<?xml version="1.0" standalone="yes"?>
<Paper uid="E99-1039">
  <Title>New Museums Site</Title>
  <Section position="3" start_page="0" end_page="261" type="metho">
    <SectionTitle>
2 Default Inheritance and YADU
</SectionTitle>
    <Paragraph position="0"> In this work, a default multiple orthogonal inheritance network is used to represent lexical information. Thus, with different subnetworks used to encode different kinds of linguistic knowledge, the idea is that linguistic regularities are encoded near the top of the network, while nodes further down the network are used to represent sub-regularities or exceptions. Such an approach to representing the lexicon has some advantages, like its ability to capture linguistic generalisations, conciseness, uniformity, ease of maintenance and modification, and modularity (Daelemans et al, 1992).</Paragraph>
    <Paragraph position="1"> This default multiple inheritance network is im- null Proceedings of EACL '99 plemented using YADU (Lascarides and Copestake, 1999), which is an order independent default unification operation on typed feature structures (TFS). YADU uses an extended definition of TFSS called typed default feature structures (TDFSs), to explicitly distinguish the non-default information from the default one, where a TDFS is composed by an indefeasible TFS (I), which contains the non-default information and a defeasible TFS (D), which contains the default information, with a '/' separating these two TFSS (I on the left-hand and D on the right-hand). As a consequence, during default unification non-default information can always be preserved and only consistent default information is incorporated into the defeasible TFS. Another important point is that default unification of two feature structures is deterministic, always returning a single value. Moreover, default specifications can be made to act as indefeasible information, using YADU's DefFill operation (Lascarides and Copestake, 1999), that has a TDFS as input and returns a TFS by incorporating all the default information into the indefeasible TFS, say at the interface between the lexicon and the rest of the system. YADU also provides the possibility of defining defaults that are going to persist outside the lexicon, with the p operator (Lascarides et al, 1996b), which was already shown to be significant, for example, for the interface between the lexicon and pragmatics, where lexically encoded semantic defaults can be overridden by discourse information (Lascarides et al, 1996a). Furthermore, YADU supports the definition of inequalities, which are used to override default reentrancies when no conflicting values are defined in the types involved (Lascarides and Copestake, 1999).</Paragraph>
    <Paragraph position="2"> YADU (~'~) can be informally defined as an operation that takes two TDFSS and produces a new one, whose indefeasible part is the result of unifying the indefeasible information defined in the input TDFSs; and the defeasible part is the result of combining the indefeasible part with the maximal set of compatible default elements, according to type specificity, as shown in the example below.</Paragraph>
    <Paragraph position="3"> Throughout this paper we adopt the abbreviatory notation from (Lascarides et al, 1996b) where Indefensible/De feasible is abbreviated to Indefeasible if Indefeasible = Defensible and T/Defeasible is abbreviated to ~Defensible.</Paragraph>
    <Paragraph position="5"> For a more detailed introduction to YADU see (Lascarides and Copestake, 1999).</Paragraph>
  </Section>
  <Section position="4" start_page="261" end_page="263" type="metho">
    <SectionTitle>
3 The proposed lexical network
</SectionTitle>
    <Paragraph position="0"> The proposed verbal subcategorisation hierarchy 1, which is based on the sketch by Pollard and Sag (1987) is shown in figure i. In this hierarchy, types are ordered according to the number and type of the subcategorisation arguments they specify. The subcategorisation arguments of a particular category 2 are defined in its SUBCAT feature as a difference-list. Thus, the verbal hierarchy starts with the intrans type, which by default specifies the need for exactly one argument, the NP subject, where e-list is a type that marks the end of the subcategorisation list: (1) intrans type: \[SuBCAT: &lt;HEAD: np, TAIL: /e-list&gt;\].</Paragraph>
    <Paragraph position="1"> Now all the attributes specified for the subcategorised subject NP in intrans are inherited by instances of this type and by its subtypes 3, namely, trans and intrans-control. However, since these types subcategorise for 2 arguments, they need to override the default of exactly one argument, specified by the e-list value for TAIL, and add an extra argument: an NP object for trans, and a predicative complement for intrans-control.</Paragraph>
    <Paragraph position="2"> In this way, the specification of the trans type is: (2) trans type: \[SUBCAT:&lt;TAIL: /e~list&gt;\].</Paragraph>
