SEMANTIC RELATIONS AMONG NOUNS IN POLISH WORDNET GROUNDED IN LEXICOGRAPHIC AND SEMANTIC TRADITION

The paper describes a system of lexico-semantic relations proposed for the nominal part of plWordNet 2.0 — the largest Polish wordnet. We briefly introduce a wordnet as a large electronic thesaurus. We discuss sixteen nominal relations together with many sub-types proposed for plWordNet 2.0. Each relation is based on linguistic intuition and supported by a set of tests which facilitate its identification. There are two main groups: pure lexico-semantic relations and semanticderivational relations.


Introduction
There was no publicly available wordnet 1 for Polish before the development of plWordNet began in October 2005. Soon thereafter an early version became accessible via Internet. Work on plWordNet has continued unabated. At present, plWordNet with about 65 000 lexical units 2 (LU) described has become one of the 1 A wordnet is a large electronic thesaurus whose construction follows the main design principles of Princeton WordNet (Fellbaum, 1998). Wordnets have been constructed for more than 50 languages, including all most widely used ones. A large wordnet is a very useful language resource for Natural Language Processing and is often included among so-called basic language resources. 2 A lexical unit is a pair: lemma and one of the senses represented across different occurrences of this lemma in language utterances. In a wordnet, we do not make any assumptions on the nature of the senses. A lemma here is understood, a little technically, as a morphological word-form selected as a representative of the whole set of word forms of the same grammatical class. largest wordnets in the world. Such large scale of a wordnet is its very required property, because wordnets are used as a basic lexical semantic resource in Natural Language Processing. The vast majority of LUs are nouns; this conforms to the needs of the intended applications of plWordNet. The verbal and adjectival parts also grow gradually.
Lexico-semantic relations are a key design consideration for the structure of a wordnet -a network of LUs. LUs are the basic building blocks, and the network of relations is the only means of defining the meanings of LUs in plWordNet. 3 From the lexicographic point of view, linking is recognized as one of the most important characteristics of an electronic dictionary (Svensén 2009: 443). In the perspective of structuralism, the vocabulary of a given language constitutes a system (McCarthy 2003: 76). Although de Saussure's ideas are hard to implement, the very need is to describe words in the light of paradigmatic and syntagmatic relations. In this paper we present nominal lexico-semantic relations of the Polish Wordnet in the perspective of general linguistics and traditional lexicography. We will show that new types of relations introduced in the nominal part of plWordNet 2.0 are well grounded in both disciplines.

Overview of the plWordNet Nominal Relations
Due to limited funding, plWordNet 1.0 was planned as a wordnet of a very moderate size, delivering the basic description of the most frequent Polish lemmas 4 . Thus the set of lexico-semantic relations of plWordNet 1.0 was restricted to the most common wordnet relations. In the case of plWordNet 2.0 our goals are much more ambitious. We do not only want to achieve a wordnet very large when measured in the number of LUs. We also aim to construct a rich description of the Polish system of lexical meanings, still based on the relational paradigm, which is in some way useful for language processing. As a result, several new relations have been introduced, and the existing system of relations has been revised.
The system of lexico-semantic relations of plWordNet 2.0 has been built as a direct extension of the plWordNet 1.0 system. A brief comparison of both systems of lexico-semantic relations is presented below. Definitions of all relations have been revised but the additions have been concentrated mainly in the area of those relations which are formally expressed via derivational relations. The relation system must be perceived as a whole, so in the following sections we will discus the present state of all relation definitions. For the sake of presentation clarity, the relations can be divided into two main groups: 1. pure lexico-semantic relations which are not primarily and obligatorily expressed via derivational relations, 2. semantic-derivational relations for which the primary vehicle is an obligatory formal derivational association between word forms representing LUs from the relation instance (a pair).
The above distinction is mainly based on the presence or absence of a formal derivational link in the background, but we do not exclude the situation in which the relations of the second group are also applied to LU pairs which are derivationally linked. In the following sections we will discuss both groups of relations.

Pure semantic relations
Relations of this group are identified exclusively by semantic criteria. LU word forms do not deliver, in general, any clues to support recognition of their relation. Most relations identified in plWordNet are well grounded in the linguistic and lexicographic tradition, but also often present in wordnets and traceable back to Princeton WordNet (Fellbaum 1998) -the first wordnet ever constructed. Word-Net is focused more on psychological premises, but its structure refers to the relations identified in linguistics as language markers of conceptual relations between concepts lexicalised via particular LUs.

