meaning of its counterpart (Apresjan 2000: 269, 270, 273).The compatibility of senses between antonyms comes from the fact that they have a common hypernym (Lewandowska-Tomaszczyk 1998: 124).
Antonymy is one of the most prominent adjective relations in PWN.Relations in PWN are psycholinguistically motivated: "The importance of antonymy first become obvious from results obtained with word association test: when the probe [i.e., an adjective given to a respondent] is a familiar adjective, the response commonly given by adult speaker is its antonym.For example, to the probe good , the common response is bad ; to bad , the most frequent response is good " (Miller 1998: 48).Descriptive adjectives which express the value of the same attribute are divided into antonymic clusters.
X and Y are antonyms if they are co-hyponyms or co-meronyms.This condition guarantees that they have a common semantic element.Our tests include yet another condition, after Cruse's proposal: "[N]ot all lexical items are felt to have opposites.Ask someone for the opposite of table, or gold , or triangle, and he will be unable to oblige.Some lexical items, it seems, are inherently non-opposable" (Cruse 1997: 257-8).We have decided to check for every pair whether it is an example of opposition: • Is he/she/it X? -No, on the contrary: he/she/it is Y.This helps distinguish antonyms from co-hyponyms.Here is the whole test.

TEST III. Complementary antonyms
• X, Y to przymiotniki 'X, Y are adjectives'.
• Jeżeli ktoś/coś jest X, to nie może być Y 'If someone/something is X, then he/she/it cannot be Y'.
• Jeżeli ktoś/coś nie jest X, to musi być Y 'If someone/something is not X, then he/she/it must be Y'.
Jeżeli ktoś jest martwy, to nie może być żywy 'If someone is dead, then he/she cannot be alive'.Jeżeli ktoś nie jest martwy, to musi być żywy 'If someone is not dead, then he/she must be alive'.

TEST IV. Gradable antonyms
• X, Y to przymiotniki 'X, Y are adjectives' • X i Y muszą być kohiponimami lub kowartościami tej samej cechy 'X, Y must be co-hyponyms or co-values of the same atrribute' Converseness is another kind of semantic opposition within which the interrelated words have reversible characteristics: they become synonymous when one changes their actants.For example, A is a husband of B implies that B is a wife of A. Similarly, A sells B to C means that B buys C from A (Lyons 1977: 274;Lyons 1995: 467-469;Malmkjaer 341;Apresjan 2000: 241-265;Kreidler 1998: 97;Kempson 1996: 85;Wanner 1996: 8-9).
Jeżeli A jest lepszy niż B, to B jest gorszy niż A 'If A is better than B, then B is worse than A'.
Jeżeli B jest gorszy niż A, to A jest lepszy niż B 'If B is worse than A, then A is better than B'.

Gradation
We have seen that lack of complementarity (defined by the implication X ¡ ∼Y) is the property not only of antonyms but also co-hyponyms (bluegreen) or even unrelated words (dogtree).To Lyons, it seems that the paradox could be resolved with the notion of gradability rather than truth values of implication (Lyons 1977;Miller 1998: 52-53

TEST VI. Gradation
• X, Y to przymiotniki będące wartościami cechy N 'X, Y are adjectives which express values of attribute N'.
• Przymiotniki X i Y nie wyrażają podobnego natężenia cechy N 'adjectives X and Y do not express similar degrees (intensity) of attribute N'.8 • Przymiotnik X wyraża większe natężenie cechy N niż przymiotnik Y 'adjective X exhibits exptesses a intensity of attribute N than Y'.
• Żaden inny przymiotnik nie może zastąpić ani X, ani Y w poprzednim zdaniu 'No other adjective can replace X or Y in the preceding statement'.
Adjectives gorący i ciepły do not express similar degrees of attribute TEMPE-RATURA.
Adjective gorący exhibits higher intensity of the property TEMPERATURA than ciepły.
No other adjective can replace gorący or ciepły in the preceding statement.
The example is supported by lexicographic data.The UDP defines gorący as «mający wysoką temperaturę, bardzo ciepły, silnie nagrzany, rozpalony» 'having high temperature, very warm, strongly heated, glowing'.ODP gives a definition relativised to the scale: «Coś, co jest gorące, ma temperaturę wysoką, bardzo wysoką lub wyższą od normalnej» 'Something that is hot has high temperature, very high or higher than average'.For ciepły the dictionary has following definition: «Coś, co jest ciepłe, ma temperaturę wysoką, ale niezbyt wysoką, albo taką, jaką zwykle mają rzeczy tego rodzaju» 'Something that is warm has a high temperature, but not very high, or such temperature as is common for things of that sort'.Both dictionaries agree that gorący has a higher intensity of TEMPERATURE than ciepły.

