The paper presents an overview of an attempt at the semantic classi cation of adjectives in the Bulgarian Wordnet based on the information that is already available in WordNet, and other classications proposed in the literature (classi cations in the linguistic literature for Bulgarian and approaches implemented by other wordnets, more precisely, the Wordnet for German). The proposed approach to adjective classi cation is a work-in-progress and involves using information from other synsets that are linked to the adjective synset at hand via lexico-semantic relations as introduced by the Princeton WordNet, their semantic primes and the texts of de nitions and usage examples.
The paper proposes a semantic classi cation of adjectives in WordNet and an
approach to its application to the Bulgarian Wordnet drawing upon existing
classi cations proposed in the linguistic literature for Bulgarian and the classi cation
of adjectives (in German) as found in the Wordnet for German (GermaNet, cf.
[
WordNet concepts are nodes linked to each other via lexico-semantic relations
(edges) that encode relationships between concepts such as hypernymy/hyponymy,
meronymy (membership, partiality, etc.), antonymy, synonymy, similarity,
derivativity, etc. [
Additionally, each synset is classi ed by a semantic primitive [
The organization of the information and the structure of the Bulgarian Wordnet
follows, in principle, those of the Princeton Wordnet [
Some of these classes were also found in the classi cations in the linguistic
literature on Bulgarian following the observation that there is no clear-cut
division between qualitative and relational adjectives [
As a result of the observations on the literature, we embarked upon outlining semantic classes of adjectives covering: social and community a liations (socialrelated, Gesellschaft in GermaNet (GN)); place or location; local time (time, Zeit in GN); weather (natPhaenomen in GN); physical characteristics (body, Koerper); movement (motion, Bewegung in GN); knowledge (cognition, Geist in GN); attitude (relation, Relation in GN); feeling (Gefuehl in GN); behavior (Verhalten in GN). Three of the semantic classes in GN were too general and included adjectives that can be considered attributes: General (Allgemein), Pertainyms (Pertonym), Privative (Privativa). The classi cation of adjectives in WordNet needs also to comply with the classi cations of nouns and verbs already implemented in WordNet through introducing the semantic primes to noun and verb synsets. When formulating semantic classes, we had to combine all this information.
We expanded the class of physical characteristics with adjectives associated with physical properties of animals (coating, fur, tail, color, etc.) and plants ( owering, roots, etc.). The substance class was additionally re-organized to cover ingredients; and material (for man-made objects). Additionally, we 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 expresses quality characteristics of animate and inanimate objects.
The attempted semantic classi cation of the adjectives as applied to the
Bulgarian Wordnet combines some of the classes outlined above (mostly adopted
from the GermaNet classi cation) plus some information from the classi cation
of verbs and nouns (following the Princeton WordNet classes { the semantic
primes have been previously validated and some changes have been introduced
into BulNet and Princeton WordNet { the e ort is described in [
Ex.: fa: pronitsatelen:3; shvatliv:3g / fa: perceptive:1g having the ability to perceive or understand; keen in discernment antonym: a: unperceptive:2; unperceiving:1 [adj.cognition] eng derivative: v: perceive:1 [verb.cognition] eng derivative: n: perceptiveness:1 [noun.attribute] eng derivative: n: discernment:3; perceptiveness:2 [noun.cognition] eng derivative: n: insight:4; perceptiveness:4; perceptivity:1 [noun.feeling] (3) Adj.feeling: adjectives for feelings and emotions of a person or related entities.
Ex: fa: otchayan:7g / fa: despondent:1g without or almost without hope eng derivative: v: despond:1 [verb.emotion] eng derivative: n: despondency:1; despondence:1; heartsickness:1 [noun.feeling] similar to: a: hopeless:2 [adj.feeling] (4) Adj.behavior: adjectives expressing behaviors, behavioral symptoms, etc. Ex.: fa: rezerviran:2g / fa: reserved:1g marked by self-restraint and reticence also see: en - a: backward:5 [adj.behavior] also see: en - a: undemonstrative:1 [adj.behavior] (5) Adj.perception: adjectives for seeing (color), hearing (voice, sound), and perception (taste, sense, etc.) and, rarely, estimation (liking/disliking, etc.) Ex.: fsmradliv:2; zlovonen:2g / fa: fetid:1g o ensively malodorous eng derivative: n: harmfulness:1; noisomeness:1; noxiousness:1 [noun.attribute] eng derivative: n: olfactory property:1; smell:7; odor:1; odour:1 [noun.attribute] eng derivative: n: malodorousness:1; stinkiness:1; foulness:1 [noun.attribute] similar to: a: malodorous:1; malodourous:1; unpleasant-smelling:1 [adj.perception] (6) Adj.time: adjectives expressing age, historical period, succession in time, longevity, occurrence in a speci c time period.