    <Paragraph position="3"> HEAD: rip, TAIL: TAIL: Similarly, the instances and subtypes of trans inherit from intrans all the attributes for the subject NP and from trans the attributes for the object NP, in addition to their own constraints. With the use of defaults there is no need for specifying a type like strict-trans, as defined in Pollard and Sag's hierarchy, since it contains exactly the same information as their trans type, except that the former specifies the SUBCAT For reasons of space we are only showing the parts of the lexical hierarchy that are relevant for this paper. 2Linguistic information is expressed using a simplified notation for the SUBGAT list, and for reasons of clarity, we are only showing categories in an atomic form, without the attributes defined.</Paragraph>
    <Paragraph position="4"> 3In this paper, we are not assuming the coverage condition, that any type in a hierarchy has to be resolvable to a most specific type.</Paragraph>
    <Paragraph position="5">  attribute as containing exactly two arguments: (3) Pollard and Sag's strict-trans type:  \[SUBCAT: &lt;HEAD: rip, TAIL: HEAD: np, TAIL: TAIL: e-list&gt;\], while the latter works as an intermediate type, where SUBGAT contains at least two arguments, as shown in (4), offering its subtypes the possibility of adding extra arguments.</Paragraph>
    <Paragraph position="6">  (4) Pollard and Sag's trans type:  \[SUBCAT: &lt;HEAD: rip, TAIL: HEAD: np&gt;\], Defaults automatically provide this possibility, by defeasibly marking the end of the subcategorisation list, which defines the number of arguments needed, avoiding the need for these redundant specifications, where the information contained in one lexical sign is repeated in others. Furthermore, these defaults are used to capture lexical generalisations, but outside the lexicon, we want them to act as indefeasible constraints; therefore, we apply the DefFill operation to these default specifications, except where marked as persistently default. In this way, a type like trans, after DefFill, has the consistent defaults incorporated and specifies, indefeasibly the need for exactly two arguments, as Pollard and Sag's strict-trans shown in (3): (5) trans type DefFilled: \[SUBCAT: &lt;HEAD: np, TAIL: HEAD: np, TAIL: TAIL: e-list&gt;\].</Paragraph>
    <Paragraph position="7"> Apart from supporting this kind of generalisation, defaults are also used to express sub-regularities, as, for example, in the case of super-equi and subject-control verbs, which are both exceptions to the general case specified by trans-equi. The type trans-equi encodes transitive-equi verbs by specifying that the predicative complement of the transitive verb is by default controlled by the object (e.g. The teacher persuaded the doctor to go): (6) trans-equi type: \[SUBCAT: &lt;TAIL: HEAD: np/\[\], TAIL: TAIL: HEAD: vp( INF, SUBCAT:&lt;HEAD: np/\[\] &gt;), TAIL: TAIL: TAIL: e-list&gt;\].</Paragraph>
    <Paragraph position="8"> For super-equi verbs, the predicative complements can be controlled by either the object or the subject. Therefore, the default object-control in the super-equi type, inherited from trans-equi, should be explicitly marked with the p operator to persist until discourse interpretation, as shown in (7), since all other features are made indefeasible prior to parsing.</Paragraph>
    <Paragraph position="9"> (7) super-equi type: \[SUBCAT: ~TAIL: HEAD: np/v \['~, TAIL: TAIL: HEAD: Yp( INF, SUBCAT: ~HEAD: np/v \[\] &gt;) &gt;\].</Paragraph>
    <Paragraph position="10"> This default would only survive in the absence of conflicting discourse information (as in e.g.: They needed someone with medical training. So, the teacher asked the doctor to go (since she had none), which is object-controlled). Otherwise, if there is conflicting information, this default is rejected (as in e.g.: They needed someone with teaching experience. So, the teacher asked the doctor (to be allowed) to go, where the control is by the subject). A description of the precise mechanism to do this can be found in (Lascarides et al, 1996a). Transitive subject-control verbs follow the pattern specified by trans-equi, but contrary to this pattern, it is the subject that controls the predicative complement and not the object (e.g. The teacher promised to go): (8) subject-control type: \[SUBCAT: &lt;HEAD: np\[\], TAIL: HEAD: np/ff\], TAIL: TAIL: HEAD: vp( INF, SUBCAT: &lt;HEAD:</Paragraph>
    <Paragraph position="12"> In this case, the constraint on subject-control specifies that the coindexation is determined by the subject, and as it does not conflict with the default coindexation by the object-control, inequalities (~) are used to remove the default value.</Paragraph>
    <Paragraph position="13">  Proceedings of EACL '99 As a result of using default inheritance to represent information about verbal subcategorisation, it is possible to obtain a highly structured and succinct hierarchy. In comparison with the hierarchy defined by Pollard and Sag (1987), this one avoids the need of redundant specifications and associated type declarations, like the strict-trans type, which are needed in a monotonic encoding.</Paragraph>
    <Paragraph position="14"> In this way, while Pollard and Sag's hierarchy is defined using 23 nodes, this is defined using only 19 nodes, and by defining 2 more nodes, it is possible to specify subject-control and super-equi types. By avoiding this redundancy, there is a real gain in conciseness, with the resulting hierarchy extending the information defined in Pollard and Sag's, with the addition of sub-regularities, in a more compact encoding.</Paragraph>
  </Section>
class="xml-element"></Paper>