The paper represents an outline of a new technique aimed at improvement of lexicon-to-ontology mapping technology. The technique is integrated within the Ontolex-Lemon lexicon model by supplying a lexical concept of an ontology class with a description logic based formal definition. A natural language definition used to describe the lexical concept is transformed into a DL-definition; the resulting DL-definition is associated with a graph-like join of domain ontology properties. As a result, a related lexical unit is mapped to several bound ontology units rather than to a single ontology class. DL-based and graph-based lexical meaning formal representations are applied for lexical sense disambiguation within a lexicon and for extension of class and property taxonomies of an associated ontology.
Due to extensive growth of amount and variety of World Wide Web content, the task
of relevant information retrieval is becoming more and more challenging. Traditional
keyword-based search engines show high recall but low precision since the search
technology does not provide any formal account of semantics of keywords [
Ontologies are supposed to be the key to an advanced search technology, providing
formal specifications of vocabulary units used to represent a domain [
Numerous ontology-based search engines have been developed during the recent
years in order to improve the precision of traditional keyword-based search: MIRO
[
Within the process of ontology-based semantic search, users’ queries and Web
documents acquire ontologically motivated semantic representations [
The major bottleneck in ontology-based semantic search development lies in
designing an efficient technology of semantic analysis of users’ queries, which provides
accurate matching of lexical units with units of an ontology. The vast variety of lexical
means of expression represented by natural languages along with highly developed
homonymy, synonymy, and polysemy within lexical systems result in the exemplified
use of ontology-based search engines being limited to a particular domain, for example,
a soccer domain [
The rest of the paper is organized as follows: in sections 2 and 3 we discuss lexical analysis implementation in operation of natural language interfaces to ontologies and lexicon models applied during the analysis. Section 4 is devoted to basic drawbacks of the current lexicon-to-ontology mapping technology. Section 5 introduces our approach towards associating lexicon units with units of an ontology. Section 6 represents the results of experimental implementation of the novel technique. Section 7 provides some brief conclusions as well as a vision of perspectives and future work. 1 http://m.wolframalpha.com 2 http://www.kngine.com 3 https://www.google.com
Natural language interfaces (NLI) are developed to conduct semantic analysis and to
provide a formal representation of a query’s semantics [
NLIs typically involve an extensible lexicon, which is at least partly generated
automatically from an ontology, in the process of semantic analysis of a query. The
involvement of linguistic data on lexemes provided by a lexicon in the process of
semantic analysis of a query differs from one NLI system to another. The NLI of
FREyA [
The NLIs of ORAKEL [
The ORAKEL parser [
Elementary trees are combined by the parser to produce a parse tree. The head nodes of the tree representing nouns, verbs, and adjectives are provided semantics by virtue of lambda expressions describing an unary or a binary predicate or a constant which corresponds to a unit of an underlying ontology. The Query Interpreter of the ORAKEL system combines semantic representations of tree nodes to produce a formal representation of a query’s semantics. The formal representation is provided in the form of a FOL-like formula augmented with query, count, and arithmetic operators, which is transformed into a SPARQL query. 2.2
The NLI of Pythia [
Elementary LTAG trees representing lexical units of a query are used by the inbuilt parser to obtain an LTAG-derivation tree, and entries of a lexicon provide enough data to build a parse tree by implementing substitution and adjoin operations. DUDEs are merged to form a Discourse Representation Structure (DRS), the obtained DRS of a user’s query is converted into a SPARQL query. 3
LexInfo [
Both lexicon models are intended chiefly to provide for a domain-specific ontology lexicon. The core difference between LexInfo and LexOnto lies in the mode of lexiconto-ontology units mapping. In LexOnto subcategorization frames introducing arguments attached by verbs and relational nouns are associated with single or joined properties of a domain ontology. In LexInfo verbs and relational nouns are mapped to object and datatype properties in a straightforward fashion, whereas their syntactic behavior is specified by virtue of specific subcategorization frames with arguments being mapped to domains and ranges of the properties associated with the lexemes. In both frameworks all lexemes acquire lemmas and a list of form variants distinguished by mood, gender, number, case, degree, person, etc.