Similarity relations
Several forms of meaning similarity of nominal LU pairs have been named in the linguistic tradition, e.g.: synonymy, near synonymy and hyponymy. What they share in common is some ability to be used exchangeably across different contexts of use. What differentiates them are conditions imposed on their exchangeability among contexts of use.

Synonymy
From the lexicographic point of view, synonyms occupy a significant position in definitions of word meanings (Svensén 2009: 240-1). They are traditionally attached to intentional 6 definitions; in cumulative synonym definitions they are even the main means of a word's semantic characterization (Svensén 2009: 250).
Synonymy is crucial for the idea of WordNet as a network of synsets linked by semantic relations. Typically, a synset is only vaguely defined as a set of near synonyms assumed to "lexicalise" some shared concept -see the discussion in (Piasecki et. al. 2009). It is worth emphasizing at the very beginning that WordNet has never been a network of relations defined only over synsets, because relations defined directly on the set of LUs were always present. Moreover, relations introduced in WordNet for synsets have been clearly motivated by lexico-semantic relations (of the same name) used in linguistics.
In plWordNet we decided to adopt a unified model. All relations are defined at the level of LUs, see (Derwojedowa et. al. 2008, Piasecki et. al. 2009).
WordNet, interpreted as a thesaurus, is often used also as source of sets of synonyms, so the proper identification of synonymy is very important for its construction. Here we face a well known problem: there are many types of synonyms and the term "synonymy" is defined in many ways (Lyons 1995b: 60-1). The most important division is that of absolute synonyms and partial synonyms (or near synonyms) (Gouws 1996: 118-120;cf. more specific Lyons' typology -Lyons 1995b: 60-3). In lexicography three major aspects of the word meaning are commonly distinguished, namely: (1) descriptive meaning, (2) connotative meaning, (3) pragmatic characteristics (Svensén 2009: 214-5). The difference between absolute and partial synonyms lies in the connotative meaning and pragmatic characteristics: «Absolute synonyms do not only have the same denotation but also the same connotation and similar stylistic values. They are items that can be used in the same register without loss of communicative success» (. . . ) (Gouws 1996: 119;cf. Sterkenburg 2003: 389).
We avoid defining synonymy directly in plWordNet. Instead, synonymy is defined by synsets and synsets are determined by the structure of selected wordnet relations. Two nominal LUs are synonymous, i.e. they belong to the same synset, if they share the same hyponyms/hypernyms and meronyms/holonyms, thus they are located in the same area of the hypernymy graph (Piasecki et. al. 2009). Only LUs sharing links of these relations can be grouped into one synset. This is the basic rule of constructing a synset in plWordNet. In addition to structural clues, linguists are supported in synset construction by substitution tests defined for plWordNet synonymy and based on the idea of interpreting synonymy as mutual hyponymy (Piasecki et al. 2009: 23-4; cf. test I on p. 185): • Jeśli jest X-em, to jest też Y-em If he/she/it is X, then he/she/it is also Y, • Jeśli jest Y-em, to jest też X-em If he/she/it is Y, then he/she/it is also X This test set, however, is only a secondary criterion for identifying synonymous LUs, i.e. for putting two LUs into the same synset. The primary criterion stays unchanged since plWordNet 1.0: LUs from the same synset must share hyponyms/hypernyms and meronyms/holonyms. Thus synonymy, which is encoded by synsets, is to some extent a secondary relation in plWordNet. It is entailed by the remaining lexico-semantic relations. As a result we obtain a relation that captures both absolute synonyms and part of near synonyms.
In plWordNet 2.0 one more substitution test -a necessary condition -was added to the test set for synonymy. The additional test is intended to emphasize the association of synonymy with the hypernymy hierarchy and other synset relations: • X i Y mają wspólny hiperonim i wchodzą w identyczne relacje semantyczne X and Y have a shared hypernym and participate in the identical synset relations.

