In this paper we describe the semi-automatic process of transforming the Russian language thesaurus RuThes to WordNet-like thesaurus, called RuWordNet. In this procedure we attempted to achieve two main characteristic features of WordNet-like resources: division of data into part-of-speech-oriented structures with cross-references between them and providing a set of relations similar to WordNet-like resources.
WordNet-like resources [
At least four attempts to create a Russian wordnet are known. RussNet [
For Russian, there exists the RuThes thesaurus, a linguistic ontology, which structure has differences from the WordNet approach. RuThes is a more ontologyoriented resource: thesaurus concepts have unique names, text entries of all parts of speech can be linked to the same concept. The RuThes relations are more formal conceptual relations. The current size of the published version of RuThes (RuThes-lite 2.0), accessible for non-commercial use, is more than 115 thousand text entries. RuThes was specially created for information retrieval and natural language applications, it can be used in most applications where WordNet is usually utilized, but researchers and practitioners want to have a Russian wordnet.
In this paper, we describe the transformation of RuThes data to WordNetlike resource, called RuWordNet. In this process we try to reproduce two main features of the Princeton WordNet structure such as the organization in the form of part-of-speech lexical nets and the basic set of relations. The current volume of RuWordNet is the same as the published version of RuThes-lite 2.0 (115 thousand entries). It can be seen in Internet and can be obtained in the XML format.
The paper is organized as follows. The second section reviews the related work. The third section considers main features of the WordNet structure. The fourth section describes the main structure of RuThes and its differences from WordNet. The fifth section presents the transformation process from RuThes to RuWordNet and achieved results. The sixth section compares web-representations of RuThes and RuWordNet. 2
The most straightforward approach to the development of WordNet-like
resources from scratch is a difficult task, which usually takes years of work. The
approach to fasten the creation of a national wordnet is to translate Princeton
WordNet to the target language [
An intermediate approach between the above-mentioned ultimate points,
which can be considered as quite usual, is to translate the top 5000 concepts
of the Princeton WordNet (core WordNet) and then extend this hierarchy
manually, using local dictionaries. This approach was accepted in the development
of EuroWordnet [
Analysing previous approaches for national wordnet development, authors
of FinnishWordNet (FiWN) decided to use manual translation of Princeton
WordNet synsets by professional translators. The direct translation approach
was based on the assumption that most synsets in PWN represent
languageindependent real-world concepts. Thus, the semantic relations between synsets
were also assumed mostly language-independent, so the structure of PWN could
be reused as well. In such a way, Finnish wordnet, FinnWordNet (FiWN), was
created by translating more than 200,000 word senses in the English Princeton
WordNet (PWN) 3.0 in 100 days [
Braslavski et al [
The structure of Princeton University’s WordNet (and other wordnets) is based on sets of partial synonyms – synsets, organized in hierarchical part-of-speechbased lexical nets for nouns, adjectives, verbs, and adverbs. Each part-of-speech net has its own system of relations between synsets.
The most frequent relation between noun synsets is the hyponym-hypernym
relation. Also since 2006 in Princeton WordNet class-instance relations denoted
as Instance Hypernym and Instance Hyponym [
The noun relationships also include part-whole relations, which are subdivided into proper part-whole relations (wing is a part of bird), member parts (tree is a member of forest), and material (snow is a substance of snowball). Parts can have several wholes (wing is a part of bird, bat, insect, or angel).
For all parts of speech, the lexical relation of antonymy can be established. Lexical relations link lexemes, not whole synsets. In Princeton Wordnet, the antonymy relation is the main type of relations for descriptive adjectives [11], which were described only with the relations of antonymy and similarity. For example, for the word heavy, the word light is indicated as an antonym, such words as hefty, ponderous, massive are linked to heavy with the relation "similar to". Other wordnets, such as GermaNet [12] or Polish WordNet (PlWordNet) [13], changed this approach and introduced taxonomic relations (hyponymyhyperonymy) for adjectives.
Verbs in WordNet are mainly linked with hyponym-hypernym relations.
Besides, they have their own unique relations not described for nouns or adjectives:
entailment (buy – pay) and causation (give – have, kill – die). The WordNet
entailment relation is a relation between two verbs V1 and V2 that holds when the
sentence "Someone V1" logically entails "Someone V2" and there is the temporal
inclusion of eventV1 into V2 or vice versa [
RuThes and WordNet are both thesauri that are lexical resources where semantically related words and expressions are collected together into synsets or concepts between which formalized relations are set. When applying both thesauri to natural language processing, the same steps should be made such as matching between a text and a thesaurus and employing the described thesaurus relations if necessary. The most evident differences between the two types of resources are as follows.
