<?xml version="1.0" standalone="yes"?>
<Paper uid="E91-1042">
  <Title>Modelling Knowledge for a Natural Language Understanding System</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
~&amp;quot;LILOG Expetimelltal EnvlronmenC'
</SectionTitle>
    <Paragraph position="0"> Our approach embodies modules oriented towards levels oflingadstic investigation like morphology, syntax and semantics, h~ addition the modules differentiate between analysis and the generation processes. In the ideal case, all processes and modules will be supported by conunonsense \]ulowledge.</Paragraph>
    <Paragraph position="1"> A crucial problem in tiffs context is the construction of an adequ'ate background knowledge base.</Paragraph>
    <Paragraph position="2"> The: need for a methodology is obvious. First steps have been made in expert system research, where both domain andtask are for the most part clearly specifiable. This does not hold for systems with natural language - and conunon sense orientation.</Paragraph>
    <Paragraph position="3"> In what follows, we will outline the lulowledge engineering approach in LILOG along three dimensions. null Task: Domain and te~cts were selected in order to cover a wide variety of lingalistic phenomena to be handled by the linguistic parts of the system (i.e. parslug:and generating components). Iat order to prove the;appropriate understanding of the texts, the at.</Paragraph>
    <Paragraph position="4"> chitecture was d~sigqled a.o. as a question/answer system. Hence, we get the additional task to generate language.</Paragraph>
    <Paragraph position="5"> Domain: For LEU/2, the domain was restricted to travel guide information about the city center of Diisseldorf. As a first step, a set of written data was obtained by travel guides, supplemented by travel agencies and a local inspection of Dfisseldorf city center.</Paragraph>
    <Paragraph position="6"> The set of different entities was to meet the following eonditlona: it should be large enough for a relevant size of the knowledge base, interconnected enough to allow for interesting inferences but at the same thue small enough for being handled within a In'ototypical implementation.</Paragraph>
    <Paragraph position="7"> We decided to work with a couple of short texts (frequently found in travel guides), which describe - 239 particular sightseeing items, and a one page narrative text about a group of people on a prototypical sightseeing tour. In the next step, the chosen texts were classified according to lingnistic criteria and analyzed for their propositional contents.</Paragraph>
    <Paragraph position="8"> Granularity: hi order to obtain a first hint at the variety of text understanding tasks whidi LEU/2 was intended to deal with, native speakers were asked to formulate questions and to provide acceptable answers concerning the contents of the texts.</Paragraph>
    <Paragraph position="9"> The selection of items aud the way these native speakers talked about them, served as guideline to determine an appropriate granularity of the luiowledge base.</Paragraph>
    <Paragraph position="10"> The overall performance of the system is determined by the interaction of it's components. Due to the modular approach, the relevant subtasks of the kno,~ledge base had to be separated from those of the lezical, syntactic, semantic analysis components and the generation module. As a result of this prelinfinary investigation, three dimensions of knowledge turned out to be crucial to the modelling process.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Dimensions of Knowledge
</SectionTitle>
    <Paragraph position="0"> We will discuss knowledge from two different perspectives. On the one hand we have those:conditions which lead to qualitative requirements concerning the contents of the lu~owledge base. The other perspective concerns aspects induced by forreal devices, i.e. the kalowledge representation fornmlism used.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1 Qualitative Dimensions
</SectionTitle>
      <Paragraph position="0"> If you consider knowledge representation as a special case of model theory, you will get a hint of how to proceed. As to the breadth of the model, the first dimension at issue, this means: The job of the representing world is to reflect some aspects of the represented wodd in some fashion.\[Palmer, 1978\] As regarding grcznularily, the second dimension, a model reflects only a subset of the characteristics of the entities it represents. This, in turn, determines the depth of the model A tldrd dimension is given by the complexity o\] the task the model is intended to cover.</Paragraph>
      <Paragraph position="1"> All three dimensions are shown in picture 1.</Paragraph>
      <Paragraph position="2"> Some of the consequences for the model in LILOG following from this view of knowledge representation are described below.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.2 Formal Devices of Representa-
</SectionTitle>