Inter-register synonymy
Inter-register synonymy, introduced in plWordNet 2.0, can be characterised briefly as synonymy that occurs between lexical units which have different stylistic register. It is a kind of near synonymy. (1) descriptive meaning (denotation) + + (2) connotative meaning (connotation) -/ + -/ + (3) pragmatic characteristics (stylistic markedness) + - The difference between inter-register synonyms (in a sense defined here) and synonyms depends on the difference of stylistic registers to which two LUs belong: inter-register synonyms belong to significantly different registers, while synonyms (LUs of one synset) must be in the same register. Thus the difference results from the pragmatics characteristics of LUs, while their denotational meaning is identical or very close ( Table 2). The difference in registers directly influences the structure of the relation graph. It is caused by our understanding of synonymy as topological identity (i.e. network location identity) with respect to the network of relations. Inter-register synonyms do not share locations in the network, because they do not belong to identical lexico-semantic relations. It is not possible to link by hyponymy an unmarked LU (general language LU) with a marked LU (e.g. a vulgar LU), for instance: {orlę «odważny chłopiec, zwłaszcza elew lub podchorąży szkoły lotniczej» (USJP) 'a brave boy, especially graduate or cadet of an aviation school'} Henceforth, we use the sign "*" to mark incorrect synset or relation instance. Orlę is a hyponym of chłopiec, but not of gówniarz. Inter-register synonymy makes it possible to avoid this contradiction, i.e.: Marked near-synonym inherits the hyponymy relation of its unmarked nearsynonym. Such view is familiar for lexicography: «General-language expressions must not be explained by means of technicallanguage synonyms (. . . ). Similarly, words and phrases belonging to normal prose should not be defined by means of synonyms that belong to statistically marked language varieties (*boy lad). The opposite procedure, on the other hand, is normal, with stylistically marked expressions being explained by unmarked near synonyms; however, it is then necessary to specify that the expression is marked, and in what way» (Svensén 2009: 216-7;cf. 249).
Substitution tests for inter-register synonymy are presented below: • X i Y mają ten sam hiperonim, zbiory ich hiponimów nie pokrywają się 'X and Y share a hypernym, their sets of hyponyms do not overlap' • X i Y nie są synonimami 'X and Y are not synonyms' • Jeżeli jest X, to także jest Y [pomijając różnicę rejestrów stylistycznych] 'If he/she/it is X, then he/she/it is also Y [to the extent of the stylistic register difference]', • Jeżeli jest Y, to także jest X [pomijając różnicę rejestrów stylistycznych] 'If he/she/it is X, then he/she/it is also Y [to the extent of the stylistic register difference].

Hyponymy/hypernymy
Linking words by the hyponymy/hypernymy relation is the main way to define lexemes in dictionaries: a head of an intentional definition is usually a hypernym, whereas an extensional definition lists a number of hyponyms. Sometimes both hyponyms and hypernyms are built into the definition structure (Svensén 2009: 218-9, 249).
Hyponymy/hypernymy symmetric relations constitute the skeleton of the wordnet structure (Piasecki et al. 2009: 28). Hypernymy, for example, can be characterised by a pair of implications: p ⇒ q, ∼q ⇒ p, where q is a hypernym, and p -a hyponym (Lyons 1995b: 127).
Hyponymy kobieta 'woman'człowiek 'man' is sometimes identified with the implication p ⇒ q, and the mutual implication p ⇔ q, used here to characterise synonymy, is called mutual hyponymy. Thus hypernymy perceived in this way becomes a primary relation in comparison to synonymy (Lyons 1995b: 127-8).
Substitution tests for hyponymy/hypernymy have been changed slightly in comparison to plWordNet 1.0 (Piasecki et al. 2009: 187, test X). In order to express better an aspect of the implication "⇒" in the tests, we introduced modal verbs musieć 'have to' i móc 'be able': • Jeżeli ktoś/coś jest X-em, to musi być Y-em (X ⇒ Y) 'If he/she/it is X, then he/she/it must be Y', • Jeżeli ktoś/coś jest Y-em, to niekoniecznie jest X-em 'If he/she/it is Y, then he/she/it not necessarily is X' (Y ⇒ ∼ X), • If he/she/it is not Y, then he/she/it cannot be X.
As a result, unacceptable implications similar to those listed below are systematically blocked: *Jeżeli jest orlęciem, to musi być gówniarzem. 'If he is brave boy, then he must be a squirt' (vulgarism)', *Jeżeli jest istotą żywą, to musi być bytem. 'If it is a living being, then it must be an entity' (phil. «everything what exists in some way; main subject of ontological investigations»), *Jeżeli jest taksówką, to musi być furą. 'If it is a taxi, it must be a a good, expensive car (joke, informally)'.
Blocking of inter-register hypernymic links together with a new relation of interregister synonymy resulted in a kind of hierarchy of stylistic registers. Linking a marked LU and an unmarked LU with hypernymy is excluded because of a general rule. This rule fits the linguistic intuition. In that way LUs of the Polish wordnet are divided into two main separate classes: general LUs and other. LUs of the general register encompass literary language and colloquial words. The set of other registers includes: scientific, technical, informal, vulgar, in jest, outdated (old use), archaic and historical, regional (dialects).

