Concept Formation in Linguistic Ontologies
Abstract. Problems of conceptualization in linguistic ontologies are discussed We show that it is necessary to form concepts of a linguistic ontology as close as possible to the meanings of linguistic units, because excessive generalization and clustering of meanings necessarily lead to distortions in the system of relations, excessive problems in a specific subject field, or an application. At the same time it is important to ensure that concepts can be distinguished from superconcepts and sibling concepts. The usage of really existing multiword expressions helps us mitigate these contradictory requirements. The introduction of concepts on the basis of multiword expressions does not change the essence of a linguistic ontology, but also makes the distinction between the concepts much clearer.
Key words: thesaurus, linguistic ontology, conceptualization
1 Introduction
An ontology is often considered to be independent of a natural language [1,2,3]. D. Lenat [4] emphasizes that taking the meaning of words into account can only confuse (“words are often red herrings”), the meanings of words divide the world ambiguously, and the division lines come from a variety of reasons: historical, physiological, etc.
From another point of view an ontology can not be fully independent of natural language. Names of the concepts in ontologies are often formulated in natural language – this is a standard practice that has been used in knowledge representation systems in artificial intelligence [2,5]. Breuster et.al [6] stress that people manipulate concepts through words. In all known ontologies the words are used to represent concepts. Therefore, phenomena that are not verbalized, can not be modelled. The Breuster et.al characterize this phenomenon as the Ontological Whorf-Sapir hypothesis, i.e. “that which can not be captured by words cannot be represented in an ontology.” Y. Wilks [2,5] asserts that the symbols in representation languages are fundamentally based on the natural language, that a representation language is a means of human communication with the inherent dynamics, polysemy and possibility of extended interpretation.
Moreover, in ontological resources developed for specific domains most ontology concepts are related to the meanings of domain terms. So J. Tsujii and S. Ananiadou [7] stress that “in many fields of application, knowledge to be shared and integrated is presented mostly in text”. Many ontologies in biology, such as GO (Gene Ontology) [8], actually represent the information-retrieval thesauri (controlled vocabularies), and they by nature differ from ontologies, required within the formal ontological approach to knowledge description.
As a result a paradoxical notion of linguistic ontology emerges, i.e. an ontology, concepts of which are considerably related to the meanings of linguistic units, the terms of the subject field [9, 10]. Linguistic ontologies cover most of the words of the language or the subject field, and at the same time they have an ontological structure represented in relations between the concepts. Therefore, a linguistic ontology can be considered as a special kind of a lexical database and a special type of an ontology. Linguistic ontologies are relatively weakly formalized, i.e. they belong to the “terminological” ontologies according to J. Sowa [11]. The role of the ”linguistic ontologies” increases greatly in applications related to natural language processing.
Examples of linguistic ontologies are the Princeton WordNet [12] and wordnets of other languages. Information retrieval thesauri can also be considered as linguistic ontologies.One of the serious problems in the linguistic ontology development consists in formulation of the principles for the concepts formation, since the relations between the concepts and lexical meanings are quite complex. Understanding these issues is important for developers of any types of ontologies, because creation of any ontology deals more or less with lexical or terminological meanings.
In this paper we will describe problems of concept formation in linguistic ontologies. We will consider principles for introducing a new concept in such linguistic ontologies as Princeton WordNet, MikroKosmos ontology [2], information-retrieval thesauri. Finally, we present our approach to description of concepts in Thesaurus of Russian language - RuThes, which we also consider to be a linguistic ontology [13].
RuThes is used in information-retrieval applications, such as conceptual indexing; automatic text categorization, document clustering, automatic text summarization, question-answering. At present it includes more than 50 thousand concepts and more than 140 thousand Russian words and multiword expressions. It was translated into English and comprises almost 130 thousand English words and expressions.
2 Principles of Concept Introduction and Lexical Senses
The general recommendations on the ontology concepts formation are usually described as follows [3,14,15]. One needs to distinguish the concept and its name, i.e. synonyms of the same concept do not represent different classes, synonyms are just different names of the concepts.
A child concept should be distinctly different from the parent one. This difference can be expressed in the form of a distinctive property that the child concept bears, in limitations on slot fillers that are distinct from limitations of other classes; or in the existence of additional relationships with other concepts. A concept must be clearly distinguished from the concepts at the same level (sibling concepts).
These recommendations are not easy to realize if the developed ontology is based on the existing linguistic meanings. First of all, it is not easy to distinguish a concept and its names, working with linguistic meanings. Secondly, a critical challenge for resources designed for the natural language processing is the presentation of ambiguous words, especially if the meanings are closely related to each other. Moreover, a serious problem is caused by the words with similar meanings, or near-synonyms, the meanings of which can differ in several features (conceptual content, speaker’s attitude, collocations, etc.), and be dependent on the context.
In the following sections we will discuss the way these problems are solved in specific linguistic ontologies. We will use the following notation: CONCEPT, term or word, ‘meaning’, ’transcription of Russian words’.
3 Confusion of a Concept and its Name in Linguistic Ontologies
3.1 Confusion of a Concept and its Name in WordNet
Initially, WordNet was considered to be a lexical rather than an ontological resource. However, over time, the growing importance of the ontological research, as well as the similarity of the WordNet nouns hierarchy with an ontology became apparent [16]. At the same time there exist a lot of deficiencies of WordNet descriptions from ontological point of view [17].
Numerous examples of confusion between concept and its name can be found in WordNet. This is due to the fact that the basic relation in WordNet is the synonymy. Sets of synonyms – synsets – are the main structural elements of WordNet. Definitions of synonyms in synsets are based on the principle of substitution of one for another in sentences [12]. This basic principle of the WordNet construction leads to the situation when different synsets are introduced for different ways of naming same entities.