      <Paragraph position="0"> tion In the field of logic based formalisms for coding background knowledge in natural language process-Breadth of the domain Depth of the model -&lt;__  ing systems, there is some controversy on tile design and use of formal constructs. Topics in this debate are tile function of axioms compared to recent expert system teHmology, the function of structured concept hierarchies \[Monarch and Nirenburg, 1987\], the quality and number of additional attributes (roles in KL-ONE like systems) or syntactic validation Criteria \[Horacek, 1989\]. Our approach aims at finding useful sdectional criteria for different expressive means of the formalism LLwoo in order to bridge the actual \[gap between problem driven and technology driven ~ research.</Paragraph>
      <Paragraph position="1"> We can make use of two kinds of formal constructs: null s A frmne-des.cription language similar to KL-ONE (cf. e.g. \[Brachman and Schmolze, 1985\]), which serves to represent tile terminob ogy of the domain by means of ! sort expressions for classes of entities, organized hierarchically as sets and subsets (i.e. the logical subsumption relation), mid - two Place predicates and functions (i.e.</Paragraph>
      <Paragraph position="2"> features and roles), attached to specific sorts and constituting functional and relational connections between sorts, and * axioms of first order predicate logic, expressing inferential dependencies between domain terms hi form of the axiomatic semmltics for those terms.</Paragraph>
      <Paragraph position="3"> So the formalism used here' is colnparable to e.g. KRYPTON (s.e.g. \[Brachman et ~., 19851).</Paragraph>
      <Paragraph position="4"> In the following, we will discuss the qualitative dimensions of knowledge in more detail. We will focus the qualitative criteria by differentiating them according to our scenario.</Paragraph>
      <Paragraph position="5"> SSee \[Lehnert, 1988\] for that distinction.</Paragraph>
      <Paragraph position="6"> tFor a detailed description of the formalism LLILOO see</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Demands Resulting from the
Task
</SectionTitle>
      <Paragraph position="0"> As mentioned above, the task of our system is to simulate text understanding. This requires a transfer of insights from linguistic research into knowledge engineering. In the ideal case, structures of the model will be strongly influenced by natural language analyses.</Paragraph>
      <Paragraph position="1"> Linguistic knowledge is relevant in various respects: null Word orientation, for example, implies close hxterrelationships with research on lexical knowledge: afrdiated generic terms, discriminating features, idiosyncratic aspects of use, etc. However, you may run into difficulties by relating syntactic categories (like word classes) with conceptual structures. So thematic roles cannot be directly trmtsformed into ontological roles as a part of the background knowledge.</Paragraph>
      <Paragraph position="2"> In the sentence The bus took the participants of the conference to the city center, s the 'bus' is ml agent of art. event from the syntactic point of view attd at the same time conceptualized as instrument (and not agent) of mx event in an ontological sense.</Paragraph>
      <Paragraph position="3"> Sentence oriented linguistic investigation implies the reconstruction of knowledge on the sentence level, as opposed to the meaning of single words or of textual structures. As all illustration might serve temporal information about the progress of actions or situations. Theoretical: work in this field was initiated e.g. by Z. Vendlcr \[Vendler, 1967\] with his analysis of verbs and times.</Paragraph>
      <Paragraph position="4"> His differentiation of states, activities, accomplishments and achievements has been established as a well known classification of verbs. One important criterion for this disthtction is the goal-orientedness of the concerned verbs: states and activities are by definition not goal-oriented, whereas accomplishments and achievements are goal-oriented in a temporaUy extended or punctual way, respectively.</Paragraph>
      <Paragraph position="5"> The aspect of goal-orlentedness turned out to be central in our domain, e.g. as to directional verbs of movement. The sentence The tourists took the bus to the Rhine and went for a boat trip. s SThe German version of the sentence is part of the text (c)orptm of LEO/2: &amp;quot;Der Bus brachte die Teilnehmer der Konferens in die Innenstadt&amp;quot;.</Paragraph>
      <Paragraph position="6"> e&amp;quot;Die Touristen nalunen den Bus bls EUlI1 R\]lel,ll tad much, an elnen Boo,san,flag.&amp;quot; allows to hlfer by default that the tourists reached their goal (the Rhine), because the location of the following event (the boat trip) is the stone as the arrival point of the bus ride.</Paragraph>