Instance-of and type relations
In plWordNet we tried to avoid introducing proper names and linking them to common nouns. In the relation structure of plWordNet 2.0 we made one exception to this rule. A proper name is included in the wordnet only if there is a common noun derived from it. The noun must be also relatively frequent in the corpus. Such nomina propria are included in the wordnet structure by means of the type relation (from an instance to superior category -a type LU) and the instance of relation (from the superior category to an instance), e.g. Wrocław is an instance of miasto 'a city' and miasto 'a city' is a type of Wrocław. The substitution test for the instance of relation received the following form:

Contrast and meaning opposition relations
A wordnet is built not only on lexico-semantic relations expressing different kinds of similarity, but also on opposition. Antonymy is commonly used in dictionaries; it occurs in definitions and entries of many dictionaries (Svensén 2009: 248, 251).
The relation is of high importance in semantics: «Antonymy, or 'oppositeness of meaning', has long been recognized as one of the most important semantic relations» (Lyons 1995b: 460). Its status is not disputed: «Unlike synonymy, everyone agrees that antonymy exists, and it is robustly evident in natural language» (Murphy 2003: 169). Lyons distinguishes three types of opposition: • complementaritysingle : married, man : woman, • gradable opposition, • converseness. Following Lyons we divide antonymy into the three categories.
The complementarity is caused by the fact that X and Y belong to a two-term set (Lyons 1995b: 461)  • 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'.
The answer to the latter question is not obvious for complementary terms: Q: If someone/something is not a man, then he/she/it is a woman. A: It depends (think of a child, neither a man, nor a woman). «The theoretical problem arising at this point is to make it [sc. formal definition of antonymy] narrow enough, that is, rule out words that are not true opposites but simply instances of incompatibility of sense (Lyons 1981: 154-5)» (Stępień 2008: 228).
To answer "yes" we must know that we are talking about adults and we put only the matter of gender in focus. In addition we must ensure that words in question are strongly semantically associated, i.e. they share a substantial part of their meanings. Antonyms -a little paradoxically -have a lot in common. An antonym negates only part of the definition of its counterpart, what is left remains unnegated (Apresjan 2000: 269, 270 and 273). That is why we add to the test set an additional necessary condition: • X and Y must be co-hyponyms or co-meronyms of the same lexical unit(s).
This condition guarantees that the words tested for antonymy are compatible in meaning.
To test them for gradable antonymy, we use Lyons's second question and the necessary condition: • X i Y muszą być kohiponimami lub komeronimami tej samej jednostki leksykalnej.
'X and Y must be co-hyponyms or co-meronyms of the same lexical unit(s)'.
'If someone/something is X, then he/she/it cannot be Y.' But it is not enough. All co-hyponyms of X pass that test, although they are merely its antonyms. To focus the linguist's attention on the contradiction of the senses, we broaden the test with this sentence: • Is he/she/it X? -No, on the contrary: he/she/it is Y.
The motivation for this question is similar to Cruse's proposal: «Ask someone for the opposite of table, or gold, or triangle (. . . )» (Cruse 1997: 257).
To distinguish between complementary terms and gradable antonyms we must also ask the question: • Jeżeli ktoś/coś nie jest X, to musi być Y If someone/something is not X, then he/she/it must be Y'.
The answer for gradable antonyms is: NO.