First, in RuThes there is no division into subnets according to different parts of speech that is words of any part of speech can be linked to the same concept if they mean the same (so called derivative or part-of speech synonyms).
Therefore, second, in RuThes it is often very difficult or even impossible to apply traditional tests of synonymy detection such as substitution of synonyms in sentences [14,15]. Tests checking the denotational scope of lexemes are usually applied in the following way: "if entity X can be called with word W1, then we can call it also with word W2" and vice versa regardless of specific context. The second test consists in formulation of explicit differences of one concept from other concepts. These differences can be fixed in the unique concept name. Thus, above-mentioned issues of RuThes such as denotational tests, denotational distinctions between concepts, and unique names of concepts make RuThes much closer to ontological resources in an imaginary scale from lexical resources to formal ontologies than WordNet-like thesauri. RuThes can be called a linguistic (lexical) ontology for natural language processing.
Third, the relations in RuThes are only conceptual, not lexical (as antonyms or derivational links in wordnets). They are constructed as more formal, ontological relations of traditional information-retrieval thesauri [16], which were designed to describe very broad, unstructured domains. The set of conceptual relations includes: - the class-subclass relation; - the part-whole relation applied with the following restriction: the existence of the concept-part should be strictly attached to the concept-whole. For example, trees can grow in many places not only in forests therefore concept tree cannot be directly linked to concept forest with the part-whole relation, the additional concept forest tree should be introduced;
- the external ontological dependence when the existence of a concept depends on the existence of another concept (in such a way forests depend on the existence of trees) [17]. In RuThes we denote this relation as association with indexes: asc1 is directed to the main concept, asc2 – to the dependent concept; - In the very restricted number of cases symmetric associations between concepts can be established.
The main idea behind this set of relations is to describe the most essential, reliable relations of concepts, which are relevant to various contexts of concept mentioning. Also this set of relations allows us to describe domain terminologies or domain-specific ontologies, combine descriptions of lexical and domain-specific knowledge in the same resource.
The relation of ontological dependence is very convenient for describing conceptual relations between concepts corresponding to multiword expressions and concepts of their component words (such as nature protection and nature), which allows easier introducing such concepts and describing useful "horizontal" relations.
Thus, RuThes has considerable similarities with WordNet: the inclusion of concepts based on senses of real text units, representation of lexical senses, detailed coverage of word senses. At the same time the differences include attachment of different parts of speech to the same concepts, formulating of names of concepts, attention to multiword expressions, the set of conceptual relations, etc. The more detailed description of RuThes and RuThes-based applications can be found in [18] or [19].
At present RuThes includes 54 thousand concepts, 158 thousand unique text entries (75 thousand single words), 178 thousand concept-text entry relations, more than 215 thousand conceptual relations. The published version of RuThes, RuThes-lite 2.0, contains 115 thousand text entries. It was singled out from full RuThes on the basis of words and phrases used in current Russian news flows with exclusion of several specific domains [20]. 5
According to the guidelines of world-known WordNet thesaurus, the first version of Russian wordnet (RuWordNet) was created.
In our opinion, one of the most distinctive features of WordNet-like resources is their division into synset nets according to parts of speech. Therefore all text entries of RuThes-lite 2.0 were subdivided into three parts of speech: nouns (single nouns, noun groups, or preposition groups), verbs (single verbs and verb groups), adjectives (single adjectives and adjective groups). We have obtained 29,297 noun synsets, 12,865 adjective synsets, and 7,636 verb synsets.
This subdivision was based on the morphosyntactic representation of RuTheslite 2.0 text entries, which was fulfilled semi-automatically. Therefore, a small number of mistakes because of particle treatment (verbs or adjectives) or substantivated adjectives can appear. Currently all found mistakes are corrected. The divided synsets were linked with the relation of part-of-speech synonymy.
The hyponym-hypernym relations were established between synsets of the same part of speech. These relations include direct hyponym-hypernym relations from RuThes-lite 2.0. In addition, the transitivity property of hyponymhypernym relations was employed in cases when a specific synset did not contain a specific part of speech but its parent and child had text entries of this part of speech. In such cases the hypernymy-hyponymy relation was established between the child and the parent of this synset.
Similar to the current version of Princeton WordNet, in RuWordNet classinstance relations are also established. By now, they had been generated semiautomatically for geographical objects.