      <Paragraph position="7"> By introducing goal-orientedness as a part of the definition of events, it will hence be possible to give an afflrmativc answer to the question null Were the!ourists at the Rhine? T s Moreover, a text necessarily involves discourse oriented information. Text understanding phenomena like annphora resolution can only be accounted by accessing background knowledge concerning interconceptual relation.</Paragraph>
      <Paragraph position="8"> The tourists went for a boat trip. They took the seats on the sundeck, s In order to capture the meaning of these sentences, three:steps have to be inferred: A boat trip is usually undertaken with a boat; a boat often has a sundeck; and a sundeck mostly offers seats.</Paragraph>
    </Section>
    <Section position="4" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.2 Demands Resulting from the
Domain
</SectionTitle>
      <Paragraph position="0"> In the LEU/2 context, we have to deal with the comprehensive task of text understanding mtd a relatively narrow domain. Consequently, the general problem of conceptualization is limited by a restricted number of entities relevant to our field. Modelling these entities includes both the selection of concepts which appear ill the domain, and the plausible combination and sununhlg of recurrent concepts. The plausibility of modelling decisions in this sense can be judged from an engineering point of view in terms of optimizing search space (system performance) and from a philosophical point of view in terms of the principle &amp;quot;of economy o/ the ontolog~ The concepts RBSTAURATION, CONSTRUCTION and RBNOVATION nlay serve as ml illustration taken from our domain. As they share simJlar aspects anti inferences, we decided to introduce the supersort MODIFICATION (see section 4).</Paragraph>
    </Section>
    <Section position="5" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3,3 Granularity: Depth of Mod-
elling and Inferenclng
</SectionTitle>
      <Paragraph position="0"> In the third qualitative dimension of knowledge we have to face the problem of dellmitatlng the depth of the model in order to reduce complexity. As it is not possible to give r deg'Waren die To~tEisten ant Rheln?&amp;quot; tThe German version of the sentence is part of the text corpus of LEU/2t ~Die Tot~isten macttten elnen Boo,taut. flux. Sic nahmea dl.'e Platte auf dem $onnendeck ein&amp;quot;. - 241 an exhaustive system of categories o, it Seems legitimate to deternfine primitive concepts dependent on the chosen task and domain. In addition, selectional criteria for clusters 'of inferences have to be determined. (See example in section 4). As a possibility of measuring the depth of a model, tlayes ({llayes, i979\]) proposed a ratio of axioms per concept. : Aside from measuring the expression of dimensions of knowledge by me/ms of quantitative data, it is important to consider qualitative dependencies between the depth and task Of the model on the one hand and between the depth and domain on the other.</Paragraph>
      <Paragraph position="1"> Depth in relation to the task Within the task of text understmlding,i some requirements of representation are e.g.: goal orientation, cuhnination, causal connections, intention, etc. \[Trabasso and Sperry, 1985\].</Paragraph>
      <Paragraph position="2"> lal all these cases the dlosen granularity has strong impact upon the resohttion of interrelations in the texts) deg Depth in relation to the domain This connection cmt be illustrated by the following exmnple: A typical event of our domain is RBSTAURATION. ~n our scenario, t0uristic aspects like the architect (agent), the time and the object concerned (e.g.,, tlle facade) will be of crucial importance. Given a different scenario like the protection of historical monuments, we would have to face an interest in considerably more details, requiring the cholcc of a deeper granularity.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 Design of the Knowl-
</SectionTitle>
    <Paragraph position="0"> edge Base Ill this section, we first want to give a brief survey of the ontology. After that, wc will take up the sorts and regularities mentioned so far and present a structured exemplary mo~lel formalized in LLXLOO * Sort expressions arc used to represent the categories of our domain model. The upper strutlure of the resulting ontology portrays some generaLized schemes of organigation of relative domain-independence. When descend!ng the model towards the lower #fracture, the categories arc defined much closer to the word level and therefore domain-specific in the sense of ezplicit text \]ulowledge. |I As already nlentioncd, we want to simulate understanding of basically two different types of .......................</Paragraph>