There are several types of confusion of concepts and their names in wordnet-like resources.
First of all, the confusion of concepts and their names shows itself in the support of different hierarchies for different parts of speech. Indeed, if, for example, the Privatization concept (privatize, privatization) is mentioned in a text using whatever parts of speech – it is always a reference to the same concept with the help of different lexical means, the parts of speech changing should not affect the relations between this concept and other concepts.
The first Wordnet followers (EuroWordNet project) considered the integration of all parts of speech–derivatives in a single synset, since such division is contrary to the principles of the development of ontological resources. However, the decision to connect parts of speech in the same hierarchy was not made [18].
The second type of confusion of a concept and its name in WordNet is the usage of different synsets to describe the old and new names, the names of concepts in different dialects of the language, in different text genres, etc.
According to ontological principles of the distinction of an entity and its name, all these different names should not produce new units of representation or concepts; they should remain to be textual expressions of the same concept. To differentiate the features of their linguistic use, they can be provided with extra labeling.
Thus, in the Princeton WordNet, one can find numerous examples of synsets, which appeared due to different characteristics of the use of words. For example, for describing slang synonyms of the word “nose”, a special synset is introduced. This synset is presented as hyponym of nose synset.
beak, honker, hooter, nozzle, snoot, snout, schnozzle, schnoz -- (informal terms for the nose)
Informal words related to money are also collected in a separate synset:
boodle, bread, cabbage, clams, dinero, dough, gelt, kale, lettuce, lolly, lucre, loot, moolah, pelf, scratch, shekels, simoleons, sugar, wampum -- (informal terms for money)
Several synsets fix specific features of English dialects, as a special synset for domestic ass in British English:
Moke 1 -- (British informal)
domestic ass, donkey, Equus asinus -- (domestic beast of burden descended from the African wild ass; patient but stubborn)
Next example of confusion of a concept and its names consists in description of monetary units with the same name used in different countries such as franc or centime:
franc -- (the basic monetary unit in many countries; equal to 100 centimes)
centime -- (a fractional monetary unit of several countries: France and Algeria and Belgium and Burkina Faso and Burundi and Cameroon and Chad and the Congo and Gabon and Haiti and the Ivory Coast and Luxembourg and Mali and Morocco and Niger and Rwanda and Senegal and Switzerland and Togo)
From the ontological point of view such synsets are not valid, because similarity between different monetary units is only their names, they are different in value . Therefore concepts should be introduced for such entities as Swiss franc, French franc, American dollar, Canadian dollar, and so on.
3.2 Differentiation of Ontology and Lexicon in the MikroKosmos Ontology
The authors of the MikroKosmos [3] ontology make a clear distinction between an ontology and a lexicon. The concepts of the ontology are described as frames – sets of slots. The system’s lexicon describes the meaning of words and phrases, by establishing links from them to the ontology concepts. This division mainly prevents confusion of a concept and its name.
The MikroKosmos ontology is relatively small; it contains about 6 thousand concepts. The lexicon contains tens of thousand linguistic expressions. A lexicon entry can have a simple structure - a reference to an ontology concept, or a rather complex structure, containing both a reference to an ontology concept and features of a particular lexical unit. The names of a concept in the ontology can look like English words or phrases, but their semantics is expressed by a set of well-defined relations between concepts.
The authors declare the independence of ontology from a specific natural language that manifests itself in two aspects:
1) the ontology contains no units specific to a language such as English or Spanish, although the names of concepts are given in English for the sake of convenience.
2) the concepts of ontology do not have one-to-one mapping to word senses in natural languages. Many concepts may not be mapped to any word in the language; other concepts may correspond to several words in the same language and vice versa.
The main stages of the ontology development include:
· assessment whether the meaning of a word gives a sufficient ground for the introduction of a new concept;
· location of the concept in the ontology, identification of existing concepts which can be described as generic concepts for the new one;
· description of the new concept features that should differ from the properties of superconcepts and subconcepts. These features are given not just by slots filling, but also in a more informative way, e.g., in the availability of other properties or relations to other concepts.
Thus, the proclaimed linguistic independence should not be misleading. At its core, the MikroKosmos ontology is certainly a linguistic ontology, because, the basic principle, which justifies the introduction of new concepts, is the existence of words with the same meaning in many languages.
At the same time, the principle of linguistic independence of this ontology stresses that in the construction of a linguistic ontology it is not necessary to follow the system of meanings of a specific language. A linguistic ontology can take into account the system of meanings of a particular language or an aggregate of languages, and in doing so adhere to the ontological principles of the concepts introduction.
3.3 Concepts and Terms in Information Retrieval Thesauri
Information retrieval thesauri are usually considered as a kind of ontological resources [19]. In addition, thesauri are based on the terms of a subject field, so they can also be considered as linguistic ontologies. Conventional information-retrieval thesauri regulated by national and international standards [20, 21] are intended to be used in manual indexing by human indexers.
The basic thesauri units are terms that are divided into descriptors (= authorized terms) and non-descriptors (= ascriptors). Most standards for information-retrieval thesauri highlight the connection between the terms and concepts of a subject field. The American standard points out that a term is one or more words referring to a concept. The ISO standard [20] emphasizes that an indexing term is a concept presentation, preferably in the form of a noun or a noun phrase. A concept is considered as a unit of a thought, mentally formed to reflect some or all of the properties of a concrete or an abstract, real or mental object. Concepts exist as abstract entities, regardless of the terms that express them.
It should be noted that not all thesauri developers distinguished concepts and terms. Thus, developers of AGROVOC thesaurus characterize their resource as term-oriented; it is manifested in the fact that a term cannot be complemented with synonyms. This feature of the thesaurus is considered by the authors as a disadvantage that must be corrected [22].