LexInfo and LexOnto specify syntactic behavior of a verb or a relational noun through a verbal or a nominal subcategorization frame accordingly. A verbal subcategorization frame encodes number and sort of a verb’s arguments: a subject, a direct object, and optionally one or two prepositional complements for a transitive verb; a subject and a prepositional complement for an intransitive verb. A nominal subcategorization frame includes one or two prepositional complements, and an argument position titled as external subject is filled whenever a relational noun is used in the role of a predicate, which is related to a subject by virtue of copula. A subject is included into nominal subcategorization frames within the LexInfo system, but it is stored separately by LexOnto. Apart from that, LexOnto provides subcategorization frames for participles attaching a prepositional complement. A binary frame is mapped either to a single property or to a 2 x 2-Join of properties; a ternary frame is mapped either to a 3 x 2-Join or to a 2 x 2-Join’ of properties whenever a joined position is mapped to an argument of the frame; a quaternary frame is associated with a 4 x 2-Join or with a 3 x 2-Join’ of ontology properties.
Adjectives used as adjectival modifiers of a noun are supposed to subcategorize for a modified noun. An adjective and a modified noun are mapped to ontology classes undergoing intersection whenever the adjective is associated with a particular class entering a class taxonomy of an ontology. Adjectives of this kind are referred to as intersective or class adjectives and are successfully handled by LexInfo. Subjective adjectives, on the other hand, do not have fixed extensions on a domain and their interpretation is context dependent. An adjective of this kind is mapped to a property of an ontology with special constraints being imposed on the property’s range, whereas a modified noun is associated with a subclass of the domain of the property.
A lexicon-to-ontology mapping scheme depends on a kind of a property which is associated with an adjective. LexInfo maps literal adjectives such as blue or skillful to object properties. Evidently, an ontology-based interpretation of the adjective skillful has to be different if it subcategorizes for the noun gardener or for the noun surgeon. Yet, the value constrains that have to be imposed on a domain of an object property a literal adjective is mapped to are not proposed.
Within the frameworks of LexInfo and LexOnto, a scalar adjective like long or big is mapped to a datatype property on occasion particular constraints are imposed on the data value range. Positive or negative polarity has to be set for this data value range to indicate if the value has to increase or to decline to give an appropriate formal account of comparative and superlative forms of the adjective. One should bear in mind that the constraints might be regionally or culturally specific and take into account the units of measurement. LexInfo also proposes value constraints imposed on a datatype property domain to model a scalar adjective’s semantics since data value constraints are supposed to be different, for instance, when the noun man or the noun woman is modified by the adjective tall. 4
Within the framework of Natural Language Interfaces to ontologies, lexemes acquire semantics with reference to ontology units. Whenever an NLI integrates a lexicon model instantiated by LexInfo and LexOnto, word sense disambiguation is done by virtue of semantic analysis of a target lexeme’s syntactic behavior. On occasion the syntactically bound lexemes of a target lexeme respect the ontological restrictions imposed through lexicalization of an ontology unit, the target lexeme is supposed to refer to the ontology unit. These restrictions expect lexical fillings of slots in lexemes’ subcategorization frames to be coreferential on a domain with particular class names or data value ranges that are predefined by one-to-one correspondence between syntactic role slots and domain/range constituting classes/data value ranges. This correspondence is set within the process of lexicalization and is scripted by virtue of elementary trees with ontological restrictions being imposed on particular nodes that acquire syntactic roles defined in a subcategorization frame.