Oppositeness of senses
Putting things together, the final set of substitution tests for the semantic contrast is presented in the table below:

TESTS
Complementary Antonymy X and Y must be co-hyponyms or co-meronyms of the same lexical unit(s) If someone/something is X, then he/she/it cannot be Y If someone/something is not X, then he/she/it must be Y It is not true, that if a is X (Praep) Gradable Antonymy X and Y must be co-hyponyms or co-meronyms of the same lexical unit(s) Is he/she/it X? -No, on the contrary: he/she/it is Y If someone/something is X, then he/she/it is not Y If someone/something is not X, then he/she/it must be Y [NO] It is not true, that if a is X (Praep) b, then b is Y (Praep) a Converseness X and Y must be co-hyponyms or co-meronyms of the same lexical unit(s) If a is X (Praep) The last test in the Gradable Antonymy and Complementary Antonymy group was added in order to exclude the possibility of classifying proper gradable antonyms as converses.
We can notice that test groups for all three opposition relations can be merged into one sequence of tests -presented in the table below -defining a kind of algorithm -shown below.
line TEST 1 X and Y must be co-hyponyms or co-meronyms of the same lexical unit(s) 2 If a is X (Praep) b, then b is Y (Praep) a 3 Is he/she/it is X? -No, on the contrary: he/she/it is Y 4 If someone/something is X, then he/she/it cannot be Y This cascade-like system of testing sentences enables linguists to distinguish between words of contrast (antonyms and converses) and words which are related in some other way.
We distinguished five subtypes of meronymy/holonymy in plWordNet 1.0, namely: part, place, portion, element of a collection and substance (Piasecki et al. 2009: 31-2, test XIII). In plWordNet 2.0 this set has been extended with an additional subtype of taxonomic unit. It is motivated by the needs of expressing lexico-semantic relations inside scientific taxonomies, especially biological taxonomy, e.g. kotowate 'felidae'kotokształtne 'feliformia'. The test set for meronymy/holonymy of the taxonomic unit type is presented below: • X jest elementem taksonomicznym Y, 'X is a taxonomic unit within Y' • Y reprezentuje poziom taksonomiczny, którego elementem jest X, 'Y represents a taxonomic level, whose unit is X' • Y nie jest elementem taksonomicznym X 'Y is not a taxonomic unit of X'

Fuzzynymy
The space of different kinds of lexical semantic association seems to be a continuum. Some of its more prominently delimited subspaces have been named by wordnet relations. However, we can still expect many distinguishable but less frequent types of associations, e.g. informatyk 'computer scientist'komputer 'computer'. In order to collect instances of those associations, we introduced the fuzzynymy relation in plWordNet 1.0. Fuzzynymy links LUs both inside a part of speech, e.g. Noun -Noun, as well as in a cross-categorial way, e.g. Noun -Verb or Adjective. We follow here the practice of EuroWordNet. Fuzzynymy is underspecified and expresses a kind of lexical semantic association visible for native speakers. Concerning dictionary definitions, fuzzynymy corresponds to the expressions like "about someone, about something", "connected with" used to refer to a semantic field, to focus the reader's attention on them or to characterise use of a word in particular subject field (Svensén 2009: 210;Piotrowski 2001: 151), e.g.

Semantic-derivative relation
The word formation process is semanticaly fertile, but ways of creating new meanings are vague and sometimes erratic (Malmkjaer 2004: 359). Suffixation and composition remain the basic ways of contemporary Polish noun formation (Bajerowa 2003: 62-67;Grzegorczykowa, Puzynina 1998: 389), but the meaning of a particular suffix varies: for example, in some deverbal derivatives the -ca suffix means 'person who does something': wystawca 'a person who shows to the public industrial goods, works of art' < wystawiać 'to show to the public industrial goods, works of art', zbawca 'a person who saves someone from danger' < zbawiać 'to save someone from danger', whereas in compounds the -ca suffix has different sense 'animal': roślinożerca 'phytofag, animal that eats only plants' < roślina 'plant', żreć 'to eat', mięsożerca 'carnivore, animal that eats meat' < mięso 'meat' (Grzegorczykowa, Puzynina 1998: 373-4). Some suffixes are more productive than others, some loose their power and popularity, while another gain it, which is a complex historical process (Malmkjaer 2004: 359;Bajerowa 2004: 67-69).
Because the meaning of the derivational base usually becomes part of a derivative, it seams natural to define semantics of the latter by the former. Giving a meaning paraphrase with derivational characteristics is an ordinary way to define morphologically related words via morpho-semantic definition (Svensén 2009: 227;Sterkenburg 2003: 88-91, 93;Szymczak 1982: XVIII-XIX).
In plWordNet 1.0 derivational relations were divided simply into two large groups providing a coarse-grained level classification of relations, named relatedness and pertainymy (Piasecki et al. 2009: 32-34;tests XV, XVIII). This division was intentionally provisional and motivated to a very large extent not by semantic but by formal criteria (e.g. regularity).
In plWordNet 2.0 we replaced this classification with a set of more detailed lexico-semantic relations which are expressed by means of derivational transformations but have a clear semantic motivation.