Distributional properties
It is quite a common situation when an adjective serves as a modifier of a limited group of nouns.A case in point are Polish adjectives which describe equine coat colours -see examples in Table 4.The adjective kary 'black (with regard to a horse)' has its synonym czarny 'black'.They differ in their distributional properties.The usage of kary is limited to nouns denoting horses, the applicability of czarny is much wider.

Table 4. Examples of equine coat colours
kary 'black' «o maści konia: czarny; wrony» UDP 'of horse coat: black; like crow's' dereszowaty 'roan' «o koniu lub jego maści: właściwy dereszowi, charakterystyczny dla deresza» UDP 'of horse or its coat: characteristic of roan' gniady 'bay' «o maści konia: jasnobrązowy z odcieniem czerwonobrunatnym, z czarną grzywą, czarnym ogonem i czarnymi dolnymi częściami kończyn» UDP 'of horse coat: light brown with a red-brown tint, with a black mane, black tail and black lower parts of limbs' kasztanowaty 'maroon' «o koniu lub jego maści: brązowożółty» UDP 'of horse or its coat: brown-yellow' The same property is exhibited by adjectives piwny '∼hazel', pierworodny 'firstborn' or podkasana '± frivolous'.They point only to the specific types of objects, animals and people.5 illustrates.The following test will be used to check whether an adjective modifies a given noun or a noun class.If a pointer of the relation links the adjective with a certain noun, it also means that all its hyponyms may be modified with the adjective.Note that only gradable adjectives pass the test.

TEST VII. Modifier
• X to przymiotnik, N, P to rzeczowniki 'X is an adjective, N and P are nouns'.
• X wchodzi w relację wartości cechy z P 'X is a value of attribute P'.
• Mówi się o N, że jest X w odniesieniu do jego/jej P 'It is said about N that it is X with regard to its/his/her P'.
• Jeżeli N jest X, to ma określone P 'If N is X, then it has a specific P'.
• Jeżeli N ma określone P, to niekoniecznie jest X 'If N has a specific P, then it/he/she need not be X'.

Inter-register synonymy
Lexical units with a very similar denotation (or even true synonyms) often cannot be used in the same contexts -for stylistic reasons or because of differences in connotation.Yet if the LUs are very close in meaning, we want them to be somehow interchangeable, just as proper near-synonyms are.The construction of plWordNet relies on grouping into synsets LUs with the same "topology" in the network: with the same links via instances of constitutive relations.
Thus for example automobil 'automobile' (a marked LU, labelled in the UDP as «przestarzały» 'obsolete') and samochód osobowy 'passenger car' have different relations, but almost identical meaning.Samochód osobowy has a large set of hyponyms (placed in 32 synsets), whereas automobil has none.In Figure 1 we present relations of two synsets: {samochód osobowy, auto 2, osobówka} and {automobile}.The hyponyms of the former are sedan 1 'sedan' and garbus 'Volkswagen Bug'.It is improbable for hyponymy to link sedan 1 and automobile or garbus and automobile.It is, however, quite natural to connect sedan and garbus with two synonyms of samochód osobowyauto 2 and osobówka.The four inter-register synonyms, {automobil} and {samochód osobowy, auto 2, osobówka}, share connection to their hypernyms, {samochód, auto 1} 'wheeled vehicle'.Note that sedan 1 and garbus are not near-synonyms because meronymy differentiates the relation between sedan 1 'car' and sedan 2 'sedan' -a hyponym of nadwozie 'car body type'.This is where inter-register synonymy, a relation which we have already defined for nouns and verbs, comes in.
Inter-register synonymy relaxes one condition on synonymy: stylistic registers of the two LUs differ significantly.The relation links LUs which can have the same direct hypernym and share instances of other constitutive relations except hyponyms, so they do not meet the criteria for belonging to the same synset.LUs linked by inter-register synonymy can also have different connotations, collocations and sentiment (emotional polarity).
The following is a test for inter-register synonymy adapted to adjectives: TEST VIII.Inter-register synonymy • X, Y to przymiotniki 'X, Y are adjectives'.
• X i Y w istotny sposób różnią się rejestrami stylistycznymi 'X and Y do significantly differ in stylistic registers'.
Jeżeli coś jest marne, to jest także badziewne 'if something is poor, then it is also shoddy [disregarding the stylistic register difference and connotation]'.
Jeżeli coś jest badziewne, to jest także marne 'if something is poor, then it is also shoddy [disregarding the stylistic register difference and connotation]'.
Marny is not a hyponym of badziewny, badziewny is not a hyponym of marny.
Marny and badziewny do significantly differ in stylistic registers.