The technology of meaning acquisition and disambiguation provided by lexicon-toontology mapping frameworks poses a high demand on taxonomy organization and entity coverage of a lexicon. In other words, all classes and properties that can possibly be lexicalized by a user in a query should be included in taxonomies of a domain ontology, and all possible contexts of lexemes’ use should be taken into account in the process of ontology units’ lexicalization. These demands appear to be virtually unfeasible, yet, their necessity is easily illustrated by the following examples of problematic cases of lexicon-to-ontology mapping.
An intransitive verb pass, which subcategorizes for a subject and a prepositional
complement attached by the preposition through, is provided as an instance of an
ambiguous verb by ORAKEL developers [
Ontological restrictions imposed on a verb’s arguments by a join of associated
properties should also be subjected to thorough reification. For instance, an alternative
OntoSem lexicon model [
The authors of LexOnto [
In order to resolve the issues of semantic ambiguity and inaccuracy of ontology-based
lexical semantics representations, we propose to enhance the technology of
lexicon-toontology units mapping by introducing description logic based formal definitions of
ontology class representing lexemes in the scope of semantic data provided by an
ontology lexicon. These definitions are intended to provide a formal account of a
lexeme’s meaning. In the current research the notion of lexical meaning is equated to
intension, which is understood as ‘a function from a set of possible worlds to a set of
all subsets of homogeneous n-ary relations on a domain: : → 2 ’ [
Resulting DL-definitions are presumed to be introduced by virtue of the
OntoLexLemon model developed to represent data on ontology units’ lexicalizations [
Figure 1 provides an instance of a lexical entry representing the noun coach, which evokes two lexical concepts defined by Open Multilingual Wordnet 2.0 (OMW)5: private instructor and manager. Private instructor, a concept of the DBpedia6 class Coach, is associated with the first sense of the lexeme coach, which could be defined as a person who gives private instruction. Manager, a concept of the DBpedia class 4 Please note that the set of transformation rules has been augmented with the solutions for formalization of scalar and literal adjectives proposed within LexOnto and LexInfo systems (see Section 3). 5 http://compling.hss.ntu.edu.sg/iliomw/omw 6 http://mappings.dbpedia.org/server/ontology SportsManager, is associated with the second sense of the lexeme coach, which could be defined as someone in charge of training an athlete or a team. Yet, the class Coach might be related to both concepts, which results in semantic ambiguity for the second sense of the lexeme coach. The lexical sense acquires reference to the ontology classes Coach and SportsManager despite the fact that reference is a functional property.
In order to overcome the ambiguity, the NL-definitions used to characterize corresponding lexical concepts have been formalized by virtue of NL-DL definition transformation. Concepts and roles forming complex descriptions in DL-definitions have been associated with units of the DBpedia ontology. As a result, the lexeme coach has acquired meaning by being mapped to joins of ontology units rather than to a single class of the ontology. The joins form graph structures with properties corresponding to edges, classes and data value ranges corresponding to vertices, therefore the joins are referred to as graph-definitions representing lexical meanings of associated lexemes. PrivateInstructor ≡Person⊓∃give. (Private⊓Instruction)
Manager ≡Person ⊓∃in_charge_of.(∃training.
(Athlete⊔Team))
A DL-definition is a terminological axiom stating concept equivalence, which binds an
atomic concept with a complex description obtained by means of specific concept
constructors. In the current research the constructors presumed by SROIQ(D) syntax
[
Atomic concepts, roles, or their combinations introduced in different joints of a chain of intersections are associated with subclasses of domains and ranges of object and datatype properties of an ontology. The ontology properties form a graph-definition associated with a lexical concept of an ontology class. DL-based and graph-based definitions related to the OMW lexical concept archeologist are given in the Figure 2 for the purpose of illustration. All lexical senses associated with the lexical concept archeologist are supposed to refer to the interconnected properties forming a graph rather than to the DBpedia class Archeologist. The resulting graph-definition specifies the ontological restrictions for syntactic arguments of the verbs associated with the following DBpedia object properties: activity (Person, Thing), focus (Thing, Thing), related (Thing, Thing), era (Thing, Thing). The constraints are supposed to be applied within discourse on human occupations.
