File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/92/j92-3002_metho.xml
Size: 70,713 bytes
Last Modified: 2025-10-06 14:13:11
<?xml version="1.0" standalone="yes"?> <Paper uid="J92-3002"> <Title>Hewlett-Packard Laboratories John Nerbonne t Deutsches Forschungszentrum ffir Kfinstliche Intelligenz</Title> <Section position="3" start_page="270" end_page="273" type="metho"> <SectionTitle> 2. Grammatical Theory </SectionTitle> <Paragraph position="0"> The phenomena involved in the analysis of the easy adjective class illustrate (obligatory and optional) subcategorization, control, long-distance dependence, optional modification, and specification (the last in its interaction with adjectival gradation with too and enough). As such, it represents an excellent demonstration vehicle for the lexical demands of grammatical analysis. Our analysis is formulated within head-driven phrase structure grammar (HPSG), the grammatical theory developed by Carl Pollard and Ivan Sag during the mid and late 1980s. See Pollard (1984; 1985; 1988; 1989) and Pollard and Sag (1987; 1988; 1991). As the lengthy list of publications might suggest, this grammatical theory is well enough documented so that we may restrict our Ross (1967), Postal (1971), Bresnan (1971), Chomsky (1973), Lasnik and Fiengo (1974), Jackendoff (1975), Chomsky (1977), Fodor (1978), Brame (1979), Nanni (1980), Schachter (1981), Jacobson (1982, pp. 221-223), Sag (1982), Maling and Zaenen (1982, pp. 253--254), Kaplan and Bresnan (1982, pp. 255-263), Culicover and Wilkins (1984), Jacobson (1984), Gazdar et al. (1985, pp. 150-152), Jacobson (1990), Jones (1990), Bayer (1990), and Hukari and Levine (1991). None of these works has attempted a thorough descriptive analysis of the range of data we address here, though we are of course indebted to these studies for much of the data and many of the generalizations we seek to express. In particular, our account is consistent with the brief generalized phrase structure grammar (GPSG) analysis of these adjectives given in Gazdar et al. (1985, pp. 150-152) though we embrace a larger range of data and extend the analysis to related nouns, a topic rarely discussed since its introduction by Lasnik and Fiengo (1974).</Paragraph> <Paragraph position="1"> Computational Linguistics Volume 18, Number 3 remarks here to the distinctive characteristics of the assumptions used here. We assume familiarity with feature-based grammars and basic familiarity with HPSG as well.</Paragraph> <Paragraph position="2"> In all linguistic theories there is a division of labor between grammatical rules and the lexicon, and this concerns the amount of information contained in each. At the rule-based extreme lie non-feature-based context-free grammars, where the lexicon merely links lexical items to nonterminals; in these grammars it is indeed customary to view the lexicon as a set of unary rules. The grammatical rules thus effectively encode all linguistic information. At the lexical extreme we find feature-based categorial grammars, which allow function argument application as the only grammatical rule. Here the lexicon bears the burden of encoding linguistic information, and the contribution of rules is marginal. We emphasize that HPSG is found very close to the lexical extreme, because this highlights the significance of the present work--HPSG is a framework whose lexical demands are very nearly maximal.</Paragraph> <Paragraph position="3"> Subcategorization information is lexically based in HPSG, much as it is in Categorial Grammar (Bach 1988). Grammatical heads specify the syntactic and semantic restrictions they impose on their complements and adjuncts. For example, verbs and verb phrases bear a feature SUBCAT whose content is a (perhaps ordered) set of feature structures representing their unsatisfied subcategorization requirements. Thus the feature structures associated with transitive verbs include the information: \[subcat: /\[ NPe:acc \]' \[NPcase: nom \]/ \] (where hip abbreviates a substantial feature structure.) Applied to adjectival VP complementation, this treatment of subcategorization leads naturally to the postulation of adjectives that subcategorize for VPs, etc. (details follow).</Paragraph> <Paragraph position="4"> The significance of subcategorization information is that it represents a (perhaps ordered) set of grammatical categories with which a subcategorizer combines in forming larger phrases. When a subcategorizer combines with a subcategorized element, the resultant phrase no longer bears the subcategorization specification--it has been discharged. Compare Pollard and Sag (1987, p. 71) for a formulation of the HPSG subcategorization principle.