<?xml version="1.0" standalone="yes"?>
<Paper uid="J92-3003">
  <Title>A Practical Approach to Multiple Default Inheritance for Unification-Based Lexicons</Title>
  <Section position="3" start_page="0" end_page="312" type="intro">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> In essence, the primary task of a computational lexicon is to associate character strings representing word forms with various types of information, able to account for their distribution within a sentence and for their contribution to the meaning of a text.</Paragraph>
    <Paragraph position="1"> Economical lexical organization appears to require the ability, on the one hand, to make general statements about classes of words, and, on the other, to express exceptions to such statements affecting both individual words and subclasses of words.</Paragraph>
    <Paragraph position="2"> These considerations have provoked interest in applying to the domain of the lexicon AI knowledge representation techniques involving the notions of inheritance and defaults. ~ Unfortunately, many of the schemes that have been proposed are highly complex; departing from simple tree-form taxonomies dramatically increases the problems of dealing in a systematic fashion with default inheritance. Not only are the intuitions underlying the behavior of systems such as those of Touretzky (1986) and Sandewall (1986) unstable (as Touretzky et al. \[1987\] show), Selman and Levesque (1989) demonstrate that certain of them are NP-hard and thus effectively intractable.</Paragraph>
    <Paragraph position="3"> In a well-defined domain such as the lexicon there remains the possibility of mitigating these problems by adopting a principled compromise; it may be advantageous to surrender some of the generality of an inheritance system, if the more restricted version that results retains sufficient expressive power for the domain in question. We expand on this below and propose a number of simplifications that are motivated by the particular task that the system is intended to perform.</Paragraph>
    <Paragraph position="4">  Daelernans and Gazdar 1990, and the contributions to Briscoe et al. 1991. We discuss some of this work below.</Paragraph>
    <Paragraph position="5"> (~ 1992 Association for Computational Linguistics Computational Linguistics Volume 18, Number 3 The system described here has been implemented as part of the ELU 2 unification grammar development environment for research in machine translation, made up of parser, generator, lexicon, and transfer mechanism. The user language resembles that of PATR-II (Shieber 1986a), but provides a larger range of data types and more powerful means of stating relations between them. Among the requirements imposed by the context within which this system is used are (i) the ability to both analyze and generate complex word forms, (ii) full integration with existing parts of the ELU environment, and (iii) the ability to accommodate a relatively large number of words. In particular, an important objective is to preserve as far as possible the flavor of this type of environment: a specialized programming language for linguistic descriptions, suitable for interpretation by a variety of programs performing tasks such as analysis and generation of sentences and phrases, lexical lookup, etc., which associates with natural language expressions representations defined by the writer of a linguistic description and employs unification as the method of combining information and enforcing constraints. The intention is that the ELU language should be as general as possible, in the sense of imposing minimal restrictions on the form of linguistic analyses and descriptions.</Paragraph>
    <Paragraph position="6"> Unification is an attractive choice as the basic computational mechanism of such a system, for well-known reasons: its algebraic properties permit a declarative, monotonic semantics for the language that is independent of whatever programs interpret it. However, it does not always meet the requirements of practical linguistic descriptions. The treatment of exceptions is a case in point; negation and conditional statements of the kind excluded by pure unification frequently play a large part in accounts of natural language phenomena, and the best-motivated area for the application of these ideas appears to be the lexicon. It is this problem that the current work seeks to address. We would hope that the interest of the system presented here extends beyond the particular environment for which it has been developed; as an attempt to reconcile two apparently disparate paradigms, the combination of inheritance and unification may be of interest in quite different contexts.</Paragraph>
    <Paragraph position="7"> The remainder of the paper is organized as follows. Section 2 discusses general issues concerning the organization of the system: the basic formalism is introduced, together with the notions of multiple inheritance, default inheritance, and class precedence. Section 3 describes the system in greater detail, showing how the form of the inheritance hierarchy determines how information in lexical specifications is combined, and Section 4 provides some comparisons with related work. Some more extended linguistic examples appear in Section 5; these serve to illustrate the various mechanisms described earlier and suggest methods that might be adopted in formulating other analyses. Finally, in Section 6, the current implementation is discussed.</Paragraph>
    <Paragraph position="8">  2. The Lexicon as a Hierarchy</Paragraph>
    <Section position="1" start_page="0" end_page="312" type="sub_section">
      <SectionTitle>
2.1 An Overview of the Formalism
</SectionTitle>
      <Paragraph position="0"> An ELU lexicon consists of a number of 'classes,&amp;quot; each of which is a structured collection of constraint equations encoding information common to a set of words, together with links to other more general &amp;quot;superclasses.' For example, if an 'intransitive&amp;quot; class is used to express the common syntactic properties shared by all intransitive verbs, then particular instances of intransitive verbs can be made to inherit this informa2 &amp;quot;Environnernent Linguistique d'Unification&amp;quot; (Estival 1990). See also Johnson and Rosner (1989) for a description of the earlier UD system on which ELU is based.</Paragraph>
      <Paragraph position="1">  Graham Russell et al. A Practical Approach to Multiple Default Inheritance tion by specifying the 'intransitive' class as one of their superclasses--it then becomes unnecessary to specify the relevant properties individually for each such verb. Similarly, the 'intransitive' class need not express any of the more general properties of the 'verb' class. The lexicon may be thought of as a tangled hierarchy of classes linked by inheritance paths, with, at the most specific level, lexical classes and, at the most general, classes for which no superclasses have been defined; these therefore inherit no information from elsewhere. Basic lexical entries are themselves classes, 3 and any information they contain is standardly specific to an individual word; lexical and non-lexical classes differ in that analysis and generation take only the former as entry points to the lexicon.</Paragraph>
      <Paragraph position="2"> It is possible to define certain properties of a class as defeasible. Such a property will be inherited from a superclass only if it is consistent with all properties contributed by more specific classes; if it is not consistent, then it is ignored. Defeasible constraints imposed on a class of words may thus be overridden by information associated with individual members or subclasses of that class; this is the notion of default that underlies the system. It is possible, for example, to place in the class expressing general properties of verbs an equation such as '&lt;aux&gt; = no' (i.e. &amp;quot;typical verbs are not auxiliaries&amp;quot;), while placing the contradictory specification '&lt;aux&gt; = yes' in a sub-class from which only auxiliaries inherit. The ability to encode exceptional properties of lexical items is extremely attractive from the linguistic point of view; the lower the position in the hierarchy at which the property appears, the more exceptional it may be considered.</Paragraph>
      <Paragraph position="3"> Like other parts of an ELU description, a lexicon consists of a set of statements that are compiled into an internal format for use in analysis and generation of linguistic expressions. Section 2.6 describes some relevant aspects of the lexical compilation  process.</Paragraph>
      <Paragraph position="4"> 2.1.1 Class Definition. A class definition consists of the compiler directive '#Class' (for a nonlexical class) or '#Word' (for a lexical class), 4 followed by: (i) the name of the class (ii) a (possibly empty) list of its direct superclasses (iii) a (possibly empty) 'main' equation set (iv) zero or more (possibly empty) 'variant' equation sets.</Paragraph>
      <Paragraph position="5">  direct superclass of the current class. This is used in computing the relative precedence of classes in the lexicon for the purpose of default inheritance (see Section 2.5); it may be empty if the class has no superclasses, i.e., if it is one of the most general in the lexicon, and thus inherits no information. Precedence of direct superclasses is</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>