Procesamiento del Lenguaje Natural ISSN: 1135-5948 [email protected] Sociedad Española para el Procesamiento del Lenguaje Natural España
Periñán-Pascual, Carlos; Arcas-Túnez, Francisco Ontological commitments in FunGramKB Procesamiento del Lenguaje Natural, núm. 44, marzo, 2010, pp. 27-34 Sociedad Española para el Procesamiento del Lenguaje Natural Jaén, España
Available in: http://www.redalyc.org/articulo.oa?id=515751744006
How to cite Complete issue Scientific Information System More information about this article Network of Scientific Journals from Latin America, the Caribbean, Spain and Portugal Journal's homepage in redalyc.org Non-profit academic project, developed under the open access initiative Procesamiento del Lenguaje Natural, Revista nº 44, marzo de 2010, pp 27-34 recibido 08-01-10 revisado 21-02-10 aceptado 05-03-10
Ontological commitments in FunGramKB
Criterios ontológicos en FunGramKB
Carlos Periñán-Pascual Francisco Arcas-Túnez Universidad Católica San Antonio Universidad Católica San Antonio Departamento de Idiomas Departamento de Informática de Sistemas Campus de los Jerónimos s/n Campus de los Jerónimos s/n 30107 Guadalupe - Murcia (Spain) 30107 Guadalupe - Murcia (Spain) [email protected] [email protected]
Resumen: Cualquier trabajo en ingeniería ontológica debe estar fundamentado en un protocolo de directrices bien definidas que no sólo organicen la estructuración de la ontología sino que además ayuden a determinar sus unidades ontológicas y propiedades. Una sólida metodología para el desarrollo de ontologías exige la eliminación de muchos de los errores e inconsistencias que se suelen cometer en el modelado ontológico, facilitando así la interoperatibilidad y el conocimiento compartido—especialmente útil cuando se diseña un recurso multipropósito. En el contexto del procesamiento del lenguaje natural, este artículo describe los compromisos ontológicos que la Ontología de FunGramKB debe cumplir. Palabras clave: ontología, FunGramKB, herencia, postulado de significado
Abstract: Ontology engineering should be grounded on a protocol of well-founded guidelines concerning the structuring of the ontology as well as the elements to be included and their ontological properties. A sound methodology for ontology development involves a dramatic reduction of many common errors and inconsistencies in conceptual modelling, facilitating thus interoperatibility and knowledge sharing—particularly useful when a multipurpose resource is designed. In the natural language processing context, this paper describes the ontological commitments to which the FunGramKB Ontology is subject. Keywords: ontology, FunGramKB, inheritance, meaning postulate
incorporated as a conceptual unit, where to 1 Introduction place it, how to represent its meaning and how to organize the structure of the whole ontology A key aspect in knowledge engineering is the (Mahesh, 1996). design and construction of an ontology model under a series of well-founded guidelines, 2 FunGramKB particularly when you want to reuse it in different natural language processing (NLP) FunGramKB (Periñán-Pascual and Arcas- applications, e.g. document retrieval, Túnez, 2004, 2005) is a multipurpose lexico- information extraction, text categorization, data conceptual knowledge base for NLP systems. mining, etc. Ontology structuring must be On the one hand, FunGramKB is multipurpose supported by some theory about the elements in in the sense that it is both multifunctional and the domain, their inherent properties and in multilingual. Thus, FunGramKB has been which way these elements are related. Since designed to be potentially reused in many NLP knowledge engineers face numerous problems tasks and with many natural languages.1 in the conceptual modelling of ontologies, it is necessary to work with some underlying 1 English and Spanish are fully supported in the ontological commitment. These guidelines current version of FunGramKB, although we have should help us to make decisions on what to be just begun to work with other languages, i.e. German, French, Italian, Bulgarian and Catalan.