</Paragraph> <Paragraph position="5"> We shall in general present subcategorization specifications in a slightly different way from that above, i.e., not as a single feature whose value is a list, but rather as a collection of complement features with category values. Compare Borsley (1987) for a development of this approach, which we shall not attempt to justify here. We will therefore reorganize the information above in the following way: subject: \[ NP case: nom degbject: \[ NP \] case: acc We choose this representation here only because we find the keywording of grammatical functions, subject, etc., more perspicuous than an encoding in terms of list positions, but nothing in the analysis hinges on the one or the other representation.</Paragraph> <Paragraph position="6"> We shall furthermore allow that subcategorized elements be either obligatorily sub-categorized or optionally subcategorized. Optionally subcategorized elements need not be discharged from subcategorization specifications. (This necessitates an obvious change to the principle that subcategorization must be satisfied in independent Dan Flickinger and John Nerbonne Inheritance and Complementation utterances.) In case an element is not discharged, something must be said about its semantics. Here we borrow an idea from Situation Theory, and specify that unsaturated predicate argument structures (or infons; see Devlin 1991) may hold when there is some way of filling out the unfilled argument positions so that the result holds. This has the effect of existentially quantifying over unfilled argument positions. Linguistically, there are many other ways in which arguments may be omitted (cf. Fillmore 1985), but this seems to suffice for the adjectives under examination here.</Paragraph> <Paragraph position="7"> Control and modification, the latter being the relation between an adjunct and a head, are both lexically realized in the case of the easy adjectives. We regard there as being a control relation between for Smith and toget in complex adjectivals such as easy for Smith to get (cf. Gazdar et al. 1985: 83ff). Modification plays a role when complex adjectivals appear in construction with nominal heads, as in easy job for Smith to get.</Paragraph> <Paragraph position="8"> These are common assumptions in the analyses of control and modification.</Paragraph> <Paragraph position="9"> Long-distance dependence is treated in HPSG in much the same way it was treated in GPSG (cf. Gazdar et al. 1985), and we assume basic familiarity with this type of analysis. We recall that the site of a missing element in a &quot;gappy&quot; constituent bears a feature SLASH whose value is a specification of the expected material. The SLASH specification is propagated by general principles (which we shall not elucidate) to the higher level constituents, until it is matched by a &quot;filler&quot; or a subcategorizing element. When the gappy constituent is adjoined to a filler or subcategorizing element, the result no longer bears the SLASH value.</Paragraph> <Paragraph position="10"> Important for our purposes is the possibility of a lexical entry specifying that a dependent may contain a gap. (Cf. Gazdar et al. 1985, pp. 150-153 for the first mention of this suggestion.) We shall exploit this in the analysis of several word classes below, viz., the ones that subcategorize for a VP with an NP in SLASH. It is unusual to find a subcategorization specification for SLASH, but not unique: comparatives likewise subcategorize for gappy complements, as in seen in examples such as taller than it is A wide. We shall require lexical specifications that lead to feature structures of the following form: stem: easy sere: easy( ~\], ^ ~\] ) syn.loc.subcat: sub j: \[ syn: NP-nom \] sem: ~-~ pp-for: \[ sem:\[-~ \] xcomp: syn: VP-inf sem: ~\] slash: \[ sem:\[~\] \] The tag \[~ in the diagram above shows that the semantics of the SLASH value and the adjectival subject semantics have been identified. Thus, once a VP/NP has combined with this adjective, the semantic contribution of the SLASH element is assumed by the subject. Figure 1 shows an analysis tree for an example containing a long-distance dependency.