THE SEMANTIC CLASSIFICATION OF ADJECTIVES IN THE BULGARIAN WORDNET : TOWARDS A MULTICLASS APPROACH

The paper presents an attempt at semantic classification of adjectives in the Bulgarian wordnet. Although designed for the Bulgarian wordnet, the classification can be applied to other wordnets which are developed in parallel to the Princeton WordNet. The classification relies on information that is already available in WordNet from other synsets (noun, verb, and other adjective synsets) that are linked to the adjective synsets via lexicosemantic relations — including their semantic classes, as well as definitions and usage examples. The first stage of the work was already presented at the workshop “Challenges for WordNets” within the conference “Language, Data and Knowledge 2017”. The continuation of the effort as described in this article, covers a proposal for introducing additional semantic classes to the adjective synsets (if applicable).


Introduction
Each synset in the Princeton WordNet is assigned a semantic class 1 (Fellbaum, Osherson, & Clark, 2009;Miller et al., 1993).However, unlike noun and verb synsets which are subjected to elaborate semantic classifications nouns are organized into 25 semantic classes such as noun.person, noun.animal, noun.plant, noun.process, noun.event, noun.act, noun.state, noun.group, noun.location, noun.time, noun.vehicle, etc., and verbs into 15 semantic classes verb.stative, The semantic classification of adjectives in the Bulgarian Wordnet: Towards a multiclass approach verb.communication, verb.change, verb.cognition, verb.contact, verb.perception, verb.cause, etc., the classification of adjective is too sparse only three labels are applied to the adjective synsets adj.all (mainly) for descriptive adjectives, adj.pert for pertainyms, and adj.ppl for adjectival participles.In consideration of this, we propose a more detailed semantic classification of adjectives to be introduced into the Bulgarian wordnet.Our effort which is a work-in-progress draws upon a number of existing classifications proposed in the linguistic literature for Bulgarian as well as the semantic classification of adjectives in German as being applied to the Wordnet for German (GermaNet, see Hamp & Feldweg, 1997).
The approach relies on the relational network of the Princeton WordNet because the Bulgarian Wordnet (Koeva, 2008;Коева, 2014) largely keeps the original structure of the Princeton WordNet (while introducing closed class words such as pronouns, prepositions, conjunctions, particles, and interjections, as outlined in (Koeva, Tinchev, & Mihov, 2004).
The paper is organized as follows.In Section 2, we briefly discuss our starting point (namely, the classifications of adjectives in other wordnets, as well as in the linguistic literature on Bulgarian).Section 3 presents in more detail the implemented classification of adjectives covering: the ways in which we have identified candidate adjective synsets (which, at this point, include only adjectives marked with the generic label adj.all ) in the Bulgarian wordnet that will be further classified by taking into account the relevant information from nouns, verbs and other adjectives linked via lexico-semantic relations; an overview of the co-occurrence of adjective and noun classes within the synsets' trees (within the verification process); and a detailed review of the classification at present here, we exemplify only adjective synsets labeled with more than one semantic class label following the comments on the first stage of our effort described in Stefanova and Dimitrova (2017).Some directions for future work are given in Section 4.