ISSN 1135-5948 © 2010 Sociedad Española para el Procesamiento del Lenguaje Natural Carlos Periñán Pascual, Francisco Arcas-Túnez
On the other hand, our knowledge base any language input conceptually. In this comprises three major knowledge levels, scenario, the Ontology becomes the pivotal consisting of several independent but module for the whole architecture, which interrelated modules: explains why this model is conceptually rather than lexically-driven. Lexical level: • The Lexicon stores morphosyntactic, 3 Ontological commitments pragmatic and collocational information Nowadays there is no single right methodology about lexical units. for ontology development. Ontology design • The Morphicon helps our system to handle tends to be a creative process, so it is probable cases of inflectional morphology. that two ontologies designed by different people have a different structuring (Noy and Grammatical level: McGuinness, 2001). To avoid this problem, the • The Grammaticon is composed of several ontology model should be founded on a solid Constructicon modules whose methodology. The remaining of this section constructional schemata help Role and describes seven methodological criteria applied Reference Grammar to build the semantics- to the FunGramKB Ontology, some of which syntax-semantics linkage (Van Valin and are based on principles implemented in other LaPolla, 1997; Van Valin, 2005). NLP projects (Bouaud et al., 1995; Mahesh, 1996; Noy and McGuinness, 2001). Conceptual level: • The Ontology is presented as a hierarchical 3.1 Commitment #1: universality and catalogue of the concepts that a person has linguistic motivation in mind when talking about everyday FunGramKB is provided with a universal, situations. Here is where semantic linguistically-motivated and general-purpose knowledge is stored in the form of meaning ontology. postulates. Firstly, the FunGramKB Ontology takes the • The Cognicon stores procedural knowledge form of a universal concept taxonomy, where (e.g. how to fry an egg, how to buy a “universal” means that every concept2 we product, etc) by means of scripts, i.e. imagine has, or can have, an appropriate place conceptual schemata in which a sequence in the Ontology. A universal approach is of stereotypical actions is organised on the adopted on the relation between language and basis of temporal continuity, and more conceptualization, where cross-lingual particularly on the basis of Allen's temporal differences in syntactic constructions do not model (Allen, 1983; Allen and Ferguson, necessarily involve conceptual differences (cf. 1994). Jackendoff, 1990). In our case, the relation • The Onomasticon stores information about between language structures and conceptual instances of entities and events, such as Bill constructs is mediated by conceptual logical Gates, Taj Mahal, or 9/11. This module structures (CLS), where phenomena such as stores two different types of schemata (i.e. diathetic alternations are directly reflected. The snapshots and stories), since instances can role of a CLS is to serve as a bridge between be portrayed synchronically or the more abstract level as represented in the diachronically. Ontology and the particular idiosyncrasies as coded in a given linguistic expression. In the FunGramKB architecture, every Therefore, CLSs are used as the interface lexical or grammatical module is language- between the semantic structure and the syntactic dependent, whereas every conceptual module is representation of sentences. To illustrate, the shared by all languages. In other words, computational linguists must develop one 2 Terms such as “class”, “category” or “semantic Lexicon, one Morphicon and one Grammaticon type” are often used in ontology engineering to refer for English, one Lexicon, one Morphicon and to elements such as FunGramKB “concepts”. one Grammaticon for Spanish and so on, but However, we prefer the latter, since it better knowledge engineers build just one Ontology, describes the domain of processing in the two-tier one Cognicon and one Onomasticon to process model of our NLP knowledge base, i.e. lexical level and conceptual level.
28 Ontological commitments in FunGramKB
CLS of sentence (1) is presented in (2):3 (e.g. #ABSTRACT, #COLLECTION, #EMOTION, #POSSESSION, #TEMPORAL (1) Betty asked Bill for the aspirin. etc), constitute the upper level in the taxonomy. (2)
29 Carlos Periñán Pascual, Francisco Arcas-Túnez