    <Paragraph position="1"> degSee for example \[Tamas, 1986, p. 509\] tdegFor a more detailed di.cu.sion, see \[Pirleln, 1900\]. t l This differentiation between upper and lower ~tructure of the model is introduced by \[Maim et al., 1985\]. texts, i.e. short texts describing single sightseeing items and narrative texts dealing with sequences of events. This leads us to the requlremcnt of both all object-oriented and all event-oriented part of the eoncephlal hierarchy. null Consequently, one of our basic design decisions is due to J. Hobbs (cf. \[llobbs et al., 1987\]) and results ill a reification of predicates. So in our model all events, states etc. have concept status on their own.</Paragraph>
    <Paragraph position="2"> This technique enables us to model the case frames for verbs in all analogical manner to the lexical entrie~ of the analyzing component as well as to incorporate tile structures for events etc. within the categories alike the definitions for objects 121 It makes sense to think about objects as wcU as about events in terms of their spatial mad temporal environment, although these knowleklge specifications will obviously be quite different.</Paragraph>
    <Paragraph position="3"> An example taken from the event cluster may serve as an illustration of several consequences of the criteria mentioned above. As to the breadth of the model, the relevance of the event part of the ontology appears intuitively plausible with respect to our domain, namely a scenario of cities, with modifying events. We have to deal with sights of the city like facades of important.buildings, and the events of modification related to them show a considerable resemblance of ilnportant features of meaning - although the verbs are no real synonyms in the linguistic sense.</Paragraph>
    <Paragraph position="4"> Figure 2 shows a screen dump witll the relevant part of' the concept hierarchy. The picture illustrates the effect of bundling that the introduction!: of adequate superconcepts has, and which allows for structured inferencing in terms of system efficiency, hi this part of our concept llicr~chy the boarderlinc between Upper Structure attd Lower Structure is clearly identifiable. When descending the hierarchy, the sort KONSTRUKTIVSIT falls out into several domain-dependent subsorts.</Paragraph>
    <Paragraph position="5"> The figure is followed by the respective sort expressions written in the bt.xLoo list structure(the sort KONSTRUKTIVSIT in the figur(c) corresponds to CONSTRUC~PION in the English list of sort expressions), expanded by roles and features which do not appear in tile graphic representatimt. It should be noted here that a third kind of information is omitted even in the list notation. More general roles and fentures (llke e.g. agent, time m~d so on) are inherited by supcrconcepts and not visible in neither presentation. (The short lille in tile upper left corner of some concept boxes indicate the existence of additional hidden superconcepts.)  The definition of the relevant event concepts in LLILOO is followed by an axiom which transfers information about the time of a construction event to the beginning of lifetime of the concerned object. This kind of structured modelling allows to dispense with writing similar axioms for a number of resembling events. In order to demonstrate task orientation, it would be necessary to consider a broader part of the ontology, because aspects like intention, causality or culmination have been modeUed separately. In addition, one would have to take a closer look at the ensemble of connected components in the system. The limitation of the depth of the model can be seen from the fact that the event concepts discussed do not have more differentiated subconcepts mid, of course, from the fact that not nil possible roles and features have been integrated into the model.</Paragraph>
    <Paragraph position="6"> In a scenario &amp;quot;protection of historical monuments&amp;quot;, for example, the instruments of renovation might be central and would induce a partly different granularity in the model.</Paragraph>
    <Paragraph position="8"> sort building &lt; material C/onstruotioa.</Paragraph>
    <Paragraph position="9"> The twofold ~nodelllng of PHYSICAL and MBN-TAb CONSTRUCTION is e.g. necessary to distinguish ideas developped by an architect from the realization of the building. 13 For constructive events one can define the following regularity (axiom): axiom rule30 gerall DI : construction,</Paragraph>
    <Paragraph position="11"> moets(DI,T3).</Paragraph>
    <Paragraph position="12"> The relation meet,: is one expression of our axionlntizati0n of Allen's time interval logic \[Allen, 19831 in LLILOG . Rule30 exemplifies a transformation rule between the clusters of events nnd objects, respectively.</Paragraph>
    <Paragraph position="13"> Our task setting implies certain ways of interaction between Knowledge Engineering and the generation component. 1t&amp;quot; you want to obtain flexibility for the generation component with respect to the possible diversity of answers, information should be available in cases of object centered questions (&amp;quot;What do you know about object xy ...&amp;quot;) as well as in comparable event oriented re-</Paragraph>
  </Section>
class="xml-element"></Paper>