<?xml version="1.0" standalone="yes"?>
<Paper uid="W91-0207">
  <Title>Representation of Semantic Knowledge with Term Subsumption Languages</Title>
  <Section position="3" start_page="67" end_page="68" type="metho">
    <SectionTitle>
2 Knowledge representation and TSLs
</SectionTitle>
    <Paragraph position="0"> In the field of AI, many researchers have addressed the problem of knowledge representation; in this area semantic networks played an important role. In semantic networks the knowledge is described by nodes and links. While Quillian aimed at the representation of word meanings \[Quillian 68\], semantic networks have also been used to model propositions, events, spatial relationships and so on. Since semantic networks failed in providing a unique semantic interpretation, several researchers examined the &amp;quot;semantics of semantic networks&amp;quot; (\[Woods 75\], \[Brachman 79\]).</Paragraph>
    <Paragraph position="1"> Another approach is to organize knowledge in chunks called frames \[Minsky 75\] which are used to represent &amp;quot;stereotypical situations&amp;quot;. Frames typically allow the specification of default slot values, perspectives and attached procedures. Collections of frames can be combined to frame-systems. The expressive power of frame systems makes it impossible to provide a well defined semantics for them.</Paragraph>
    <Paragraph position="2"> Both, elements of different network formalisms and basics of the frame theory, have influenced the structural inheritance networks and the subsequent implementations (KL-ONE, \[BrachmanSchmolze 85\]). The basic idea is to postulate a level of knowledge representation with &amp;quot;knowledge structuring primitives, rather than particular knowledge primitives&amp;quot; (\[Brachman 79\]), the so-called epistemological level. The basic buildung blocks of KL-ONE representations are &amp;quot;concepts&amp;quot;, i.e. structured conceptual objects. &amp;quot;Roles&amp;quot; are possible relationships between two concepts. The subsumption relation organizes the concepts in a concept taxonomy. Concepts are described with respect to their superconcepts by restricting and differentiating roles. In particular, roles can be restricted by the number (number restriction) and the range (value restriction) of allowed role fillers. If the specified restrictions constitute necessary and sufficient conditions for the concept, it is called a defined concept, wheras primitive concepts only need necessary conditions. Classification, an important inference mechanism of KL-ONE like systems, inserts concepts at the correct place in the concept hierarchy.</Paragraph>
    <Paragraph position="3"> A logical reconstruction of KL-ONE revealed that the semantic status of a number of notions of KL-ONE was rather unclear. TSLs are formal knowledge representation languages derived from KL-ONE providing well-defined semantics which enables the decision whether the inferences are sound and complete. A number of KR systems based on TSLs have been developed, for instance, Krypton \[Brachman et al. 85\], KL-Two \[Vilain 85\], Back \[Peltason et al. 89\], Loom \[MacGregorBates 87\]. Besides a component for defining concepts and reasoning about the relationships between concepts (terminological component, TBox) these systems include an assertional component (ABox) that allows the definition of assertions about individuals.</Paragraph>
  </Section>
  <Section position="4" start_page="68" end_page="72" type="metho">
    <SectionTitle>
3 Representing word meanings with TSLs
</SectionTitle>
    <Paragraph position="0"> While aspects of syntactic structure are rather well understood in NLP, the problem of representing semantic information is far from being solved. Scientists from various research areas, e.g., linguistics, philosophy of language, lexicology and artificial intelligence, are dealing with problems concering the nature of word meanings and means for their representation.</Paragraph>
    <Paragraph position="1"> In the following, we make some general remarks on semantic description without going into details of any semantic theory. A particular aspect all semantic theories are concerned with is the principle of compositionality: the meaning of a sentence is a function of the meaning of each of its components and its context. As a first approximation, one could abstract away from the context or assume a typical context (paradigmatic analysis). A good semantic theory, however, must allow the notion of semantic variation. So, in addition to the enumeration of possibly different senses of a word, we have to examine the meaning of a word in varying contexts (syntagmatic analysis). Also, a finite enumeration of word senses does not suffice to explain the creative use of words (\[BoguraevPustejovsky 90\]).</Paragraph>
    <Paragraph position="2"> In addition to syntactic and semantic knowledge, the process of understanding natural language involves extralinguistic (encyclopedic) knowledge. Mechanisms for the combination of these types of knowledge are an important prerequisite for natural language understanding.</Paragraph>