Ex.: fa: dvumesechen:2g / fa: bimestrial:1g two months long; lasting two months eng derivative: n: bimester:1 [noun.time] similar to: a: long:5 [adj.time] (7) Adj.location: adjectives expressing spatial properties, placement, succession in space, etc.
Ex.: fa: tropicheski:1g / fa: tropical:3g relating to or situated in or characteristic of the tropics (the region on either side of the equator) eng derivative: n: Torrid Zone:1; tropical zone:1;tropics:1 [noun.location] eng derivative: n: tropic:3 [noun.location] similar to: a: equatorial:1 [adj.location] (8) Adj.motion: adjectives related to manners of motion (vehicle, speed, etc.). Ex.: a: visokoskorosten:1, barzohoden:1 / a: high-speed:1; high-velocity:1 operating at high speed similar to: a: fast:5 [adj.motion] (9) Adj.social: adjectives that express relations resulting from social norms and principles or concern entities or phenomena that are part of the social structure (incl. religion, ideology, marriage, etc.) Ex.: fa: politicheski:1g / fa: political:1g involving or characteristic of politics or parties or politicians antonym: a: nonpolitical:1 [adj.social] similar to: a: governmental:1 [adj.social] (10) Adj.substance: adjectives expressing relation to substances.
Ex.: fa:kristalen:3, kristalinen:1g / fa: crystalline: 1g consisting of or containing or of the nature of crystals similar to: a: crystallized:1; crystalised:1 [adj.quality] (11) Adj.material: adjectives expressing materials used for production of manmade objects.
Ex.: fa: darven:1g / fa: wooden:1g made or consisting of (entirely or in part) or employing wood similar to: a: woody:3 [adj.substance] (12) Adj.weather: adjectives related to climate conditions (a limited number). Ex.: fa: tropichen:1g / fa: tropical:1g of weather or climate; hot and humid as in the tropics eng derivative: n: tropic:3 [noun.location] similar to: a: hot:9 [adj.perception] (13) Adj.quantity: adjectives expressing quantity, size, degree, range, etc. Ex.: fa: minimalen:1g / fa: minimum:1g the least possible also see: a: meager:1; meagre:1; meagerly:1; stingy:1; scrimpy:1 antonym: a: maximal:1; maximum:1 [adj.quantity] similar to: a: negligible:1[adj.quantity] (14) Adj.state: adjectives expressing states of a person or an entity which are more or less stable for a period of time but can be subjected to change (physical, cognitive, etc.) Ex.: fa: tuberkulozen:2g / fa: tubercular:1g constituting or a icted with or caused by tuberculosis or the tubercle bacillus eng derivative: n: tuberculosis:1; TB:3; T.B.:1 [noun.state] eng derivative: n: tubercle:3 [noun.state] similar to: a: ill:8; sick:7 [adj.state] (15) Adj.cause: adjectives expressing abilities relating to change of state. Ex.: fa: stimulirasht:3g / fa: stimulative:1g capable of arousing or accelerating physiological or psychological activity or response by a chemical agent eng derivative: v: stimulate:1; arouse:3; brace:1; energize:1; energise:1; perk up:1 [verb.body] eng derivative: v: stimulate:2; excite:1 [verb.change] eng derivative: v: stimulate:7; excite:8; stir:9 [verb.perception] (16) Adj.quality: adjectives expressing a property of an entity that is considered more or less an inherent attribute of this entity.
Ex.: fa: kachestven:1g / fa: qualitative:1g involving distinctions based on qualities antonym: a: quantitative:1 (17) Adj.relation: adjectives denoting an explicit relation to an entity such as possession, purpose, function, composition, similarity, etc.