Adjectives in WordNet
The concepts in WordNet (called synonym sets) are nodes linked to each other via lexico-semantic relations (edges) that encode relations between the synsets such as hypernymy/hyponymy, meronymy (membership, partiality, etc.), antonymy, synonymy, similarity, derivation, as well as a number of morpho-semantic relations between nouns and verbs such as agent, undergoer, instrument, event, result, location, material, etc.
In addition, each WordNet synset is classified by a semantic class (Miller et al., 1993;Fellbaum et al., 2009) noun synsets are organized into 25 semantic classes, while verb synsets are labeled under 15 semantic classes.Adjectives (in the Princeton WordNet) are classified into two larger classes: descriptive and relational adjectives; plus adjectival participles (Fellbaum, Gross, & Miller, 1993).They are accordingly labeled adj.all, adj.pert and adj.ppl and are organized into separate non-intersecting structures each class is linked to other synsets via different sets of relations.Descriptive adjectives (labeled by the semantic class adj.all ) are organized into clusters based on similarity of meaning (synonymy) and binary opposition (antonymy).Relational adjectives (labeled adj.pert) are (derivationally) related and subsequently linked to the synset which contains (as a literal) their source noun.Although the Bulgarian wordnet follows the structure of the Princeton WordNet, there are a number of language-specific concepts and features that are to fit in.For example, relational adjectives in Bulgarian often have no morphological equivalent in English (which often uses nouns as modifiers, as in dog food in English kucheshka hrana in Bulgarian) when (solely) introduced into the Bulgarian wordnet, they are linked to the respective noun via the pertainym relation.Adjectival participles are marked as adj.ppl and are related via participle relation to synsets containing the verbs they are derived from.
Thus, adjectives are organized via a set of relations encoding their properties of attribution, antonymy, similarity, derivation; fuzzynymy and thematic category (which have been added in the EuroWordNet (Vossen, 2002), etc.; some relations are specific for one of the two classes (attribute, similarity, fuzzynymy for adj.all ; pertainym for adj.pert; participle for adj.ppl ) (Dimitrova The semantic classification of adjectives in the Bulgarian Wordnet: Towards a multiclass approach & Stefanova, 2016).
Adjectives in the Polish WordNet (plWordNet) are categorized into four classes based on derivativity: deadjectival (derived from other adjectives), deverbal adjectives, quality adjectives (mostly denominal that are often derived from nouns via the attribute relation when express the values of an attribute), and relational adjectives.In plWordNet 2.0, a set of lexical-semantic relations was introduced to connect adjectives to the other parts of the wordnet as discussed in (Maziarz et al., 2012).The WordNet for Russian (RussNet; Azarova & Sinopalnikova, 2004) sticks to the traditional terms and organizes adjectives into four classes: descriptive adjectives that are subdivided into adjectives expressing size, and qualitative adjectives; relational adjectives; pronominal adjectives (derived from pronominal base); and ordinal adjectives (numerals).
As we have opted for a semantic classification, however, our approach partly channels the one that has been adopted by the WordNet for German (GermaNet).It is based on a classification proposed by Hundsnurscher and Splett (1982) which employs the modification property of the adjective a (modifying) adjective is (semantically) linked to a certain (modified) noun to form a separate semantic entity.Thus, thirteen semantic fields (Perzeption, Ort, Zeit, Bewegung, Substanz, natPhaenomen, Koerper, Gefuehl, Geist, Verhalten, Gesellschaft social, Gesellschaft social quantity, Relation) and three general semantic fields (Allgemein, Pertonym, Privativa) are hierarchically organized and further divided according to sub-features expressed by 70 (sub)classes organized around a specific feature.
Some of these classes were also found in the linguistic literature on Bulgarian where, traditionally, the adjectives are divided into two large classes qualitative and relational adjectives (roughly, in parallel to adj.all and adj.pert classes in WordNet) and are analyzed as a class of dependent lexical items whose semantic and syntactic properties are fully realized only in relation to the noun they modify Radeva (В.Радева, 1991).Thus, even when an adjective expresses a property of being related to an object or an event (as is the case with relational adjectives), it expresses a relational property of another object or event that manifests in a certain way, to a certain degree or in relation to a certain internal property of the modified object or event.
In Radeva (В. Радева, 1991), an adjective is claimed to denote a property that is "permanently inherent for the entity and is attributed to it in its entirety.This is also the reason to define the adjective as part-of-speech whose denotative function is realized through its connection to the noun.2According to the author, however, there is no clear distinction between qualitative and relational adjectives.Qualitative adjectives can be derived from nouns and refer to an attribute property of the defined entity (expressed by the noun) through its relationship to another entity.The property qualifies and characterizes the entity expressed by the noun from which they are derived (e.g., pitiful -pity, malicious -malice, etc. examples are from Radeva; В. Радева, 1991).Hence, the hypothesis is that the adjective expresses one-sided relationship with the entity denoted by the motivating noun; the property is perceived as permanent and inherent (В.Радева, 1991, p. 161).
In another study Radeva (Б. Радева, 2011) proposes a combined class of qualitative-relational adjectives due to the lack of clear distinction between the two traditional classes, on the one hand, and, on the other, because many relational adjectives acquire qualitative meaning when combined with certain nouns or even in consideration with the source word.Earlier, Trifonova (Трифонова, 1982, p. 169) claims that qualitative and relational adjectives can be part of a general grammarsemantic type while maintaining their relative autonomy.
The semantic classification of adjectives in the Bulgarian Wordnet: Towards a multiclass approach Radeva (Б.Радева, 2011) also states that the relation of the adjective to the motivating noun determines the class and the core meaning of the adjective, but an essential part can be attributed to the derivation, as well specific qualitative sense is realized with relational adjectives when combined with a limited class of nouns (kucheshka hrana 'dog food' -kucheshki zhivot 'doglike life' -kucheshka viarnost 'dog faithfulness' -kucheshki stud 'doggy cold' (Б.Радева, 2011, p. 29).
In the next section, we will discuss the semantic classes that we have formulated.The first stage of our effort where a set of adjective synsets (over 2,500) were manually classified, was already described in an earlier paper (Stefanova & Dimitrova, 2017).Here, we give additional details covering all the semantic classes (with the number of classified adjectives being raised to 3,110) and present our further effort for introducing additional semantic classes with some of the adjectives.
The classification needs to comply with the classifications of nouns and verbs that have been already implemented through introducing semantic classes to noun and verb synsets.For example, we expand the class of physical properties of (adj.body with adjectives associated with physical properties of animals (coating, fur, tail, etc.) and plants (flowering, roots, etc.).The substance class is additionally re-organized to cover ingredients; while material class is intended for manmade objects.Additionally, we have introduced classes for adjectives denoting a state (of a person or an object), a causing phenomenon or trigger of change of state, and adjectives that express quality properties of animate and inanimate objects.