phases. In a first stage, a shared view of the 3.3 Commitment #3: non-atomic modelled world was sought to shape the upper conceptual units level (i.e. metaconceptual level). In a second stage, the most descriptive concepts were In FunGramKB, basic and terminal concepts hierarchically arranged in a middle level (i.e. are not stored as atomic symbols but are basic level). In a third stage, the most specific provided with conceptual properties such as the concepts are incorporated as leaves of the thematic frame and the meaning postulate. Both ontology (i.e. terminal level). Thus, we started of them are conceptual schemata, since they with metaconcepts until terminal concepts were employ concepts—and not words—as the reached (globally top-down), but each ontology building blocks for the formal description of level was developed in a locally bottom-up meaning. Thus, thematic frames as well as fashion. Although bottom-up approaches tend meaning postulates become language- to lead to ad hoc ontologies, this problem is independent conceptual knowledge solved by postulating the high-level ontology representations. beforehand. Currently, the third stage of the Every event and quality in the Ontology is ontology modelling is still being developed. assigned one single thematic frame, i.e. a Indeed, the development of the FunGramKB conceptual construct which states the number Ontology is grounded on a spiral model, where and type of participants involved in the conceptual promotion and depromotion can prototypical cognitive situation portrayed by the occur between the basic and terminal levels. event. To illustrate, we present the thematic Thus, some terminal concepts can be frame of +OPEN_01 (3), to which lexical units such as open [Eng], abrir [Spa], offnen [Ger], “promoted” to basic concepts when the 6 inclusion of a new language demands a slightly ouvrir [Fre] and aprire [Ita] are linked: different view to the world model. On the (3) (x1)Agent (x2: +DOOR_00 ^ contrary, some basic concepts can be +WINDOW_00)Theme (x3)Location “depromoted” to teminal concepts when their (x4)Origin (x5)Goal expectations as meaning descriptors are not finally fulfilled. In any case, the metaconceptual Thematic frames can also include those selectional preferences typically involved in the level always remains stable. Figure 1 illustrates 7 the design of the FunGramKB Ontology. cognitive situation. Thus, thematic frame (1) describes a prototypical cognitive scenario in which “entity1 (Agent) moves entity2 M e (Theme)—being typically a door or window— t P a U c - from one place (Origin) to another (Goal), there P o G M n G L U c -O E O also being a place (Location) along which e N MT p B OT t E A TO u R L entity2 moves”. Although one or more T a I L B l C O Y D B T T subcategorization frames can be assigned to a E O O YZ B L I conceptual P P L a L - - single lexical unit, every concept is provided A s D D A depromotion i O C C c O O WW with just one thematic frame. OL L N N A meaning postulate is a set of one or more T e r logically connected predications (e1, e2... en), m conceptual in i.e. conceptual constructs carrying the generic a promotion l 8 features of concepts. Consider (4-5) as a representation of the thematic frame and
Figure 1: FunGramKB Ontology design. 6 Although FunGramKB knowledge is stored in XML, examples of thematic frames and meaning The motivation of constructing such an postulates shown in this paper are presented by ontology model responds to the need of a core means of those parenthetic string representations level of knowledge (i.e. basic concepts) playing created by language engineers through FunGramKB a pivotal role between those universal Suite (www.fungramkb.com). categories which can facilitate ontology 7 Selectional preferences are stated when they interoperatibility (i.e. metaconcepts) and those can exert some predictive power on the participant. 8 particular concepts which can grant immediate Periñán-Pascual and Arcas-Túnez (2004) applicability (i.e. terminals). describe the formal grammar of well-formed predications for meaning postulates in FunGramKB.
30 Ontological commitments in FunGramKB
meaning postulate of +INGEST_00, to which concept (i.e. specificity principle). Therefore, a lexical units drink [Eng], beber [Spa], trinken new conceptual unit can be placed within the [Ger], boire [Fre] and bere [Ita] are linked: Ontology providing that there is some information which can be expressed about that (4) (x1: +HUMAN_00 ^ concept and was not already stated about the +ANIMAL_00)Agent (x2: superordinate concept. However, it is necessary +LIQUID_00)Theme (x3: to achieve a balance in the ontology design in +THROAT_00)Location (x4)Origin (x5: order to avoid the creation of multiple +STOMACH_00)Goal subordinate concepts on the basis of minor (5) +(e1: +ABSORB_00 (x1)Agent specifications on superordinate concepts. (x2)Theme (x6)Location (x4)Origin (x7: Finally, differentiae in the meaning +MOUTH_00)Goal (f1: x6)Instrument postulates of sibling concepts must be (f2: (e2: +SWALLOW_00 (x1)Agent incompatible one another (i.e. opposition (x2)Theme (x3)Location (x7)Origin principle). In this way, every subordinate (x5)Goal))Purpose) concept has an exclusive value within the The thematic frame of +INGEST_00 depicts cognitive dimension established by the a situation in which five participants are superordinate concept. typically involved: a person or animal (x1) Strictly speaking, these principles are not makes a liquid (x2) move through their throat directly applied to the meaning postulates (x3) from one place (x4) to their stomach (x5). stored in the knowledge base but to the output If the semantic burden of this concept, and of the Microconceptual-Knowledge Spreading consequently of their corresponding words, lay (MicroKnowing), a pre-reasoning process for just on this thematic frame, then we would not the construction of the extended meaning be actually describing the conceptual content of postulate of a concept. The MicroKnowing those lexical units. This is the reason why takes place in a multi-level scenario, since it is meaning postulates were introduced as performed by the iterative application of two notational devices for the representation of types of reasoning mechanisms: inheritance and conceptual meaning in the Ontology. Unlike inference. Whereas inheritance strictly involves some other approaches in NLP, FunGramKB the transfer of one or more predications from a adopts a deep semantic approach which superordinate concept to a subordinate one in strongly emphasizes the commitment to provide the Ontology, our inference mechanism is based meaning definition via meaning postulates. For on the constructs shared between predications example, (6) presents the natural language linked to conceptual units which do not take equivalent of the meaning postulate of part in the same subsumption relation within the +INGEST_00: Ontology. It might be the case that two hierarchically-linked concepts could not be (6) A person or animal absorbs a liquid distinguished by means of their meaning through an instrument to the mouth with postulates; this is feasible providing that their the purpose of swallowing it to the extended meaning postulates are dissimilar, stomach. since different predications are inferred. Indeed, 3.4 Commitment #4: meaning this approach is an advantage of the FunGramKB Ontology model, since meaning postulates as ontology organizers postulates are underspecified in order to avoid The structure of the FunGramKB Ontology unnecessary duplication of information. complies with the similarity, specificity and Therefore, the MicroKnowing plays the role of opposition principles applied to the meaning a pre-reasoner running on the Ontology with the postulates of concepts. aim of minimizing redundancy as well as Firstly, all subordinate concepts must share maximizing informativeness in the semantic the meaning postulate of their superordinate and world knowledge repository of concept (i.e. similarity principle), though non- FunGramKB.9 monotonicity is permitted. Secondly, all subordinate concepts must have a meaning postulate which states a 9 A more accurate account of the resolution of distinctive feature (or differentiae) not present incompatibilities in the MicroKnowing is described in the meaning postulate of its superordinate in Periñán-Pascual and Arcas-Túnez (2005).
31 Carlos Periñán Pascual, Francisco Arcas-Túnez
3.5 Commitment #5: IS-A taxonomic and only one logical relation (&, | or ^) between relation the multiple parents. In other words, FunGramKB gives the possibility to state if a At first sight, it can seem that the exclusive use subordinate concept can refer simultaneously to of the IS-A relation can impoverish the the referents of all its superordinate concepts ontology model. Indeed, a consequence of this (8), or if the conceptual parents are disjuncts restriction on the taxonomic relation is found in (9). the modelling of the upper level into three subontologies, where metaconcepts #ENTITY, (8) +((e1: +BE_00 (x1: +HAIR_01)Theme #EVENT and #QUALITY arrange nouns, verbs (x2: +BODY_PART_00 & and adjectives respectively in cognitive +GROUP_00)Referent) (e2: dimensions. However, the fact that concepts +COMPRISE_00 (x2)Theme (x3: linked to lexical units of different grammatical +HAIR_00)Referent)) categories are not explicitly connected in the (9) +(e1: +BE_00 (x1: +DRUG_00)Theme Ontology doesn’t prevent FunGramKB to relate (x2: +SOLID_00 ^ +LIQUID_00 ^ those lexical units in the conceptual level +GAS_00)Referent) through their meaning postulates. Indeed, the *(e2: n +BE_01 (x1)Theme (x3: Ontology establishes a high degree of +LEGAL_00)Attribute) connectivity among concepts by taking into *(e3: +INGEST_00 (x4)Agent (x1)Theme account conceptual components which are (x5)Location (x6)Origin (x7)Goal (f1: (e4: shared by their meaning postulates. In order to +BE_01 (x4)Theme (x8: +HAPPY_00 | incorporate human beings’ commonsense, the +NERVOUS_00)Attribute))Result) Ontology must identify the relations which can be established among conceptual units, and The meaning postulate of a concept can also hence among lexical units. However, displaying help to determine the IS-A path to its root. For conceptual similarities and differences through example, as can be seen in the meaning taxonomic relations turns out to be more postulates of +DRUG_00 (9) and +BALL_00 chaotic than through meaning postulates linked (10), both of them are a kind of +SOLID_00. to conceptual units. For example, some (10) +(e1: +BE_00 (x1: +BALL_00)Theme ontologies present PART-OF as a taxonomic (x2: +ARTIFICIAL_OBJECT_00 & relation, so VEHICLE and WHEEL would be +CORPUSCULAR_00 & explicitly linked through the ontology +SOLID_00)Referent) structuring itself. However, problems arise *(e2: +BE_01 (x1)Theme (x3: when inheritance takes place, since the +ROUND_00)Attribute) properties of the superordinate are inherited by *(e3: +PLAY_00 (x4: the subordinate. On the contrary, FunGramKB +HUMAN_00)Theme (x5)Referent (f1: can retrieve any kind of relation by means of x1)Instrument (f2: (e4: +THROW_00 their meaning postulates, as can be seen in (7), (x4)Agent (x1)Theme (x4)Origin as well as maintaining the consistency of the (x6)Goal))Manner ^ (f3: (e5: +HIT_00 Ontology. (x4)Agent (x1)Theme (x4)Origin (x6)Goal (7) +(e1: +BE_00 (x1: +VEHICLE_00)Theme (f4: +FOOT_00)Instrument))Manner) (x2: +TRANSPORT_00)Referent) In turn, +SOLID_00 can be a subordinate of *(e2: +COMPRISE_00 (x1)Theme (x3: +CORPUSCULAR_00 or +SUBSTANCE_00, +WHEEL_00)Referent) both of which can be subordinated to 3.6 Commitment #6: multiple +ARTIFICIAL_OBJECT_00 or inheritance +NATURAL_OBJECT_00. The problem is that, unlike +DRUG_00, +BALL_00 does not The model of the FunGramKB Ontology allows share the conceptual path of its ontological multiple inheritance, where a conceptual unit ancestors, since it is neither +SUBSTANCE_00 can be subsumed by two or more concepts, nor +NATURAL_OBJECT_00, as illustrated in creating complex hierarchies in the form of a Figure 2. FunGramKB solves this type of mesh instead of a tree. In case of multiple problems on multiple inheritance by explicitly inheritance, the first predication of the meaning stating in the first predication of the meaning postulate always includes all the superordinate postulates the conceptual route to be taken. concepts of the definiendum, together with one
32 Ontological commitments in FunGramKB
information, so a closed-world assumption #SELF- cannot be applied. CONNECTED OBJECT There are usually three types of rules in a defeasible theory: strict rules, defeasible rules and defeaters. Strict rules are law-like rules, +ARTIFICIAL +NATURAL which have no exceptions: e.g. whales are OBJECT OBJECT mammals, circles are round. On the other hand, defeasible rules can be defeated by contrary evidence: e.g. birds typically fly. Finally, +CORPUSCULAR +SUBSTANCE defeaters are used to block some defeasible rules in order to prevent some conclusions. For example, a rule such as "if an animal is heavy, then it may not be able to fly" may override the +SOLID +LIQUID +GAS +ENERGY conclusion drawn from the defeasible rule “birds typically fly”. The superiority relation, in which a superior rule may override an inferior one, can be expressed by means of rules (e.g. r2 +DRUG +BALL > r1, in case rules can be labelled) or a superiority operator. In FunGramKB, each predication taking part Figure 2. A sample of complex multiple in a meaning postulate is preceded by a inheritance. reasoning operator in order to state if the predication is strict (+) or defeasible (*). The FunGramKB reasoning engine handles 3.7 Commitment #7: non-monotonic predications as rules, allowing monotonic inheritance reasoning with strict predications, and non- monotonic with defeasible predications. The Non-monotonicity is a key issue in human superiority relation is not explicitly stated in the reasoning, because it permits the withdrawal of predications, but the priority principle is applied conclusions which are true just for the typical at the different levels of the MicroKnowing to members of a particular class. On the contrary, resolve conflicts between predications. monotonic inheritance is not able to deal with the intrinsic properties of middle-level concepts 4 Conclusions in the ontology, because this type of inheritance doesn't admit exceptions to inherited default The definition of these methodological criteria values. in the analysis and design phases of the Non-monotonic logics is an umbrella term FunGramKB Ontology and the strict for a family of formalisms based on default application of these guidelines in the reasoning, where the system can override development phase contributed to avoid some previous beliefs in the light of further