The part-whole relations from RuThes were semi-automatically transferred and corrected according to traditions of WordNet-like resources. Now RuWordNet contains 3.5 thousand part-whole relations. The part-whole relations include the following subtypes: - functional parts (nostrils – nose), - ingredients (additives – substance), - geographic parts (Sevilia – Andalusia), - members (monk – monastery), - dwellers (Moscow citizen – Moscow), - temporal parts (gambit – chess party) - inclusion of processed, acitivities (industrial production – industrial cycle) Adjectives in RuWordNet similarly to German or Polish wordnets are connected with hyponym-hypernym relations. Adjectives often have POS-synonymy links to nouns, but also can have POS-synonyms to verb synsets.
In the current RuWordNet representation of Russian verbs, part-whole
relations can be seen. For example, synset видеть во сне, сниться, грезиться,
присниться, привидеться во сне, пригрезиться, пригрезиться во сне" [to
dream] is linked to synset спать, поспать, доспать, соснуть, досыпать, почивать,
проспать, просыпать [to sleep] with the part-whole relation. Such a relation
between the translation equivalents [to dream, to sleep] exists also in Princeton
WordNet and called ’entailment relation’. Christian Fellbaum wrote in [
Antonymy relations are conceptual relations in RuWordNet, that means they link synsets, not single lexemes. They are introduced for all parts of speech, mainly for synsets denoting properties and states. 6
RuThes-lite 2.0 and RuWordNet are published in form of static web-pages. Looking through RuThes1, the user should select a letter to begin, then choose an initial trigram of a word, and then click on a proper word. For example, selecting Russian word двор [yard] the user can find three concepts associated with this word, relations of these concepts, and other text entries attached to the same concepts. Further, the navigation through concepts or text entries is possible.
In the similar representation of RuWordNet2, there is the initial division to parts of speech, which the user should select, then the user should find a word. In the RuWordNet representation, there are no concepts, each synset contains text entries belonging to the same part of speech, POS-synonymy links to other parts of speech are indicated. Thus, in the representation RuThes looks more as an ontology, and RuWordNet is presented more as a lexical net.
In this paper we have described the semi-automatic process of transforming the Russian language thesaurus RuThes (in version, RuThes-lite 2.0) to WordNetlike thesaurus, called RuWordNet. In this procedure we attempted to achieve two main characteristic features of wordnet-like resources: division of data into partof-speech-oriented structures with cross-references between them and providing a set of relations similar to wordnet-like relations.
Both thesauri, RuThes-lite 2.0 and RuWordNet, are currently published as static web-pages. Also RuWordNet can be seen through web interface1. Researchers can obtain both types of thesauri, compare them in applications. In future, we will continue to add new types of relations to RuWordNet including the domain relation, the cause relation, the entailment relation, etc. Acknowledgments. This work is supported by the Russian Science Foundation (grant N16-18-02074).
11. Gross, D., Miller, K.J.: Adjectives in WordNet, International Journal of Lexicography, 3(4), 265–277 (1990) 12. Kunze, C., Lemnitzer, L.: Lexical-Semantic and Conceptual relations in GermaNet, In Storjohann P (ed) Lexical-semantic relations: Theoretical and practical perspectives, 163–183 (2010) 13. Derwojedowa, M., Piasecki, M., Szpakowicz, S., Zawisawska, M., Broda, B.: Words, concepts and relations in the construction of Polish WordNet, In Proceedings of the Global WordNet Conference, Seged, Hungary, 162–177 (2008) 14. Cruse, D.: Lexical Semantics. Cambridge. University Press (1986) 15. Miller, G.: Nouns in WordNet. In WordNet: An Electronic Lexical Database, Fellbaum, C (ed). The MIT Press, 23–47 (1998) 16. Z39.19. Guidelines for the Construction, Format and Management of Monolingual
Thesauri. NISO (2005) 17. Guarino, N., Welty, C.: Evaluating ontological decisions with ONTOCLEAN, Communications of the ACM, 45(2), 61–65 (2002) 18. Loukachevitch, N., Dobrov, B.: RuThes Linguistic Ontology vs. Russian Wordnets.
In Proceedings of the Seventh Global WordNet Conference (GWC 2014), 154–162 (2014) 19. Lukashevich, N.: Thesauri in information-retrieval tasks. Moscow (2011) (in Russian) 20. Loukachevitch, N., Dobrov, B. , Chetviorkin, I.: RuThes-Lite, a publicly available version of thesaurus of Russian language RuThes, In proceedings of International Conference on Computational Linguistics and Intellectual Technologies Dialog2014, 340–350 (2014)