</Paragraph> <Paragraph position="11"> Complex adjectivals such as easy subcategorize for a complement VP containing a &quot;slashed&quot; NP, i.e., a VP missing an NP (whose expected position may be arbitrarily deep).</Paragraph> <Paragraph position="12"> The variety of linguistic phenomena exemplified in the easy class of adjectives guarantees that it is a demanding testing ground for theories of lexical representation. 3</Paragraph> </Section> <Section position="4" start_page="273" end_page="275" type="metho"> <SectionTitle> 3. Adjectival VP Complementation </SectionTitle> <Paragraph position="0"> We assume familiarity with the mechanisms of lexical inheritance and lexical rules in the analysis to follow, but we provide an overview of these mechanisms for lexical 3 It is also worth mentioning that HPSG has also been the subject of intensive implementation activity during the past several years; we know of implementations at Hewlett-Packard Laboratories, The German AI Center (DFKI), Stanford University, Carnegie Mellon University, The Ohio State University, Simon Fraser University, University of Edinburgh, ICOT, University of Stuttgart, the IBM LILOG project in Stuttgart, and ATR. We may therefore safely refer the reader to documentations of those implementations, even if these are less generally available than the theoretical literature: Proudian and Pollard (1985), Nerbonne and Proudian (1987), Franz (1990), Emele and Zajac (1990), and Carpenter, Pollard, and Franz (1991).</Paragraph> <Paragraph position="1"> Dan Flickinger and John Nerbonne Inheritance and Complernentation representation in Appendix A. The fundamental data we shall be concerned with are repeated in (2): Bill is easy to talk to.</Paragraph> <Paragraph position="2"> It is easy to talk to Bill.</Paragraph> <Paragraph position="3"> Bill is easy for Mary to talk to.</Paragraph> <Paragraph position="4"> It is easy for Mary to talk to Bill.</Paragraph> <Paragraph position="5"> that show this same distribution include the following: depressing great nice difficult hard painful exhausting important tiresome fun impossible terrible good impressive tough Given pairs like (2a,b) and (2c, d), two dusters of properties begin to suggest themselves as part of the definitions of the relevant lexical entries. The first of these clusters we will associate with the class of words containing lexical entries for the easy of (2a,c) and its counterparts in (3), a class we term SLASH-EASY. The other cluster of properties we associate with a second class termed IT-EASY, containing the lexical entries for the variant of easy in (2b, d) and its counterparts in (3). We begin by simply identifying the relevant properties in each of these two classes, supported by examples as necessary; then we provide motivation for factoring these properties into several word classes linked by inheritance.</Paragraph> <Paragraph position="6"> Adjectives in the IT-EASY class have two obligatory complements, an NP subject and a verbal complement; in addition they have one optional complement, a PP headed by the preposition for. As seen in (4), the verbal complement can be either infinitival or gerundive, and (5) shows that this complement can be a VP even with a PP-for present, or an infinitival S, again with or without the optional PP-for complement. The subject NP must be the expletive it.</Paragraph> <Paragraph position="7"> (4) a. It was great working for Bill.</Paragraph> <Paragraph position="8"> b. It was great to work for Bill.</Paragraph> <Paragraph position="9"> (5) a. It's easiest for the dogs to feed them at noon.</Paragraph> <Paragraph position="10"> b. For the dogs, it's easiest to feed them at noon.</Paragraph> <Paragraph position="11"> c. It's easiest for the dogs to be chained up all day.</Paragraph> <Paragraph position="12"> d. *For the dogs, it's easiest to be chained up all day.</Paragraph> <Paragraph position="13"> e. It's easiest for me for the dogs to be chained up all day.</Paragraph> <Paragraph position="14"> f. For me, it's easiest for the dogs to be chained up all day.</Paragraph> <Paragraph position="15"> Examples (5e, f) demonstrate that not only VP complementation but also S complementation, is involved in easy subcategorization. Note that S complementation never requires a controller, and that the PP phrase in such structures is mobile (5f). In addition to the conclusion that a variety of complementation schemes are used with easy, the data above also demonstrate that the exact specification of the controller (the understood subject of the infinitival VP) is nontrivial. Example (5a) demonstrates that the PP-FOR complement need not control the VP, and (5b) suggests that noncontrolling PPs are more mobile than controllers (5d).