Semantic roles
Thematic role relations semantically characterise associations between a noun and derivationally linked verb from the perspective of a situation denoted by the verb. According to Fillmore's Frame Semantics every predicate has its own semantic frame, which consists of different semantic roles (Fillmore 1968). The approach is used also outside the area of frame-based verb description (Vossen et al. 1998: 101-2). In EuroWordNet the only limitation is the strength of semantic connection of two words: «This relation is only being used to encode data on arguments/adjuncts that are strongly implied in the meaning of a verb/noun. This is not the same as encoding arguments or adjuncts co-occurring with a verb/noun in a sentence. In the relational approach we follow, we only encode the semantic features incorporated in the meaning of a word. These certainly also determine the kind of syntactic contexts in which that word may occur, but do not necessarily coincide with them» (Vossen et al. 1998: 101-2).
The semantic roles strongly affects the word formation process (Grzegorczykowa, Puzynina 1998: 378-383;Laskowski 1973). Following solutions proposed in Eu-roWordNet and the scheme proposed for Polish by Grzegorczykowa and Puzynina (Grzegorczykowa, Puzynina 1998: 398-415), we distinguished nine roles -relation subtypes: agent, patient, instrument, location, product, time, agent of hidden predicate, object (of hidden predicate) and product (of hidden predicate). Morphological connections between deverbal nouns and their derivational bases guarantee that the meaning of a predicate is involved in the meaning of an argument.

Emotional markedness
Many Polish nouns have emotional markedness encoded in their structure via suffixes: the most frequent in the lexicon are diminutives, augmentatives and names of young beings (Grzegorczykowa, Puzynina 1998: 427-30).
The emotional markedness is similar to inter-register synonymy, because the relations both carry stylistic register markings. The difference is essential, however: the emotionally marked words are hyponyms or unmarked counterparts rather than their inter-register synonyms (kociak 'small' + 'nice' + 'cat' is rather hyponym of kot 'cat').

Others
In plWordNet there are additional four derivational relations linking nouns with nouns, adjectives and verbs: i.e., state|feature bearer, femininity, dweller and derivation.
The inhabitant relation is very productive in Polish. There are whole series of derivatives with semantic element 'inhabitant': domownik < dom, wrocławianin < Wrocław, Kanadyjczyk < Kanada etc. (Grzegorczykowa, Puzynina 1998: 437-8). The inhabitant relation links nomina propria and nomina appellativa: The derivativeness is a last-chance relation. If none of the previously mentioned derivational relation is suitable for a given pair of related words, then that morphosemantic relation should be used.

Conclusions
For linguists, a wordnet often seems to be a simple thesaurus based on simple rules and well known relations. Two aspects make a wordnet an interesting language resource: its scale and consistency in implementing the adopted set of rules. A wordnet should be large and deliver a broad picture of the lexical-semantic system of the given language. The development of plWordNet 2.0 goes exactly in this direction. It now offers a large-scale description of nouns and it will provide massscale description of verbs and adjectives soon. A description of the lexical meanings is being consequently built in plWordNet 2.0 as a network of lexical units. What makes plWordNet different from other wordnets is that all relations introduced are clearly linguistically motivated. By constructing plWordNet 2.0 we want to achieve a relatively complete picture of the Polish lexical-semantic system with respect to two aspects: coverage of LUs and richness of description.
The relational paradigm, in which the means of description are reduced to the lexico-semantic relations, introduces definite limitations on the description's richness. The amount of semantic information provided for a LU is correlated with the number of relation links which concern it. With a greater number of relation types a better wordnet-based description can be obtained. However, too infrequent relations will result in too fragmented and too accidental a description, especially from the perspective of wordnet applications in language technology. Thus, we explored a wide range of potential lexico-semantic relations. In the paper we presented a selected sub-set. Relation originating from the formal derivational associations are its important part. Their strong presence is a characteristic feature of wordnets for Slavic languages, see e.g. (Pala & Hlaváčková 2007;Koeva 2008). It is worth emphasizing that derivative associations were only the starting point and our goal was to identify several more productive and more frequent classes of semantic dependencies behind them. It is hard to evaluate the proposed system of relations. It will be possible only when we complete a large wordnet -plWordNet 2.0, which this system of relations will underlie.