Deverbal adjectives
Deverbal adjectives are usually divided into two main groups -predisposition (or potential) and activity adjectives (Puzynina 1976, Grzegorczykowa 1979, Kallas 1998, Nagórko 1987: 152).The former express the tendency of the agent or patient to act somehow or to get into a particular state (so, it is likely that the agent does something or something is done to him/her/it).The latter have the semantic property of action or process perpetuity -without any modal modifiers (so the agent simply does something or something is done to him/her/it) (Kallas 1998: 471).Predisposition adjectives are paraphrased using such terms as habitually or potentially, while activity adjectives lack this kind of modality: they can be explained by terms such as usually or often (Grzegorczykowa 1982: 51-2).
• Roles (Adj ¡ V) Activity adjectives are divided according to the semantic roles of the nouns they modify: agent, patient, instrument and so on.Among the role relations which we consider in plWordNet there are the following semantic roles of adjectives.In all tests, A is an adjective and V is a verb.
'that when it is drizzling'.
• Predisposition (Adj ¡ V) Predisposition adjectives are involved in three subtypes of predisposition relations: potential, habituality and quantification (Kallas 1998: 478-480).In all tests, A is an adjective and V is a verb.
-Potential.The relation links an adjective which could be explained by the phrase 'it is possible that'.There are a few suffixes with that function, including -ny (materiał palny 'flammable material'), -liwy (łamliwe kości 'breakable bones') and a very productive -alny roughly equivalent to English -able (poznawalny 'recognizable, that which can be recognized', mieszkalny 'livable, that in which someone can live').Potential adjectives appear in the objective case (Kallas 1998: 478-9).
-Habituality.This is described with the adverbs such as always, often, easily (Kallas 1998: 479).The noun modified is the subject.Habituality is signalled by a specialised suffix -liwy.

Deadjectival adjectives -markedness
Polish deadjectival derivatives are created not only by suffixation but also by prefixation: krótk-awy 'shortish', przy-krótki 'rather short' from krótki 'short' (Kallas 1998: 501).Linguists distinguish three main types of deadjectival derivatives: negative adjectives (for absence or lack of something), gradation adjectives (for intensification or diminution of the attribute value) and expressive markedness (Kallas 1998: 501-9).Only the latter type will be encoded with a special relation markedness.Negative and gradation adjectives can only be coded with vague derivational relation (derivativity, cf.Section 7.4.).That is because we can encode the relations without resorting to primary (more specific) derivational relations.