common errors in conceptual modelling. information. Although non-monotonic logics Moreover, the conceptual acquisition process is has been widely developed in the last thirty much easier when ontology development takes years, resulting in formalisms such as place under a user-friendly interface. Finally, circumscription (McCarthy, 1980), default logic knowledge engineers emphasize the importance (Reiter, 1980) and autoepistemic logic (Moore, of performing an ontology analysis (Noy and 1983), most of these models involve a high McGuinness, 2001). In this respect, the computational cost when being implemented in FunGramKB Ontology is provided with some working applications which require knowledge diagnosis tools for checking the most common representation. On the contrary, and despite its errors and inconsistencies during the ontology less-expressive power, the simplicity of development, and thus validating defeasible logic makes it one of the most computationally the ontological commitments efficient non-monotonic reasoning model for described in this paper. NLP (Antoniou et al., 2000; Maher et al., 2001). Defeasible reasoning allows the Acknowledgements possibility of working with incomplete Financial support for this research has been
33 Carlos Periñán Pascual, Francisco Arcas-Túnez
provided by the DGI, Spanish Ministry of Interdisciplinary Meeting. Chicago, 79-94. Education and Science, grant FFI2008-05035- McCarthy, J. 1980. Circumscription: a form of C02-01/FILO. The research has been co- non-monotonic reasoning. Artificial financed through FEDER funds. Intelligence, 13: 27-39. References Moore, R.C. 1983. Semantical considerations on nonmonotonic logic. In 8th International Allen, J. 1983. Maintaining knowledge about Joint Conference on Artificial Intelligence. temporal intervals. Communications of the Karlsruhe, 272-279. ACM, 26 (11): 832-843. Niles, I. and A. Pease. 2001. Origins of the Antoniou, G. et al. 2000. A flexible framework Standard Upper Merged Ontology: a for defeasible logics. In 8th Workshop on proposal for the IEEE Standard Upper Non-monotonic Reasoning. Breckenridge, Ontology. In Working Notes of the IJCAI- 405-410. 2001 Workshop on the IEEE Standard Bateman, JA., R. Henschel and F. Rinaldi. Upper Ontology. Seattle. 1995. The Generalized Upper Model 2.0. Noy, N.F. and D.L. McGuinness. 2001. Technical report. IPSI/GMD, Darmstadt. Ontology development 101: a guide to Bouaud, J. et al. 1995. Methodological creating your first ontology. Technical principles for structuring an ontology. In report KSL-01-05. Stanford Knowledge IJCAI'95: Workshop on Basic Ontological Systems Laboratory, Stanford University. Issues in Knowledge Sharing. Montreal. Periñán-Pascual, C. and F. Arcas-Túnez. 2004. Gangemi, A. et al. 2002. Sweetening ontologies Meaning postulates in a lexico-conceptual with DOLCE. In A. Gómez-Pérez and V. knowledge base. In 15th International Richard Benjamins, eds. Knowledge Workshop on Databases and Expert Systems Engineering and Knowledge Management. Applications. IEEE, Los Alamitos Ontologies and the Semantic Web: 13th (California), 38-42. International Conference, EKAW 2002. Periñán-Pascual, C. and F. Arcas-Túnez. 2005. Sigüenza. Microconceptual-Knowledge Spreading in Hovy, E. and S. Nirenburg. 1992. FunGramKB. In 9th IASTED International Approximating an interlingua in a principled Conference on Artificial Intelligence and way. In The DARPA Speech and Natural Soft Computing. ACTA Press, Anaheim- Language Workshop, New York. Calgary-Zurich, 239-244. Jackendoff, R. 1990. Semantic Structures. MIT Procter, P., ed. 1978. Longman Dictionary of Press, Cambridge (Mass.). Contemporary English. Longman, Harlow. Lenci, A. et al. 2000. SIMPLE: a general Reiter, R. 1980. A logic for default reasoning. framework for the development of Artificial Intelligence, 13: 81-132. multilingual lexicons. International Journal Van Valin, R.D. Jr. 2005. The Syntax- of Lexicography, 13 (4): 249-263. Semantics-Pragmatics Interface: An Maher, M.J. et al. 2001. Efficient defeasible Introduction to Role and Reference reasoning systems. International Journal on Grammar. Cambridge University Press, Artificial Intelligence Tools, 10 (4): 483- Cambridge. 501. Van Valin, R.D. Jr. and R. LaPolla. 1997. Mahesh, K. 1996. Ontology development for Syntax, Structure, Meaning and Function. machine translation: ideology and Cambridge University Press, Cambridge. methodology. Technical report MCCS-96- VOX-Universidad de Alcalá de Henares. 1995. 292. Computing Research Laboratory, New Diccionario para la Enseñanza de la Mexico State University, Las Cruces. Lengua Española. Bibliograf, Barcelona. Mahesh, K. and S. Nirenburg. 1995. Semantic classification for practical natural language processing. In 6th ASIS SIG/CR Classification Research Workshop: An
34