An important property of a descriptor is that it should be formulated explicitly, its implied meaning in the thesaurus should be clear to the user. If an unambiguous and clear descriptor cannot be found, the term, taken as a descriptor is supplied with a “relator” (a brief note) or a comment.
American standard Z39.19 [21] recommends to use relators for descriptors’ names, even when a descriptor sounds uniquely within a given subject field, but has different meanings in the general language, or other domains. This makes it easier to search through multiple databases and to compare descriptors of various subject fields. For example, it is proposed to introduce Shells (structures) descriptor for the engineering subject field, since the word shell has a lot of meanings in English.
4 Similar Meanings of Ambiguous Words
The existence of closely related meanings of ambiguous words presents major difficulties for developers of linguistic ontologies. The difficulty of automatic disambiguation of ambiguous words requires to formulate principles for the description of such sets of related meanings as concepts of a linguistic ontology.
4.1 Similar Meanings of Ambiguous Words in WordNet
Many authors admit that the differences of meanings in WordNet are too fine-grained for such applications as machine translation, information retrieval, text classification, question-answering systems, etc. In [23] it was indicated that the average number of senses in WordNet is larger than in traditional lexicographical dictionaries.
The number of senses of certain lexical items may vary in different lexical resources, dictionaries. However, a large number of meanings in WordNet causes difficulties in applications related to natural language processing and brings about the question of how and what senses can be combined (“clustered“) [24, 25] to use them in the applications.
Gonzalo [26] pointed out that the experiments on sense clustering led to a conclusion that the typology of the relations between different senses of ambiguous words is more useful than the formation of sense clusters, because sense proximity depends on the application. For example, metaphoric senses belong to different semantic fields therefore distinction of such senses is very important for information retrieval applications and question-answering systems. However, for machine translation applications, this distinction may be unimportant, as the metaphorical transfer may be similar in different languages.
Fellbaum and Miller [27] review attempts to cluster senses of WordNet. They emphasize that the sense clustering can be based on a variety of alternative criteria (semantic, syntactical, domain-oriented), which apparently confirms the significance of different sense clusters for natural language applications.The problem of automatic selection of WordNet senses in practical applications can be mitigated through the use of semantically-marked, according to WordNet meanings, corpus SemCor [28].
The OntoNotes project proposed its own way of integrating lexical meanings of an ambiguous word and concepts formation [29], which is based on the consideration of the use of an ambiguous word in a corpus, representing the majority of its senses. All examples of use are analyzed and divided into groups of senses most distant from each other, a branch node in the tree is created, and then for each node the process should be iterated.
The development of the ontology of word senses depends on the explanatory need or application requirements. A standard termination condition for the ontology development process is the absence of an obvious way of splitting the remaining group of senses into subgroups, or the existence of equally reasonable ways of splitting it into subgroups according to different reasons. The work also highlights the usefulness of a multilingual consideration for the appropriate separation of the meanings and concepts.
Consider an example of the verb drive, for which WordNet provides 22 individual senses. In OntoNotes project two independent experts defined 7 most important groups of meanings of this verb. The most frequent group of senses comprises seven senses from WordNet and can be called “Operating or traveling by means of a vehicle”:
WN1: Can you drive a truck?
WN2: drive to school,
WN3: drive her to school,
WN12: this truck drives well,
WN13: He drives taxi,
WN14: The car drove around the corner,
WN16: Drive the turnpike to work.
It should be stressed that, e.g. from the viewpoint of a Russian native speaker this group of senses is not very evident because these senses correspond to 5 different (non-synonymous) Russian words: ‘vodit’ (WN1, WN13), ‘ehat’ (WN2, WN12), ‘vezti’ (WN3), ‘povernut’ (WN14), ’proehat’ (WN16).
4.2 Similar Meanings of Ambiguous Words in the MikroKosmos Ontology
The basic rule, proclaimed for dealing with similar meanings of ambiguous words in the MikroKosmos ontology, is the reduction of polysemy rule [3]: it is necessary to decide how many dictionary meanings a particular lexical sense can represent, and to unite as many meanings as possible, so there will be as few different senses as possible.
The principles of distinguishing meanings are as follows:
· A candidate meaning must be clearly distinguishable from the already described meanings.
· It is necessary to check whether the meaning needs further clarification, if it is used in a short sentence. If one needs additional context to figure out what meaning is used, the meaning should not be introduced, but must be attributed to one of the existing meanings.
· It is necessary to check whether there is a property in the description of the meaning, which is filled with too small number of fillers. If so, the meaning must also be assigned to one of the more common meanings, or described as a multiword expression.
Seemingly, this procedure should reduce the problem of sense disambiguation, but on the other hand, it leads to violations of the ontology structure. Thus, N. Guarino [17] criticizes several existing ontologies, including the MikroKosmos ontology, for the polysemy of ontological nodes, e.g., for the treatment of the Window concept as an artifact and as a place at the same time.
The problem resides in the fact that window in different contexts may denote an opening (A man looked out of the window) and an artifact (The workers mounted a window), and these entities Window (Opening) and Window (artifact) are very closely linked to each other. This criticism relates to the fact that, in Guarino opinion, the polysemy in ontological nodes should not be permitted in any form. To conform to the principle of forbidding polysemy nodes, the ontology should have different nodes at different locations for such concepts as Window (Artifact) and Window (Opening).
To reply to N. Guarino, the authors of the ontology [3, p. 129] explain that the fact of the English word window having two meanings is not crucial for the ontology development, since it is not considered that the relations between the meanings of the natural language (or, more precisely, the meanings of all known languages) and the ontology concepts should have a unique correspondence.