Attribution and verification
First, we have classified a set of adjectives manually by applying only one semantic class to an adjective synset while taking into account the information that is already available in wordnet such as the semantic class of noun and verb synsets linked via lexico-semantic relations to the adjective to be classified, the semantic class of other adjectives in the synset structure, and definitions and usage examples.
To determine the semantic class of an adjective, we search in the database of the Bulgarian wordnet using regular expressions (the Bulgarian wordnet can be locally accessed through the Hydra system3 the regular expressions and the modal language are outlined in Rizov, 2014).We have retrieved information through: (i) Keywords in the definition the query [lang('bg')&definition(#'...')&pos('a')] searches for all synsets in Bulgarian (lang'bg') of the adjective part of speech (pos('a')) whose definition contains a string of symbols #'...'.The following keywords were used for the respective adjective classes: 'to feel', 'feeling', 'emotion' -adj.feeling;'react', 'obsession', 'behavior' -adj.behavior ; 'relates', 'is related', 'relation', 'attitude' -adj.relation;'causes', 'evokes' -adj.cause;'characterize', 'distinguish', 'quality' -adj.quality;'knowledge', 'cognition', 'experience' -adj.cognition;'suffers The semantic classification of adjectives in the Bulgarian Wordnet: Towards a multiclass approach (a disease that leads to physical changes in / on the body of a person', 'body', 'appearance'adj.body.For example, the query [sem_class('adj.all')&definition('intime')] returns a synset such as {antecedent:1, anterior:1, subsequent:1, precedent:1, previous:1}) (ii) Usage examples -they may exemplify reference to persons, objects and / or phenomena.There are adjectives that express characteristics that are only found with persons (humans) but not with objects (such as hair and skin color, feelings, etc.).Although the adjectives of the semantic class adj.behavior and adj.feeling can be modifiers to activities of a person, they are still related to persons and not objects, whereas the semantic class adj.substance expresses properties of an object.
(iii) Structure of the synset if the semantic class of an adjective cannot be easily determined, the structure of the synset is reviewed.For example, {izgladnial:1; pregladnial:1; umrial ot glad:1; osvirepial ot glad:1} / {famished:1; ravenous:2; sharp-set:1; starved:1; esurient:3} is linked to {ostarviavane:1; osvirepiavane:1} / {edacity:2; esurience:2; ravenousness:1; voracity:2; voraciousness:3} which is classified as noun.state,but its hypernym is {glad:1}, {hunger:5; hungriness:3} a physiological state manifested by an acute need of food.Therefore, the adjective can be classified as adj.state but the state is related to body adj.body.A set of semantic classes were automatically introduced through extraction from the database of synsets of derivationally related nouns and verbs (and adjectives, later on).Classified adjectives were checked and edited if the definition and usage examples did not match the selected class because there were discrepancies between the synsets.