• Negative adjectives
Negative adjectives, such as bez -dyskusyjny 'undisputed' from dyskusyjny 'debatable' or nie-pewny 'unsure' from pewny 'sure', can be explained with gradable/complementary antonymy or converseness and hyponymy (see Figure 2).They are created by prefixation (the list of prefixes includes: nie-'non-', bez-'without', a-, ir-'non-, ir-, a-' and anty-'anti-', inter alia).• Gradation adjectives A gradation adjective denotes intensification or diminution of a value expressed by the base adjective.There are either comparative or superlative adjective forms, described in Section 7.2.or less categorial derivatives, produced by absolute derivation by degree (pol.derywaty stopniowania bezwzględnego, Kallas 1998: 502-506): biał-awy 'whitish' from biały 'white', przy-głuchy 'partly deaf' from głuchy 'deaf', super -nowoczesny 'supermodern' from nowoczesny 'modern' etc.We decided to define such adjectives by other relations (for example, hyponymy and derivativity).The semantic feature 'very' (intensification) can be treated as a differentia specifica, and the meaning of the base as a genus proximum.This is the case of super-nowoczesny, which can be linked to nowoczesny by hyponymy (if something is very modern, then it is also modern).On the other hand, przygłuchy and białawy cannot be so analysed.The implication does not hold: someone partly deaf need not be deaf; something whitish need not be white.It seems that the gradation relation serves such cases better (see Figure 3).
We introduce the two derivational relations (comparative and superlative) in cases when we define a given word using adjectives in those degrees.

Denominal adjectives -relational and gradable adjectives
Denominal adjectives up 87% of all adjectival derivates in Polish (Kallas 1998: 482;cf. Urban 2006: 19-20).Semanticists find them difficult to describe.There usually is a strong context-dependence.Consider the adjective wiejski 'rural' from the noun wieś 'village'.Here is a selection of its senses: wiejski kościół 'church (located) in a village'; wiejskie dożynki 'harvest festival (celebrated) in a village'; wiejska kiełbasa 'country sausage, made in a village or in village style', 'sausage from village'; wiejska wspólnota 'rural community, community that consists of villagers and inhabits a village', 'community from a village'.
The senses oscillate around location, origin or collectiveness.Their variety is constrained by the semantic properties of wieś: it is a place where people live in a community.While the sense set is determined by the meaning of the base, sense distribution depends on the semantics of the noun modified.Thus, kościół 'church' is a building, dożynki 'harvest festival' -a feast, kiełbasa 'sausage' -food, wspólnota 'community' -a group of people.
Because of the contextual nature of the senses of denominal adjective, we must not include them in Polish WordNet.It constists of lexical units -word+sense pairs -not of contextual meanings.Relational adjectives belong rather to the world of syntax than to the lexis.Instead of giving them a specific semantic relation (e.g., "role"), we may link them by cross-categorial synonymy.In plWordNet 2.0 this relation has already been used to link transpositional deadjectival nouns with the suffix -ość (± English -ness) and their adjectival bases, as in czerwoność 'redness' from czerwony 'red' It also connects gerunds (categorical deverbal nouns with suffixes -anie, -enie, -cie) and their base verbs (Maziarz et al. 2011b: 190).
• Relational versus gradable/qualitative adjectives It is important to distinguish relational and non-relational (also called gradable or qualitative) adjectives, because only the former may be linked by cross-categorial synonymy.Several tests have been proposed to capture the difference.
The two classes are not fully separable, and some researchers posit a third class of relational-qualitative denominal adjectives (Grzegorczykowa 1979, Szymanek 2010)." [T]here are specific semantically defined subclasses of denominal For example, the adjective sławny 'famous' (from sława 'fame') is defined in the UDP as «mający sławę, rozgłos, słynący z czegoś» 'having fame, renown, being famous for something'.The substitution test for characteristic adjectives works: Jeśli ktoś jest sławny, to jest charakteryzowany przez sławę 'If someone is then he is characterised by fame'.

Derivativity
To be consistent with our verb and noun derivation systems (Maziarz et al . 2011(Maziarz et al . , 2011b) ) we also mark derivativity relation if none of the relations discussed in Section 7.1-7.3holds and if the given pair of words is connected by derivational dependency.The relation will be used, for example, to link gradation adjectives with their adjective bases (see Section 7.2).The test for derivativity is this (X is an adjective, Y is an adjective, a noun or a verb): X jest derywatem Y 'X is a derivate of Y' Thus, arcy-ciekawy 'most/extremely interesting' is a derivate of ciekawy 'interesting'.
Because of the semantics of derivation phenomena, derivativity also carries an unspecified semantic relation together with morphological processes.