As a justification for their position the authors argue that they do not know such a natural language, in which the word for the Window concept does not realize two meanings: the meaning of an opening and the meaning of an artifact. This semantic genericity is the strongest argument in favour of the fact that people can combine these two concepts. The authors of the ontology also emphasize that “an effort to split ontological concepts into ever smaller unambiguous units leads to a sharp increase in polysemy and, therefore, makes the task of disambiguation so much more difficult... So if the ontology is made less ambiguous, it only means that the ambiguity will have to be treated increasingly elsewhere” [3, p. 132].
However, we need to agree with N. Guarino that violations of the ontological structure is also a serious problem because, if it is required to use the described relationships between the concepts for the logical inference, then at first it will be necessary to determine if these relationships could be applied in current context, which means that the problem of choosing the right meaning of an ambiguous word simply shifted to another stage of text processing.
4.3 Lack of Similar Meanings of Ambiguous Words in Information Retrieval Thesauri
Recently terms were considered to be unambiguous, context-independent linguistic units, now it is known that the terms have many features of the common language units, in particular, the terms can also be ambiguous, and their meanings can be very similar, for example, industrial production (the process and the outcome).
However, as the traditional information retrieval thesauri are not intended for the automatic processing of texts, usually only one of related term meanings is presented in an information-retrieval thesaurus.
5 Near-Synonyms in Linguistic Ontologies
The problem with near-synonyms, i.e. different words with similar meanings, is that they can differ in many features: denotative content, language register, evaluation, dialect, collocations, etc. This justifies the existence of a special genre of “dictionaries of synonyms”, which explain in detail the specificity of using synonyms.
For many of these sets of near-synonyms it is extremely difficult to establish a unique correspondence in other languages, because in another language the corresponding set of near-synonyms is characterized by its own system of parametric differences and, accordingly, its own specificities.
Although the linguistic ontology takes into account the existing lexical meanings, nevertheless it should remain to be an ontology. According to general principles of ontological hierarchy (see Section 2) its main elements – the concepts -- should have clear, context-independent differences from the related concepts.
G. Hirst [30] explains that for the description of the words with similar meanings in the linguistic ontologies, it is necessary to implement one more level of representation, a conceptual-semantic level. This level should specify a relatively coarse conceptual hierarchical system, which is based on denotative, context-independent properties of words. Each concept is linked to a set of near-synonyms, and their features (stylistic, evaluation, connotations, etc.) are described in additional intra-conceptual structures.
However, the denotative component of the meaning is often very difficult to separate from other components. For example, consider the problem of determining the optimal number of concepts (and principles of it) to be associated with the following set of words with the meaning ‘error’: error, fault, omission, oversight, blunder, mistake, miss, screw-up, dereliction, defect.
The authors of work [31] point out that it is often very difficult to determine the words with similar meaning that should better be described as a part of the internal structure of concept, and which should belong to different concepts. On the one hand, linguist’s intuition can help. On the other hand, a look at the conceptual structure in terms of another language can really help to better delimit the boundaries of the concepts.
5.1 Near-Synonyms in WordNet
To describe the relationship between meanings according to the principles of the possibility of synonymous substitutions in same sentences, as it was made during the creation of Princeton WordNet, means that near-synonyms should be classified on several grounds, as the synonymous substitution of the word must take into account conceptual, stylistic, attitudinal and other components of the meaning. It is clear that the development of a hierarchy on such grounds is impossible; the whole construction becomes very volatile during the transition from language to language.
Therefore, WordNet has a large number of synsets that are difficult to distinguish from one another; this violates the ontological principles of descriptions of the concepts. For example, there are four different synsets denoting likeness, similarity, each next synset is a hyponym of the previous, which is hardly distinguishable from its hyperonym:
sameness -- (the quality of being alike)
similarity -- (the quality of being similar)
likeness, alikeness, similitude -- (similarity in appearance or character or nature between persons or things)
resemblance -- (similarity in appearance or external or superficial details)
5.2 Near-Synonyms in MikroKosmos Ontology
In the MikroKosmos ontology large sets of near-synonyms are related to the same concept of ontology, their specific features are described in the lexicon [3].
The authors provide an example: all the ‘change’ verbs are assigned to the same concept of Change-event. The features of the words are described in lexicon entries, for example, for the verb to increase it is pointed out that the THEME semantic role of the verb should be presented by a scalar_value (for example, price or height) and the value of this quantity is changed to a larger one. The meaning of the word Zionist is represented in the dictionary as a POLITICAL_ROLE, which is an AGENT_OF a SUPPORT_EVENT, the theme of which is Israel. The meaning of the word to asphalt is described as a COVER_event, an instrument of which is the asphalt concept.
Using the Web site of the ontology (http://ilit.umbc.edu), we can see that the situation with the implementation of the principles is quite complicated. The concept CHANGE_EVENT is associated to a long list of words in the lexicon. In the list there are such words as acclimatization, commerzialization, contamination, damage, deteriorate, improve, and many others – there are no separate concepts for these words.
At the same time, the following concepts can be seen lower in the hierarchy: ADJUST, CORRECT-EVENT, DIVIDE, INTEGRATE, RESTRUCTURE, etc. It is not clear why separate concepts were introduced for some of the words, but were not introduced for the others. Why the meaning of word acclimatization does not deserve an independent concept, although there are important relations to climate, biological processes, but the meaning of adjust has received a concept?
In addition to questions of consistency/inconsistency in the description, there are clear consequences for the natural language processing applications. Thus, it is difficult to establish what words from a larger list of lexical entries to the concept of CHANGE-EVENT can be regarded as synonyms, and what are the relations between other words. Besides, one cannot specify the relations between, e.g., the asphalting and road works.