Semantic classes: interdependence
The adjective classes comply with the semantic classes of nouns and verbs either the ones directly linked to the adjective or the ones that are indirectly linked via other adjectives in the synset tree.Adjectives are linked to noun and verb synsets via lexico-semantic relations: has_attribute ; eng_derivative ; category_domain ; usage_domain linking the synsets in a topical class).Table 1 contains an overview of the semantic classes of nouns to which a set of 3,110 adjective synsets are directly linked (and not indirectly, i.e., via other adjective synsets directly linked to noun synsets).4 The semantic classification of adjectives in the Bulgarian Wordnet: Towards a multiclass approach The semantic classification of adjectives in the Bulgarian Wordnet: Towards a multiclass approach An adjective synset is linked to other adjective synsets via relations of antonymy ( antonym ), and similarity (via relations also_see to link semantically related synsets, and similar_to for semantic similarity between focal and satellite synsets with close referential meaning).We assume that semantically related adjectives can be classified into the same semantic class but there are a number of exceptions as exemplified in Table 2  An adjective may express a property of an entity but the definition can be worded to include references to different entities such as human, animal, plant (for adj.body); or reference to behavior, feeling, cognition; etc.For example, {perceptive:1} 'having the ability to perceive or understand; keen in discernment' can be used both for human abilities and for actions attributed to human abilities, and the synset is linked to nouns classified as noun.attribute,noun.cognition,noun.feeling, as well as to verb.cognition.The adjective {consumptive:2} which is classified as adj.state with the definition 'afflicted with or associated with pulmonary tuberculosis', and examples 'a consumptive patient' (a state of a person and reference to adj.body) and 'a consumptive cough' (a hint at adj.cause 'a cough caused by tuberculosis').There are more than one way to The semantic classification of adjectives in the Bulgarian Wordnet: Towards a multiclass approach resolve these issues we may either split a synset, or classify it under more than one semantic class, and we have opted for the latest, as for now.
Additional semantic classes have been applied manually only to a set of adjectives that were automatically extracted from the database based on whether there is a (direct) relation to a noun synset or an (indirect) relation (no further than one step) via a related adjective in the synset tree (if there is a noun synset linked to an adjective synset linked via a similar_to or also_see relation to the adjective at hand).Some numbers for the adjective synsets that received additional semantic classes, are given in Table 3 (synsets with the most variety are given; more than two semantic classes can be applied).The classes person, animal, plant, food refer to the referent of the (modified) noun.
The example below 5 illustrates a complex adjective synset which has been classified according to its definition but whose synset tree involves adjectives for perception, state, cognition, feeling.

Semantic classes
The attempted semantic classification of the adjectives as applied to the Bulgarian Wordnet combines some of the classes outlined above (mostly adopted from the GermaNet classification) plus some information from noun and verb classes (following the Princeton WordNet classes the 5 Examples throughout the paper include: the synset in Bulgarian; the synset in English; the definition in English; the semantic class, plus additional semantic classes, if applicable; other synsets linked via lexico-semantic relation to the adjective synset at hand, with semantic classes given in italic.Beware that only one semantic class of the adjective synset if given can be viewed in the database of the Bulgarian wordnet which can be checked at: http://dcl.bas.bg/bulnet/.There is a parallel view with Bulgarian vs. English, the parallel synsets are marked by a red arrow beside the synset (for description of the viewer and its features, cf.Rizov, Dimitrova, & Barbu Mititelu, 2016).Examples throughout the paper do not give the whole synset tree.
The semantic classification of adjectives in the Bulgarian Wordnet: Towards a multiclass approach semantic classes have been previously validated and a number of changes have been introduced into the Bulgarian wordnet (as well as in the local database of the Princeton WordNet) the effort and results are described in Koeva, Leseva, Stoyanova, Dimitrova, & Todorova, 2016). 6s already stated, the continuation of the work is to introduce additional semantic classes to adjectives if: (i) they can express relations and properties that can be classified into more than one semantic class (e.g., the adjectives expressing relation to time (adj.time)can express a position in time or a duration of an act); (ii) the adjective expresses characteristics that can be attributed to a referent of specific class (described in Stefanova & Dimitrova, 2017).
In the examples below, we include only adjectives with two or more semantic classes (plus additional classes of the referent of the noun the adjective can modify (person, animal, plant) these are applied to adjective classes with a variety of modified noun's referents.
(1) Adj.body: these are adjectives that express physical characteristics of humans, animals, and plants, and the nature of the possessor of these characteristics is given as an additional semantic class.An adjective may refer to a property that can be considered under another semantic class but if it can be attributed to body part, it is still classified as adj.body.(2) Adj.cognition: adjectives denoting cognitive processes and contents and expressing cognitive abilities of a person or entities resulting from the cognitive activities of a person; in most cases, the class of the referent of the modified noun is not marked (for example, most of the adjectives related to grammatical or other linguistics abilities are classified as adj.cognition only; the class of the referent, however, is included in adjectives specifically referring to the state of the person, as in the example (2a)).

Table 1 :
Distribution of adjective classes and noun classes.

Table 2 :
for the combinations in synset trees of adj.behavior, adj.feeling, adj.perception and adj.cognition.Combination of adjective classes in adjective synset trees.

Table 3 :
Combination of adjective classes in adjective synset trees.