Derivational relations and word-sense differentiation
The system of derivational relations presented in Sections 7.1-7.3 is part-of-speechdependent.For deverbal and deadjectival adjectives, the relations are explicit and their semantics is quite well established. 13The majority of deverbal, deadjectival and simplex (underived) adjectives are qualitative (Post 1986: 109, Szymanek 2010: 83).It is the denominal adjective which is supposed to cause most sematic problems (Heinz 1957, Urban 2006).We built an algorithm for adjectives which can be used to supply the right class of adjectives with the right relations (Figure 4).
It is well known that senses are distinguished in wordnets on the basis of relations, mainly synset relations (Miller 1998: 36-7).The set of derivational relations which we have established will be useful as an auxiliary tool in distinguishing and describing various adjective senses.If a given derived adjective can be matched to a proper derivational relation, then it can be distinguished as a separate lexical unit.Two prominent exceptions are gradation and negative adjectives which have their own derivational relations which are not, however, included in our derivationalsemantic model.They also should be given appropriate lexical units in Polish WordNet.Derivational relations will help decide the number and type of meanings of a derived adjective.
For example, the denominal adjective żmijowaty is a qualitative adjective (it passes all tests described in Section 7. The similitudinal relation could be applied to sense (I) by linking this lexical unit to żmija 'viper' (something is żmijowate 'viper-like' if it resembles a viper by shape).Sense (II) would be given the same relation to another, metaphorical meaning of żmija 'someone who is bad, devious, hostile'.Since the two denominal adjectives receive their own derivational relation, according to the morphological criteria they may be distinguished as separate lexical units.
Indeed, the two meanings have different superordinates and acquire different synset relations in Polish WordNet 2.0, see Figure 4.
The third sense from the UDP ('expressing such features') will not get a separate entry in Polish WordNet, since we do not recognize any proper derivational relation for it.According to many linguists, such senses are a manifestation of adjective contextual properties and regular polysemy (Nagórko 1987: 110, Apresjan 1980: 275).
Consider the adjective październikowy (from październik 'October') which the UDP defines as "adjective [derived] from październik ".The relation/qualitative tests establish that there are: I a true relational meaning (expressed by the definition of the UDP), II a marginal (rare) qualitative sense 'having weather conditions like in October'14 , not-distinguished in the UDP.
This observations tell us that in Polish WordNet there are two lexical units październikowy, of which the first would be linked by cross-categorial synonymy to the noun październik 'October', while the second would be linked to the very same lexical unit with the similitudinal relation.
To sense (I) hardly any hyponym could be suited.Neither the adjective miesięczny 'occurring every month, lasting a month', nor the poetical (or dialectal) miesiącowy 'of the Moon' is appropriate to październikowy.15It seems that the cross-categorial synonymy would be the only lexico-semantic relation devoted to that LU.The more specific meaning (II), apart from the similitudinal relation, will also get two other relations: • październikowy 2 -value¡ warunki atmosferyczne 'weather conditions' (if something is październikowe 2 then it has weather conditions characteristic of October weather), • październikowy 2 -MODIFIER¡ czas 'time, period' (in contexts found on the Internet the adjective usually modifies durations).The final examples are the deverbal adjective odstraszający 'scary, scaring off' (e.g., odstraszająca mina 'a scary face') and for a deadjectival adjective prościutki 'very simple': Odstraszający is an adjective of the active participle form ending with -ący.The UDP gives it the definition «wywołujący strach, zniechęcający do czegoś» 'causing fear, discouraging from something' and clearly distinguishes it from the homonymous participle odstraszający.As we know, if such adverbial modifiers as teraz 'now' cannot be combined with a given word (there is no phrase *odstraszający teraz 'frightening now' in this non-participial sense), it is not a participle (Section 7.1.).The adjective odstraszający will be connected via the role-agent relation to odstraszać 'to scare someone/something off', according to the substitution test: odstraszający = taki, który odstrasza 'that which scares someone/something off'.
Prościutki has three meanings in the UDP.All of them are relativized to meanings of prosty from which the adjective was derived with a suffix -utki : I a diminutive form of prosty 'straight, not curving'; II (of objects) a diminutive form of prosty 'simple, plain, without decoration'; III 'easy to do or to understand, simple'.
None of our relations make sense for prościutki 3 'pleasantly and very easy to do or to understand'.The markedness relation to prosty 'easy to do or to understand' shall suffice.

