Apologiae Ontologiae

Treatment of meaning in NLP is greatly facilitated if semantic analysis and generation systems rely on a language-neutral, independently motivated world model, or ontology. However, the benefits of the ontology are somewhat offset in practice by the difficulty of its acquisition. This is why a number of computational linguists make a conscious choice to bypass ontology in their semantic deliberations. This decision is often justified by questioning the principles underlying ontologies and by challenging the ontology-based semantic enterprise on the grounds of its ostensible irreproducibility. In this paper we briefly illustrate the expressive power of ontological lexical-semantic descriptions used in the Mikrokosmos machine translation project and make a comparison with some of the non-ontological approaches to lexical semantics. We argue that these approaches, in reality, rely on ontologies in everything but name. We claim that no underlying principles for ontologies are possible and explain why the charge of irreproducibility is not valid. The central tenet of ontology-based semantics is "grounding" the meanings of lexical units in an independently motivated and interpreted system of symbols, an ontology. An ontology-based lexicon is an important component of the Mikrokosmos multilingual MT project (see, e.g., Onyshkevych and Nirenburg, 1994; earlier versions of the same paradigm were described in Nirenburg et al., 1992 as well as Carlson and Nirenburg, 1990 and Meyer, Carlson and Onyshkevych, 1990). Advantages of lexicons supported by ontologies include support for treatment of synonymy (and nearsynonymy), control over uniformity of the grain size of lexical descriptions, enabling the treatment of lexical gaps in translation as well as support for a variety of inferencing mechanisms necessary for lexical and other kinds of disambiguation. It has been widely accepted since Bar Hillel (1960) that without such mechanisms a large number of textual phenomena cannot be fully treated. It might, therefore, seem strange that lexicographers and MT researchers are not flocking to this method (cf. a similar lament in Bateman, 1993, p.83). But, of course, there are important reasons for that, some connected with linguistic tradition, and some others with purely engineering considerations. Very briefly, the major objections can be summarized as follows: •The Redundancy Objection: it is often possible to resolve semantic ambiguities using non-semantic (syntactic, morphological, etc.) clues. This observation has led to a variety of claims that semantics is redundant. As a practical consequence, recognition supplanted interpretation in system designs, as attempts were made at disambiguation but not at representation of the meaning of text. 1. Also of NLP Lab, Purdue University, W. Lafayette, IN 47907 USA 2. Also of Computational Linguistics Program, Carnegie Mellon University, Pittsburgh PA 15213 USA 3. Cf. the many suggestions for syntactic treatment of prepositional phrase attachment, a problem whose solution is universally accepted to require semantics.