</Paragraph> <Paragraph position="16"> Computational Linguistics Volume 18, Number 3 We accommodate these facts semantically by aUowing that easy and similar adjectives denote two-place relations between individuals and states of affairs. The relation holds between the pair, roughly, when it is easy (or convenient) for the individual when the state of affairs obtains. Examples (5e,f) show that the individual involved in the easy relation need not be involved in the state of affairs, i.e., that there is no necessary semantic control involved in this relation. 4 The control facts are clear enough: when this easy is combined with an S, there is no semantic control; and when it is combined with a VP, there is no grammatically specified controller of the VP--although there may be pragmatic inference about the understood subject.</Paragraph> <Paragraph position="17"> Adjectives in the SLASH-EASY class also have two obligatory complements, an NP subject and a verbal complement, as well as an optional PP-for complement. In contrast to the first class, this class specifies that the subject is a normal (nonexpletive) NP, and that the verbal complement must contain an NP gap. Moreover, this verbal complement must be infinitival, not gerundive, as seen in (6), and must be a VP, not an S, as shown in (7). 5 (6) a. Bill was great to work for.</Paragraph> <Paragraph position="18"> b. *Bill was great working for.</Paragraph> <Paragraph position="19"> (7) a. For me, Bill was easy to talk to.</Paragraph> <Paragraph position="20"> b. *Bill was easy for me for Mary to talk to.</Paragraph> <Paragraph position="21"> In the word class hierarchy we assume, sketched in Appendix A, there is a word class CONTROL, which introduces a verbal complement subcategorization, and which serves as the superclass from which both of the classes IT-EASY and SLASH-EASY inherit. However, neither of these classes is an immediate subclass of CONTROL; we draw on the data provided in (8) and (9) below to motivate two intermediate word classes that will stand between CONTROL and these two in the hierarchy. The English lexicon contains two more groups of adjectives that have much in common with the two variants of easy introduced above, but must be kept distinct. Lasnik and Fiengo (1974:535) identified a set of adjectives including pretty and melodious, illustrated in (8).</Paragraph> <Paragraph position="22"> (8) a. Disneyland is pretty to look at.</Paragraph> <Paragraph position="23"> b. Sonatas are melodious to listen to.</Paragraph> <Paragraph position="24"> c. *It is pretty to look at Disneyland.</Paragraph> <Paragraph position="25"> d. *It is melodious to listen to sonatas.</Paragraph> <Paragraph position="26"> e. ?Disneyland is pretty for children to look at.</Paragraph> <Paragraph position="27"> f. ?Sonatas are melodious for serious musicians to listen to.</Paragraph> <Paragraph position="28"> 4 There is an interesting pragmatic problem lurking in the control specifications involved here. If one specifies the control relationships exactly, then one needs to postulate systematic structural ambiguity in examples such as (5c), where the sequence of PP and VP may or may not be analyzed as an constituent. This seems plausible, but then we would like to have a pragmatic account of why there is normally no distinction, i.e., why the control relationship is inferred, or, equivalently for all intents and purposes, why the S reading is so strongly preferred.</Paragraph> </Section> <Section position="5" start_page="275" end_page="277" type="metho"> <SectionTitle> 5 Hukari and Levine (1991) note in passing that there is a group of closely related adjectives such as </SectionTitle> <Paragraph position="0"> worth that do take a gerundive complement instead of the usual infinitival complement, as in That article is not worth looking at. The extension of our analysis to worth is straightforward, but not given here.