Ex.: fa: mazhki:1g / fa: male:3g for or pertaining to or composed of men or boys eng derivative: n: maleness:1; masculinity:2 [noun.attribute] eng derivative: n: male:5; male person:1 [noun.person] similar to: a: male:2 The adjective expresses an attribute of an entity but in WordNet the attribute can refer to di erent entities { and this can be tracked down the synset tree as well as in the de niton (and examples). For example, the adjectives expressing physical state can express also attributes of a human, an animal, a plant, etc. Many adjectives express attributes pertaining to entities of di erent nature. For example, fperceptive:1g having the ability to perceive or understand; keen in discernment can be used both for human abilities and for actions attributed to human abilities; the synset is linked to nouns classi ed as noun.attribute, noun.cognition, noun.feeling, as well as to verb.cognition. Another example is the adjective fconsumptive:2g { which was classi ed here as adj.state { with the de niton 'a icted with or associated with pulmonary tuberculosis', and examples 'a consumptive patient' (a state of a person { referring to possible classi cation of adj.body ) and 'a consumptive cough' (a hint at adj.cause { 'a cough caused by tuberculosis). There are more than one way to resolve this { we may either split a synset, or classify it into more than one semantic class (or formulate subclasses). These approaches need further consideration. The next section presents the rst stage of our e ort. 4
At the rst stage, we have classi ed the adjectives manually taking into account additional information available from wordnet such as the semantic class of noun and verb synsets linked via lexico-semantic relations to the adjective synsets, the semantic class of the adjectives in the synset tree, and the text in the de nition. A set of 2500 adjective synsets { only the ones marked as adj.all { have been manually classi ed (as of April, 2017) experimentally.
We have used information provided in the de nitions to extract synsets of
adjectives { f.ex. the query [sem class('adj.all')&de nition('in time')] (cf. [
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: h has attribute i; h eng derivative i; h category domain i; h usage domain i { linking synsets in a topical class). Table 1 presents an overview of the semantic classes of the nouns to which some of the adjectives are directly linked (but not indirectly, i.e., via other adjective synsets directly linked to noun synsets)2. Further, an adjective synset is linked to other Adj.class All body behavior perception 265 quantity cognition time feeling cause social location adjective synsets in a synset tree via relations for antonymy (h antonym i), and similarity (via relation h similar to i { for semantic similarity between focal and 2 Table 1 gives numbers for adjective classes with over 150 synsets being classi ed. satellite synsets with close referential meaning; and relation h also see i to link semantically related synsets. Information about semantic classes of adjectives can be obtained also by analyzing these relations. It can be expected that semantically related adjectives are classi ed in the same semantic class (or similar) but there are a number of exceptions as exempli ed in Table 2 for the combinations in synset trees of adj.behavior, adj.feeling, adj.perception and adj.cognition. The proposed classi cation resolves some issues concerning the structure of the Adj.class Adj.class similar to also see adj.cognition adj.behavior 12 3 adj.cognition adj.cognition 87 34 adj.behavior adj.feeling 15 12 adj.behavior adj.behavior 104 61 adj.perception adj.perception 226 28 adj.feeling adj.feeling 72 20
Table 2. Combination of adjective classes in adjective synset trees wordnet and the synset references while leaving others aside for further consideration. First, the de nition of a synset may cover more than one class of referents, as with adj.body where the property can be attributed to bodies of humans, animals, etc. Further, there are adjectives which may express di erentially attributed properties { an issue that is mirrored in the combined synsets in the synset tree { see with adj.behavior and adj.cognition, adj.behavior and adj.feeling in Table 2. The keywords in the de nitions are a telling sign (keywords for experiences or feelings or relations to verb.emotion or noun.feeling signals for adj.feeling ; showing, reacting, behaving, etc. { adj.behavior ; words for causing or relation to verb.cause { adj.cause; animate and inanimate objects related to noun.state or verb.state fall into adj.state; etc.). 5
The paper presented the rst step of an attempt to classify adjectives in WordNet drawing upon information that is already available in the structure of the resource (both in the lexico-semantic relational structure as inherited from the Princeton WordNet, and in content of the de nitions and examples). We are still at the stage of formulating the semantic classes while classifying the candidate synsets manually. The goal of this manual e ort is to extract patterns to help us further with extraction and possible automatic classi cation. 6
The work reported in this paper is carried out within the project Semantic Classi cation of the Adjectives in the Bulgarian WordNet of the Institute for Bulgarian Language which is supported under the Program for Career Development of Young Scientists (2016) at the Bulgarian Academy of Sciences. We would also like to thank the three anonymous reviewers and the participants in the workshop Challenges for Wordnets co-located with LDK 2017: The rst conference on Language, Data and Knowledge.