Statistics and discussion
Sections 2-6 describe a system of semantic relations: hyponymy, value of the attribute, modifier, gradation, antonymy and converseness, as well as inter-register synonymy.Sections -present a system of derivational relations.Finally, Section introduces a support criterion of distinguishing senses for derivational relations: if an adjective can be attributed to one of the main adjective classes and if it can be given a proper relation (with the exception of negative and gradation adjectives), it constitutes a lexical unit, a unit of Polish WordNet 2.0.For underived (simple) adjectives the semantic relations will be decisive: different senses should acquire different instances of relations (Miller 1998: 36-7).
The system of semantic and derivational relations based on semantic, lexicographic and wordnet tradition has been tested on a small but carefully selected 100-word sample taken from the UDP.All adjectives have been passed through our substitution tests, and given the appropriate relations to lexical units.Four of 100 adjectives (żmijowaty, październikowy, odpychający and prościutki ) were featured in Section 7.5. in order to illustrate our sense recognition, division, and establishing approach.To every adjective we applied methods presented in detail in the preceding sections.The results appear in Table 6.Surprisingly, it was the Adj-N relation type that had nearly 60% of all instances of relations.The next one was the Adj-Adj type (1/3 of all instances).This explains why the set in plWordNet 1.0 could not capture all the semantic properties of adjectives -as in PWN 1.5, it was based mainly on Adj-Adj relations.
Pure semantic relations (not necessarily derivational) -antonymy, hyponymy, value of the attribute, modifier and gradation -account for 61.1% of all instances of relations.Cross-categorial synonymy for relational adjectives with 20.4% is the most frequent relation.The least frequent relation is gradation with frequency of 7.5%: still high.
We feel that the proposed system of adjectival relations in plWordNet 2.0 is not only grounded in linguistic premises but also quite well suited to the Polish language.The "relation density" now reaches 2.94 relations per LU (cf.Table 1 in Section 1).The "experiment" is clearly too small to draw far-reaching conclusions: we must repeat it on a larger sample.Constructing Polish WordNet will lead to real statistics which should prove convincingly that the relation set described in this paper is appropriate for Polish adjectives.

Figure 1 .
Figure 1.An example of incompatibility of relation sets of inter-register synonyms for nouns.

Figure 4 .
Figure 4. Derivational relations across different adjective morphological and semantic classes

Table 3 .
Relations among adjectives in plWordNet 2.0 Czy on jest dobry?Nie, wręcz przeciwnie: jest zły! 'Is he good?No, on the contrary: he is bad!' Jeżeli ktoś jest dobry, to nie może być zły 'If someone is good, then he/she cannot be bad' Jeżeli ktoś nie jest dobry, to niekoniecznie jest zły 'If someone is not good, then he/she need not be bad' • Czy ktoś/coś jest X? Nie, wręcz przeciwnie: jest Y! 'Is he/she/it X? No, on the contrary: he/she/it is Y!' • Jeżeli ktoś/coś jest X, to nie może być Y 'If someone/something is X, then he/she/it cannot be Y' • Jeżeli ktoś/coś nie jest X, to niekoniecznie musi być Y 'If someone/something is not X, then he/she/it need not be Y' Example -dobry 'good', zły 'bad' dobry 'good', zły 'bad' are adjectives which are co-values of the same attribute WŁAŚCIWOŚĆ 'quality' Verbuk 2007;antonyms express different values of the same attribute -they belong to the same scale (cf.Verbuk 2007; Croft and  Cruse 2004: 169-192).Of course, there are other adjectives connected with the same attribute(scale, van Rooij 2011: 131).The order of different values could be seen as a semantic ordering relation for adjectives pointing to the same scale.

Table 6 .
Statistics of relations in the UDP sample (antonymy is gradable and complementary, without converseness).