In addition, in domain-specific applications the relatively small size of the ontology leads to the introduction of additional concepts even for words that are already included in the lexicon.
Thus, we believe that in MikroKosmos ontology the problem of near-synonyms is being solved by overgeneralization, which can lead to problems in real subject domains. In our opinion, it is necessary to implement an additional level of concepts, which would help to divide the words more clearly, not dumping them into large chunks.
5.3 Near-Synonyms in Information Retrieval Thesauri
In information-retrieval thesauri, each descriptor, which most often corresponds to a concept, combines several ascriptors that are considered as equivalent. Ascriptors are of three sub-types [20,21]:
· actual synonyms;
· lexical variations,
· quasi-synonyms.
Lexical variants differ from synonyms in that they represent some modification of the same expression, for example, different spelling, abbreviations, etc.
Quasi-synonyms are terms, the meanings of which generally differ, but are regarded as equivalents for the purposes of the thesaurus, for example, antonyms are often regarded as quasi-synonyms (nuclear danger – nuclear safety). Other frequent type of quasi-synonyms is the case where some integrating type is considered as a descriptor, and its subspecies are described as ascriptors.
For example, in the thesaurus LIV of Research Service of the U.S. Congress [32] descriptor Transplantation of organs, tissues etc. contains such ascriptors as medical transplantation, organ transplantation, skin grafting, surgical transplantation, tissue transplantation, some of which could be considered as subordinate concepts (Skin grafting).
In the same thesaurus, the term deflation is included as an ascriptor in the thesaurus entry of the descriptor inflation, because the developers believe that these are different manifestations of a more general concept. Typically, thesauri authors prefer introducing more quasi-synonyms for concepts, regarded as peripheral to the basic domain of the developed thesaurus.
In addition, standards and guidelines for the creation of information-retrieval thesauri often recommend not to include some kinds of terms into a thesaurus.
Relatively low frequency terms can be removed from the list of candidate terms, or represented as ascriptors for more common or more frequent concepts. Too specific terms may also be excluded from the list, since it is believed that if a thesaurus contains too many hierarchy levels, it is difficult to manage. In particular indexing subjectivity increases, since indexers can use descriptors of different levels for documents indexing [21]. If several terms with similar meanings were revealed, it is necessary to choose the most representative one; the remaining terms can be partially removed and transferred to ascriptors.
Thus, in traditional information retrieval thesauri some near-synonyms are excluded from consideration, other near-synonyms are introduced as ascriptors. The exclusion of near-synonyms from a thesaurus, made for the sake of convenience and reduction of the subjectivity of manual indexing, results in decline of search quality if the information retrieval thesaurus is used in automatic modes of document processing.
Table 1 summarizes above-mentioned distinctions in concept formation approaches.
Table 1. Specific features of concept formation approaches in linguistic ontologies
|
Problems of concept formation |
WordNet |
MikroKosmos |
Information-retrieval thesauri |
|
Confusion of concepts and their names |
Concept and names are often confused |
Concept and names are rarely confused |
Concept and names are rarely confused |
|
Representation of related senses of ambiguous words |
Very detailed description of senses |
Related senses of ambiguous words are rarely described |
No related senses |
|
Relations between related senses |
There are no relations between related senses |
Related senses are generalized to the same concept |
- |
|
Near-synonyms |
Sets of near-synonyms are arbitrarily split to synsets |
Near-synonyms are generalized to the same concept |
Near-synonyms are absent or presented as ascriptors to the same descriptor |
6 Concepts and Senses in Thesaurus of Russian Language RuThes
RuThes Thesaurus of Russian language [13] can be called a linguistic ontology for the automatic text processing, i.e. an ontology, where the majority of concepts are introduced on the basis of actual linguistic expressions.
RuThes is a hierarchical network of concepts. Each concept has a name, relations with other concepts, a set of linguistic expressions, i.e., text entries (words, phrases, terms), the meanings of which correspond to the concept.
In construction of the Thesaurus we combined three different methodologies [12, 20, 21, 15, 17]:
· the methods of construction of information-retrieval thesauri (information-retrieval context, analysis of terminology, terminology-based concepts, a small set of relation types)
· the development of wordnets for various languages (word-based concepts, detailed sets of synonyms, description of ambiguous text expressions)
· ontology and formal ontology research (concepts as main units, strictness of relations description, necessity of many-step inference).
The main types of relations are taxonomic relations and a specific set of conceptual relations based on ontological dependence relations [33, 34]. This set of relations was experimentally confirmed to be effective in information-retrieval applications [35, 36].
In the following sections we will present the principles of concept description in RuThes.
6.1 Scope of Concepts in Thesaurus RuThes
Most concepts in RuThes are associated with the meanings of linguistic expressions.
Linguistic expressions that may give rise to an independent concept in the Thesaurus RuThes belong not only to the general vocabulary, but also can be terms of specific subject domains within the scope of social life (economy, law, international relations, politics), and of the infrastructure (transport, banks, etc.), so-called socio-political domain (Fig.1). This is due to the fact that many professional concepts, terms, and slang of these domains penetrate easily into the general language, and can become widely discussed in mass media [37].
Fig.1 Specific domains vs. Socio-political domain
Multiword expressions are also actively used as concept sources in RuThes. The basic principle of introducing this kind of concepts is the need to record some additional information that cannot be described on the basis of component word concepts.
6.2 Concept of Ontology is not a Synset
In RuThes, a unit is presented not by a set of similar words or terms, as it is done in the WordNet thesaurus, but by a concept – as a unit of thought, which can be associated with several synonymic language expressions.
Words and phrases, the meanings of which are represented as references to the same concepts of the thesaurus, are called text entries. Text entries of a concept can be:
· words that belong to different parts of speech (stabilization, stabilize, stabilized);
· linguistic expressions relating to different linguistic styles, genres;
· single words, idioms, free multiword expressions, the meanings of which correspond to this concept.