    <Paragraph position="3"> The analysis of word meanings is also the subject of dictionary definitions. During the last years, there has been an increasing interest in methods for extracting lexical semantic information from machine-readable dictionaries (see for example \[BoguraevBriscoe 89\]).</Paragraph>
    <Paragraph position="4"> There is, however, no consensus about the representation formalism into which the meaning descriptions should be transformed. In our opinion, the suitability of AI-based KR formalisms for the representation of semantic knowledge and as a means for the combination of linguistic and extralinguistic knowledge has to be investigated. As a first step, we are analyzing a number of simple dictionary definitions in order to derive some basic requirements which a representation language has to meet to be usable for the representation of word meanings. The results will allow us to assess the suitability of TSLs for that matter.</Paragraph>
    <Paragraph position="5"> In a dictionary, different meanings of a word are usually specified by means of definitions, examples, references and pictures. 1 Subsequently, we will concentrate on the analysis of meaning definitions. There exist different types of definitions, e.g., 2 * definition by reference to synonyms  - acclaim: applause; approval.</Paragraph>
    <Paragraph position="6"> - complaint: illness; disease.</Paragraph>
    <Paragraph position="7"> - jowl: jaw.</Paragraph>
    <Paragraph position="8"> * definition by reference to antonyms - absolute: not relative.</Paragraph>
    <Paragraph position="9">  - affected: not natural or genuine.</Paragraph>
    <Paragraph position="10"> - wild: (of plants) not cultivated.</Paragraph>
    <Paragraph position="11"> * definiton by reference to hyperonyms and modifying elements - park: public garden or public recreation ground in a town.</Paragraph>
    <Paragraph position="12"> - bobsled: large, long sleigh with brake and steering wheel, used for racing. - blackboard: board used in schools for writing and drawing on with chalk.</Paragraph>
    <Paragraph position="13">  We will have a closer look at nominal definitions of the latter type which contain a genus term of the defined word. 3 The first part of the definition of park, namely &amp;quot;public garden&amp;quot;, can be represented by a concept with superconcept garden and a relation called PROPERTY to the concept public: a park is a garden with PROPERTY public A visualization of this definition (in a KL-ONE-Iike graphical notation) is given in  The concept identifiers, e.g., garden, in the example have to be distinguished from the corresponding word forms. 4 Each concept represents one of the possible meanings of the corresponding word.</Paragraph>
    <Paragraph position="14"> This example shows at least two problems of representing word meanings with TSLs. The main problem arises from the fact, that the epistemological primitives of a TSL do not give enough specifications for the representation of word meanings. Many nominal definitions contain nouns modified with adjectives. We need a number of predefined  and modifiable roles, like the relation PROPERTY between the nominal concepts and the concepts representing these adjectives.</Paragraph>
    <Paragraph position="15"> Another important relation for the representation of noun meanings is the part-whole relation, called meronyrny. An example is the definition of bobsled:' large, long sleigh with brake and steering wheel, used for racing. The concept bobsled refers to the parts brake and steering-gheel:  This example shows a third class of important relations for the definition of noun meanings, i.e. the normal uses or functions of a thing. The relations can be further specialized, e.g., the PART relation describes different types of meronymy like COMPONENT, MEMBER, MATERIAL (see \[Miller et al. 90\]).</Paragraph>
    <Paragraph position="16"> The part-whole relation is a relation between nominal concepts and can be represented in a TSL-based KR formalism by means of roles. Number and type of given parts can be described by number restriction and value restriction respectively. Different parts of a thing have to be specified by different subroles of a more general PART role. In the example above the two components brake and steering-wheel have to be related to bobsled by  The PROPERTY relation is a relation between nominal concepts and &amp;quot;property concepts&amp;quot;, e.g., public in the first example. Such kinds of concepts do not fit into a term hierarchy because they usually do not have suitable superconcepts or individuals. Consequently, the most important inference mechanism of TSLs, namely classification is unsuitable for the representation of property concepts. We presumably need another formalism for the representation of properties, in which other relations, for example antonymy, play an important role (\[GrossMiller 90\]). This formalism has to be combined with the term subsumption formalism.</Paragraph>