This kind of specification appears to be the key to resolution of problematic cases of mapping verbs to object properties of an ontology, which are exemplified in the Section 4. Let us presume that DBpedia contains the object properties: pass_through (Thing, Place), establish (Agent, Thing), and hasBeneficiary (SpeechAct, Thing). In order to specify an ontological restriction for a lexeme associated with the domain of the object property pass_through (Thing, Place), one should produce a DL-definition bound with the OMW lexical concept itinerary associated with the domain of the property. The DL-definition: ≡ ∃ ℎ . ( ⊓ ∃ _ . ( ⊔ )), (1) acquires the following graph-based interpretation within the framework of DBpedia: establish (Agent, RouteofTransportation); pass_through (RouteofTransportation, Place). With the ontological restriction being specified as pass_through (RouteofTransportation, Place), the lexical units associated with the domain of the object property pass_through (Thing, Place) are supposed to be coreferential with subclasses of the DBpedia class RouteofTransportation, i.e. with the classes: Bridge, RailwayLine, RailwayTunnel, Road, RoadJunction, RoadTunnel, WaterwayTunnel.
In order to impose an ontological restriction on a verb argument which is associated with the range of the object property hasBeneficiary (SpeechAct, Thing), a DLdefinition associated with the OMW lexical concept addressee was mapped to the object property hasBeneficiary (SpeechAct, Agent). As a result, the DBpedia class Agent was set as an ontological restriction for a syntactic argument associated with the range of the property hasBeneficiary (SpeechAct, Thing).
The proposed attitude might be extended to improve semantic representations of adjectives. With a view to define the properties an adjective is mapped to, one has to produce a DL-definition associated with a lexical concept evoked by a modified noun. For instance, whenever the adjective big modifies the noun city, the adjective should be associated with the DBpedia datatype properties area (City, Area) and populationDensity (City, PopulationDensity). The reason is that these datatype properties compose a graph-definition bound with the OMW lexical concept urban center, which is evoked by the noun city.
In order to provide a brief description of advantages and shortcomings of the proposed approach to lexicon-to-ontology mapping, we have retrieved 50 lexical concepts from Open Multilingual Wordnet 2.0, the concepts that should be associated with subclasses of the DBpedia class Person. The NL-definitions used to characterize intensions of the lexemes associated with the lexical concepts were transformed into DL-based formal definitions.
All DL-definitions used to characterize 50 retrieved lexical concepts associated with 39 DBpedia classes were successfully linked to graph-definitions. Graphdefinitions were the keys to resolution of 11 cases of semantic ambiguity, which arose every time two or more lexical concepts were defined as concepts of the same ontology class. The list of these classes includes the classes: Coach, Judge, Politician, Referee, and Ambassador among others. The set of graph-definitions provides information on the ontological restrictions that should be imposed on arguments of the lexemes associated with the object properties: profession (Person, Thing), education (Person, Thing), activity (Person, Thing), management (Thing, Thing), specialization (Thing, Thing), created (Person, Work). For instance, within discourse on human occupations the arguments associated with the domain of the object property management (Thing, Thing) should be restricted to the ones mapped to subclasses of the DBpedia classes Person and WrittenWork. The arguments mapped to the range should be restricted to the ones associated with subclasses of the DBpedia classes: Person, Activity, Place, and MeanOfTransportation. The arguments associated with the domain of the object property specialization (Thing, Thing) should be restricted to the ones mapped to subclasses of the DBpedia class Person, whereas the lexical units associated with the range of the property are supposed to be coreferential with subclasses of the class Science.