Concepts often have more than 10 text entries including single nouns, verbs, adjectives and noun or verb groups. For example, a set of English text entries of concept JUDICIAL COURT looks as follows:
court, court authorities, court instance, court of judiciary, court of jurisdiction, court of justice, court of law, judicature, judicial bodies, judicial court, judicial organ, judicial tribunal, law court, tribunal.
NATURE PROTECTION concept comprises such English expressions as:
conservancy, conservation of nature, to conserve nature, to conserve natural environment, defense of nature, maintenance of nature, nature conservation, nature conservative, to protect nature, protection of nature
and others.
6.3 Concept Name
To work with concepts, to analyse the results of automatic text processing, each concept should have a clear, univocal and concise name. From this point of view, the synonymic sets are not very convenient to use as concept names. Moreover, a synset can consist only of a single ambiguous word and needs additional explication.
Therefore, in the RuThes ontology, each concept has an assigned name.
Name of a concept can be:
· one of unambiguous synonyms;
· an unambiguous multiword expression;
· a pair of synonyms that uniquely identifies the concept;
· an ambiguous word with a relator similar to those used in traditional information retrieval thesauri.
If necessary, a concept may have a comment, which is not a part of the concept name. This is the usual practice in the development of traditional information retrieval thesauri.
6.4 Closely Related Senses in RuThes
The problem of similar meanings of ambiguous words, which may be very hard for compilers of explanatory dictionaries, often become even more complicated for developers of computational linguistic resources. It is often assumed that the integration of similar senses would help reducing the complexity of this problem in computational vocabularies.
However, such clustering may cause other problems.
First of all, as it was already noted, the integration of similar senses may be different depending on an intended application, e.g., machine translation, information extraction, or information retrieval.
Secondly, if we look at the example with word window, the problem of confusion of the name of the concept and the actual concept itself reappears. For example, Window (opening) and Window (artifact) are distinctly different entities, which arose and existed for some time independent of each other. Integration of different entities because of their similar names presents an example of such confusion.
As we have stated in Section 4.2 the integration of two different entities results in confusion of their relations with other concepts, which eventually may affect the logical inference.
Fig.2 Set of concepts corresponding to ambiguous words window and door
Finally, each of these independent entities can be expressed precisely and unambiguously: Window (Opening) can be expressed as window opening (351,000 pages in Google), a Window (artifact) can be expressed as window pane (697,000 pages in Google). As a result of integration of the initial concepts, window opening has become a synonym of window pane.
Thus, in our opinion, if there are distinctly different entities with their own sets of relations and text entries, then they need to be represented by different concepts, even if these individual entities are closely related, and there is a word that can refer to both entities. However, since such entities as Window (opening) and Window (artifact) are closely related, there should be a relation between respective concepts.
Note that these relations should not be relations of metaphor, metonymy, homonymy, as discussed in [26] (see Section 4.1), because these are linguistic relations between the characters, not between the concepts.
In RuThes, one usually employs the part-whole and external dependence [33, 34, 35] relations to describe the relation between closely related entities, which may be named by same ambiguous words. In particular, the concept Window (artifact) is externally dependent on concept Window (opening), since to define concept Window (artifact) one should have concept Window (opening) already defined (Fig.2).
The existence of relations between the related entities simplifies disambiguation of words like window, since it is possible to give a default meaning and to choose it in complex cases.
6.5 Near-Synonyms in RuThes
To describe a set of related meanings of near-synonyms through a set of concepts of a linguistic ontology, the following procedure is applied in RuThes.
The first step is to identify components of the meaning that either always (regardless of the context of use) exist for at least one word of a near-synonyms set, or may occur in certain contexts for several words of a set. In a set of words that are close to the word similarity (see respective WordNet synsets in Section 5.1), this element of the meaning is, for example, the similarity of the external characteristics:
Likeness, alikeness, similitude – (similarity in appearance or character or nature between persons or things)
Resemblance – (similarity in appearance or external or superficial details).
The notation suggests that ‘similarity in appearance’ meaning is significant for people, and this fact should be reflected in the respective concept.
The second step is to find a suitable name for such a concept. In the case of near-synonyms to the word similarity, the name of such a concept could be Similarity in appearance (34,700 pages in the Google). The concept is introduced in the thesaurus with the chosen name.
The next step is to find different ways of expressing the same concept in the form of phrases and single words, e.g., resemblance in appearance, similarity of appearance, external resemblance, etc. All these variants are added as text entries to the concept description.
To reflect the meaning of the words that often express this concept in particular, but can also be used to express the similarity in whole, e.g., resemblance, this word is referred to as a text entry to the concept Similarity in appearance, and as a text entry to the general concept Similarity.
Fig.3 Formation of distinguishable concepts for near-synonyms of word similarity
Fig. 3 presents the resultant set of concepts and their text entries. Such words as resemblance and likeness are described as text entries for two different concepts because of their vague meaning. Thus, we are trying to create context-independent, distinguishable concepts, which are as close as possible to the linguistic meanings.
The analysis of word meanings is similar to feature-based analysis that is often accomplished with Formal Concept Analysis methods [38]. But we try to introduce additional taxonomic categories on the basis of existing language expressions.
In the analysis above it was not necessary to use representations of similar meanings in other natural languages. However, taking into account the examples from another language may be very helpful for recognition of poorly distinguishable concepts.
7 Recommendations for Developers of Formal Ontologies
As we argue that concepts of an ontology cannot be fully separated from natural language meanings we can formulate several recommendations that follow from our consideration of concept-meaning interactions in linguistic ontologies.