    <Paragraph position="17"> The FUNCTION relation relates nominal concepts to concepts representing verb meanings, e.g.,  The representation of verb meanings involves a number of further problems, e.g., the representation of space and time, that can not be investigated in this paper. Returning to the representation of nominal concepts, we try to represent the complete definition of park: public garden or public recreation ground in a town:  This example demonstrates the necessity of concept disjunction. Disjunction is frequently used in definitions. Therefore, a KR formalism adequate for the representation of semantic knowledge has to provide a form of concept disjunction. Disjunction is not allowed in TSLs because it is contrary to the claim that concepts should only be defined with respect to their superconcepts. In the example, both &amp;quot;garden with property public&amp;quot; and &amp;quot;recreation-ground with property public and with location town&amp;quot; (subsequently termed p-garden and p-recreation-ground respectively) are subconcepts of park, as illustrated in Fig. 2.</Paragraph>
    <Paragraph position="18">  This representation of park is unsuitable because it does not guarantee that all parks are public gardens or public recreation grounds in a town. A solution to this problem is the notion of a &amp;quot;covering&amp;quot;, which was inroduced in NIKL (\[Moser 83\]). If park is defined as a covering of p-garden and p-recreation-ground, every instance of park will be an  instance of at least p-garden or p-recreation-ground. In NIKL coverings are used to enhance concept specifications but they are ignored by the classifier.</Paragraph>
    <Paragraph position="19"> The investigation of further nominal definitions revealed that several other extensions to TSLs are necessary: * boil: hard (usn red, often painful) poisoned swelling under the skin, which bursts when ripe.</Paragraph>
    <Paragraph position="20"> * boulevard: wide city street, often with trees on each side.</Paragraph>
    <Paragraph position="21"> These examples indicate the existence of typical features of a concept, called &amp;quot;defaults&amp;quot;. Inferences with defaults require the use of nonmonotonic reasoning techniques which are outside the scope of a TSL classifier.</Paragraph>
    <Paragraph position="22"> The following examples show that similarity between concepts is another important relation.</Paragraph>
    <Paragraph position="23"> * lemur: nocturnal animal of Madagascar, similar to a monkey but with a foxlike face. * marimba: musical instrument similar to the xylophone.</Paragraph>
    <Paragraph position="24"> * quail: small bird, similar to a partridge, valued as food.</Paragraph>
    <Paragraph position="25"> Most of the requirements mentioned above can not be integrated into a TSL maintaining sound and complete inferences. Because these requirements seem to be necessary for the representation of word meanings, TSLs can provide only a &amp;quot;representational kernel&amp;quot;, which has to be embedded into a component with greater expressive power. This component has to allow enhanced concept descriptions, e.g., concept disjunction, defaults and similarity of concepts.</Paragraph>
  </Section>
  <Section position="5" start_page="72" end_page="73" type="metho">
    <SectionTitle>
4 An Approach to the Integration
</SectionTitle>
    <Paragraph position="0"> As a first step we have implemented a KR system that consists of a terminological and an assertional component. The formalism used is similar to the formalism described in \[Nebel 90\]. The restricted expressiveness enables inferences that are sound and complete and makes the formalism suitable as a platform for the extensions described above.</Paragraph>
    <Paragraph position="1"> A small fragment of the TBox language is illustrated in the following example. The concept bobsled is described as a subconcept of sleigh with two PART relations, namely COMPONENT1 and COMPONENT2. These roles have to defined separately as specializations of COMPONENT. The concept bobsled is primitive because the specifications are necessary but not sufficient for the definition of bobsled.</Paragraph>
    <Paragraph position="2">  The system is implemented in CLOS (Common Lisp Object System) and an overview of its syntax and semantics is given in \[Forster et al. 91\].</Paragraph>
    <Paragraph position="3"> The formalism has to be integrated into a lexicon, a possible architecture of which is shown in Fig. 3. The lexicon consists basically of two components: one containing word forms and another for the representation of word meanings. The latter has to be embedded into a component for the representation of more general world knowledge.</Paragraph>
  </Section>
class="xml-element"></Paper>