In 78% of cases newly proposed classes and properties had to be used in order to compile suitable graph-based definitions associated with particular lexical concepts. Overall, 43 classes and 9 object properties have been proposed to enhance the class and property taxonomies of DBpedia. For instance, the class Anthropologist and the class Culture have been proposed in order to develop the graph-definition represented in the Figure 2. Some of the proposed properties appear to be crucial for formal description of human occupation subdomain: instruct (Person, Person), solve (Thing, Thing), represent (Thing, Thing), find (Thing, Thing). Two object properties that have been proposed: appoint (Agent, Person) and appointed (Person, Thing), form a 2 x 2 Join’ which is supposed to be mapped to the subcategorization frame of the verb to appoint containing three arguments. 13,5 % of the proposed ontology units were used more than once for the purpose of graph-based representation of lexical concepts of human occupations. 6% of proposed units subsume at least one suggested unit. For instance, the introduced class Science possesses 6 subclasses that were suggested to represent different areas of research: Economics, Egyptology, Linguistics, Philosophy, Psychology, History. Consequently, the extension of an ontology’s taxonomies in the process of graph-definitions formation should be considered domain oriented and therefore fruitful.
Simultaneously, a soft spot of the proposed technique arises. Units composing DLbased and graph-based definitions acquire an irregular match that might complicate the technique implementation in lexicon modelling (see Table 1). For instance, the range of the object property education (Person, Thing) is associated with the atomic concept Psychology within the DL-definition related to the lexical concept psychologist. The same object property gets its range mapped to the atomic role travel_in, which is in turn associated with the domain of the property equipment (Activity, Thing) used within the graph-definition related to the lexical concept spaceman. Finally, the range of the object property education (Person, Thing) happens to be related to the complex description ∃compete_in.Sports, which in turn is used to define an existential restriction imposed on the atomic role trained_to, within the DL-definition related to the lexical concept athlete. The range of the object property created (Person, Work) is associated with an intersection of the concepts Creative and Work within the DL-definition related to the lexical concept artist. Within the same DL-definition the range of the object property picture (Thing, Thing) is mapped to the atomic concepts Sensitivity and Imagination that compose an intersection used to characterize an existential restriction imposed on the atomic role show. The case of the object property activity (Person, Thing) being mapped to the atomic concept Scientist within the DL-definition associated with the lexical concept psychologist should also be taken into consideration.
These examples reveal the necessity of making ad-hoc decisions in order to link units of DL-based definitions of lexicon units’ semantics with units of an associated ontology. Hence, even though the proposed technique seems to be an appropriate tool for revision and improvement of lexicon-to-ontology mappings conducted in relevance to a particular domain of discourse, the large-scale implementation of the technique is yet to be achieved. ≡ ≡ ⊓ ∃
⊓ ∃ . ( ⊓ _ . (∃ ⊓ ∃ . (( ⊓
_ . _ .
ℎ _ . (∃ _ . ⊓
) )) ) ) graph-definition created (Person, ArtWork); picture (ArtWork, Sensitivity); picture (ArtWork, Imagination) education (Person, Contest) activity (Person, Science); education (Person, Psychology) education (Person, Management); equipment (Management, Spacecraft) The brief study introduced in the current paper has shown that DL-based and graphbased formal specifications of a lexeme’s meaning improve the accuracy of lexicon-toontology mappings by resolving cases of semantic ambiguity among ontology class representing lexemes. At the same time, the semantics of property representing lexemes is subjected to reification through specification of ontological restrictions imposed on arguments entering the lexemes’ subcategorization frames.
The novel technique of ontology lexicon modelling is applicable under the condition of an ontology’s taxonomies being limited and allows to deny the impracticable demand for summarization of all possible grammatical contexts of a lexeme’s use. Meanwhile, the process of graph-definitions formation stimulates the development of class and property taxonomies of a domain ontology. However, a regular correspondence between units composing DL-based and graph-based definitions is yet to be found. A set of rules associating atomic roles, atomic concepts, and complex descriptions of a DL-definition with classes and data value ranges representing domains and ranges of ontology properties has to be introduced in order to make the technique applicable in a large-scale fashion.