In applying Formal Concept Analysis to ontology development one encounters the problem of huge number of taxonomic categories (formal concepts) automatically obtained from a formal context [39, 40]. An additional condition of choosing the most important concepts can be based on existence of words or multiword expressions with corresponding meanings. For example, in the lattice from [39] (Fig. 4), at least four of six intermediate concepts can be named using existing phrases: brain signal (39), MOUSE BRAIN (38), SIGNAL PATHWAY (77), VERTEBRATE BRAIN (36).
Fig.4 Some of intermediate concepts in the example from [38 ] have linguistic names
Developers of an ontology can deal with related entities named by same ambiguous word and, therefore, it can be difficult to distinguish these entities. In these case it can be helpful to find multiword expressions including the same word (or derivative words) and synonymous to this ambiguous word, which can clarify these meanings. We described this procedure for word window in Section 6.4 and used such expressions as window pane and window opening.
Another example is ambiguous word “nation”. This word can denote both political nation, which means ‘state, country’, and ethnic nation, which means ‘ethnicity’. So we can see different entities behind this word and the procedure does not require professional linguistic competence.
At last if an entity seems to have unstable attributes which appear or disappear in different contexts it is possible that the language ambiguity takes place. In such cases it is helpful to use more stable concepts fixing the attributes. We used this procedure in Section 6.5 introducing concept SIMILARITY IN APPEARANCE. And again we use a multiword expression to reveal and fix this concept.
Conclusion
Ontology developers can hardly avoid the influence of linguistic meanings, linguistic polysemy, since the names of concepts and relations in ontologies have mnemonic names, the knowledge in many subject fields is hidden in texts.
Therefore, it is important for ontology developers to understand the problems related to the formation of concepts on the basis of linguistic meanings, namely:
· The problem of distinguishing the concept and its name.
· The problem of presenting closely related meanings of ambiguous words.
· The problem of splitting meanings of near-synonyms into concepts.
In developing the RuThes as a linguistic ontology we are trying to adhere to two, generally speaking, contradictory criteria.
On the one hand, we form concepts of the thesaurus as close as possible to the meanings of linguistic units. As practice has shown, the excessive generalization and clustering of meanings necessarily lead to distortions in the system of relations, to problems in a specific subject field, or an application.
On the other hand, we try to ensure that a concept is still a concept, i.e. it is, at least, distinguishable from the superordinate concept and the sibling concepts.
Exploitation of really existing multiword expressions helps us mitigate these contradictory requirements. The introduction of concepts on the basis of multiword expressions does not only change the essence of a linguistic ontology, but also makes the distinction between the concepts much clearer.
For the concept of ontology, which is clearly distinguishable from other concepts, it is much easier to find equivalents in another language in the form of single words or multiword expressions. So distinguishable concepts do a linguistic ontology more language-independent.
Taking into account the examples from other language(s) is very useful for recognition of poorly distinguishable concepts.
References
1. Staab S., Studer R.: Handbook on Ontologies. Birkhäuser, (2004)
2. Nirenburg, S., Wilks, Y.: What’s in a symbol: Ontology, representation, and language. J. of Experimental and Theoretical Artificial Intelligence. 13(1), 9--23 (2001)
3. Nirenburg, S., Raskin, V.: Ontological Semantics. MIT Press (2004)
4. Lenat, D., Miller, G., Yokoi, T.: CYC, WordNet, and EDR: critiques and responses. Communications of the ACM. 38 (11), 45 --48 (1995)
5. Wilks, Y.: The Semantic Web as the apotheosis of annotation, but what are its semantics? IEEE Intelligent Systems. 23(3), pp. 41--49 (2008)
6. Brewster, Ch., Iria, Jose., Ciravegna, F., and Wilks, Y.: The Ontology: Chimaera or Pegasus. In: Dagstuhl Seminar Machine Learning for the Semantic Web, 89--101 (2005)
7. Tsujii, J., Ananiadou, S.: Thesaurus or logical ontology, which one do we need for text mining? In: Language Resources and Evaluation, Springer Science and Business Media B.V., vol. 39, no 1, pp. 77--90 (2005)
8. Gene Ontology, http://www.geneontology.org/
9. Magnini, B., Speranza, M.: Merging Global and Specialized Linguistic Ontologies. In: Proceedings of OntoLex 2002, pp. 43--48 (2002)
10. Gomez-Perez, A., Fernandez-Lopez, M., Corcho, O.: OntoWeb. Technical Roadmap. D.1.1.2. - IST project IST-2000-29243 (2001)
11. Sowa, J.: Building, Sharing and Merging Ontologies, http://www.jfsowa.com/ontology/ontoshar.htm
12. Miller, G., Beckwith, R., Fellbaum, C., Gross, D., and Miller, K.: Five papers on WordNet. CSL Report 43, Cognitive Science Laboratory, Princeton University (1990)
13. Loukachevitch, N., Dobrov, B.V.: Development and Use of Thesaurus of Russian Language RuThes. In: Proceedings of workshop on WordNet Structures and Standartisation, and How These Affect WordNet Applications and Evaluation. Gran Canaria, Spain, 65--70 (2002)
14. Noy, N.F., McGuinness, D.: (2001) Ontology Development 101: A Guide to Creating Your First Ontology. Stanford Knowledge Systems Laboratory Technical Report KSL-01-05 and Stanford Medical Informatics Technical Report SMI-2001-0880 (2001)
15. Bouaud, J., Bachimont, B., Charlet, J., Zweigenbaum, P.: Methodological principles for structuring an “ontology”. In: Proceedings of IJCAI-95 Workshop “Basic Ontological Issues in Knowledge Sharing” (1995)
16. Miller, G., Hristea, F.: WordNet Nouns: Classes and Instances.– J. Computational linguistics, 32, Number 1, 1-3 (2006)
17. Guarino, N.: Some Ontological Principles for Designing Upper Level Lexical Resources. In: Proceedings of First International Conference on Language Resources and Evaluation. Granada, Spain, pp. 28--30 (1998)
18. Climent, S., Rodriguez, H., Gonzalo, J.: Definitions of the links and subsets for nouns of the EuroWordNet project. Deliverable D005, WP3.1, EurоWordNet, LE2-4003 (1996)
19. Welty, C., McGuinness, D., Uschold, M., Gruninger, M., and Lehmann, F.: Ontologies: Expert Systems all over again. In: AAAI-1999 Invited Panel Presentation (1999)
20. ISO 2788-1986 - Guidelines for the establishment and development of monolingual thesauri. (1986)
21. Z39.19 – Guidelines for the Construction, Format and Management of Monolingual Thesauri. NISO (2003)
22. Soergel, D., Lauser, B., Liang, A., Fisseha, F., Keizer, J., Katz, S.: Reengineering Thesauri for New Applications: the AGROVOC Example. J. of Digital Information, Article No. 257 (2004)
23. Chugur, I., Gonzalo, J., Verdejo, F.: Polysemy and sense proximity in the Senseval-2 Test Suite. In: Proceedings of the ACL-2002 Workshop on “Word sense Disambigation:recent successes”, pp. 32--39 (2002)
24. Agirre, E., Lacalle, L.O.: Clustering Wordnet word senses. In: Proceedings of RANLP 2003 (2003)
25. McCarthy, D.: Relating WordNet Senses for Word Sense Disambiguation. In: Proceedings of NACCL Wordkshop on Making Senses of Sense, pp. 17--24 (2006)
26. Gonzalo, J.: Sense Proximity versus Sense Relations. In: Proceedings of International Wordnet Conference (GWC – 2004), pp. 5--6 (2004)
27. Fellbaum, Ch., Miller, G.: Whither WordNet. Presentation of winners of Antonio Zampolli Award. LREC-2006. (http://www.lrec-conf.org/lrec2006/article.php3?id_article=45) (2006)
28. Landes, S., Leacock, C., and Tengi, R.I.: Building semantic concordances. In: Fellbaum, C. (ed.) WordNet: An Electronic Lexical Database. Cambridge (Mass.): The MIT Press (1998)
29. . Hovy, E.: Methodologies for the Reliable Construction of Ontological Knowledge. In: F.Dau, M.-L. Mugnier, G. Stumme (eds.). Conceptual Structures for Sharing Knowledge. Proceedings of 13 annual conference «Conceptual Structures (ICCS 2005). LNAI, vol. 3596, pp. 91--106. Springer (2005)
30. Hirst, G.: Ontology and the Lexicon.. In: Staab S., Studer R. (eds.) Handbook on Ontologies in Information Systems. pp.209-230. Springer, Berlin (2003)
31. Edmonds, Ph., and Hirst, G.: Reconciling fine grained lexical knowledge and coarse-grained ontologies in representation of near-synonyms. In: Proceedings of Workshop on Semantic Approximation, Granularity and Vagueness. pp. 12-22 (2000)
32. LIV (Legislative Indexing Vocabulary). Congressional Research Service. The Library of Congress. Twenty-first Edition (1994)
33. Guarino, N., Welty, C.: A Formal Ontology of Properties. In: Knowledge Engineering and Knowledge Management: Methods, Models and Tools. 12th International Conference, EKAW 2000. LNCS, vol.1937, pp. 97--112 (2000)
34. Masolo, C., Vieu, L., Bottazzi, E. Catenacci, C., Ferrario, R., Gangemi, A., Guarino, N.: Social roles and their descriptions In: Dubois, D., Welty, C., Williams, M.A. (eds.) Principles of Knowledge Representation and Reasoning, Proceedings of the Ninth International conference KR 2004. pp. 267--277. AAAI Press (2004)
35. Loukachevitch, N.V., Dobrov, B.V., Development of Ontologies with Minimal Set of Conceptual Relations. In: Lino, M. T., Xavier, M. F., Ferreira, F. et al. (eds.) Proceedings of Fourth International Conference on Language Resources and Evaluation. vol. 6, pp.1889--1892 (2004)
36. Loukachevitch, N.V., Dobrov, B.V.: Ontological Types of Association Relations in Information Retrieval Thesauri and Automatic Query Expansion. In: Oltramari et.al. (eds) Proceedings of OntoLex 2004: Ontologies and Lexical Resources in Distributed Environments. pp. 24--29 (2004)
37. Loukachevitch, N., Dobrov, B.: Sociopolitical Domain as a Bridge from General Words to Terms of Specific Domains. In: Proceedings of Second International WordNet Conference GWC. pp.163--168 (2004)
38. Priss, U.: Linguistic Applications of Formal Concept Analysis." In: Ganter; Stumme; Wille (eds.), Formal Concept Analysis, Foundations and Applications. LHAI, vol. 3626, pp. 149--160. Springer (2005)
39. Kuznetsov, S., Obiedkov, S., Roth C.: Reducing the Representation Complexity of Lattice-Based Taxonomies, in Proc. 15th International Conference on Conceptual Structures (ICCS'07), U. Priss, S. Polovina, R. Hill (Eds.), LNAI, vol. 4604, pp. 241--254. Springer (2007)
40.Cimiano, P., Stumme, G., Hotho, A., Tane, J.: Conceptual Knowledge Processing with Formal Concept Analysis and Ontologies. In: Eklund P.W. (ed.) Proceedings of the The Second International Conference on Formal Concept Analysis (ICFCA 04). LNCS, vol. 2961, pp.189--207. Springer (2004)