</Paragraph> <Paragraph position="1"> Dan Flickinger and John Nerbonne Inheritance and Complernentation Members of this class of adjectives share much in common with the SLASH-EASY adjectives, but have two significant differences: first, as shown by (8c,d), they do not have a corresponding entry with an expletive it subject, and second, they assign a real thematic role to their subjects. That is, (8a) entails that Disneyland is pretty, while (la) does not entail that Bill is easy. The two-place relation suggested above for IT-EASY and SLASH-EASY adjectives could not account for the validity of this inference, since the subject of the adjective plays no direct role in the relation whatsoever. A distinct semantic relation is called for here, one in which the subject does play a role (which effectively makes this class a kind of EQUI adjective in contrast to the raising easy).</Paragraph> <Paragraph position="2"> It also appears that these adjectives do not permit the optional PP-for complement licensed by easy in (lc), though judgments are less clear. In order to express these differences, we introduce a class SLASH-COME which will include the entries for pretty adjectives, and which will also serve as the class from which SLASH-EASY inherits. 6 Similarly, English has a set of adjectives that have much in common with the IT-EASY adjectives of (lb, d), but with no counterparts of the SLASH-EASY type.</Paragraph> <Paragraph position="3"> (9) a. It is possible to talk to Bill only at breakfast.</Paragraph> <Paragraph position="4"> b. It is unnecessary to fire Bill.</Paragraph> <Paragraph position="5"> c. *Bill is possible to talk to only at breakfast.</Paragraph> <Paragraph position="6"> d. *Bill is unnecessary to fire.</Paragraph> <Paragraph position="7"> The second principal difference between adjectives such as possible and those of the IT-EASY class is that the former do not permit an optional PP-for phrase complement; they do allow the verbal complement to be either a VP or an S (containing a PP-for subject), but (10) shows that if a PP-for is present, it must be contained within the complement.</Paragraph> <Paragraph position="8"> (lO) a. It is unnecessary for Mary to fire Bill. (M firing B) b. *For Mary, it is unnecessary to fire Bill. (M firing B) c. *It is unnecessary for Mary for you to fire Bill.</Paragraph> <Paragraph position="9"> Again, we express the distinction between the set of adjectives like possible and the IT-EASY adjectives by introducing a fourth class IT-SUBJ parallel to SLASH-COME 7 These four class definitions, together with one supporting class, are given in (1116), with the Superclasses attribute showing the relevant inheritance relations. Computational Linguistics Volume 18, Number 3 The disjunctive specification (Complete + -) overrides the default (Complete -) specified in the CONTROL class, and means that the verbal complement may be either The SLASH feature on the XComp specifies that the VP must contain a gap for a normal (non-expletive) noun phrase, which is accusative case and which is not predicative. This nonpredicative specification serves to exclude examples like *Bill is difficult to become (assuming the complement of become is predicative), since the gap for that complement would fail to satisfy the restriction on SLASH given in (12). The SLASH specification furthermore notes that the SLASH semantic value is identical to that of Subject-Semantics. As was explained in Section 2 above, this is the form a lexical specification of semantic coindexing takes.</Paragraph> <Paragraph position="10"> The controller of the controlled complement is specified through the attribute XComp-Subj-Semantics; for example, in CONTROL, this attribute has the value Subject-Semantics, since subjects are default controllers. But the complements of SLASH-COMP are not grammatically controlled (cf. (8e,f)), a fact that requires an overwriting specification. The semantic variable x is used here because it will not represent the semantics of any grammatical complement, which ensures that no grammatical control is effected (see examples (9a,b)). This is an example of a subregularity appearing within an exceptional specification.</Paragraph> <Paragraph position="11"> The classes for the two variants of easy adjectives we have discussed have one cluster of properties in common: they both license the optional PP-for phrase seen in preceding examples. To further reduce redundancy, we define in (13) the class FOR-EXPERIENCER, from which the two classes in (14-15) also inherit.</Paragraph> <Paragraph position="12"> As expected, the IT-EASY class eases one restriction on the verbal complement; note too that no controller is specified, in keeping with remarks on (5). On the other hand, the SLASH-EASY class blocks inheritance of the subject's thematic role assignment (the default value having been specified in the INCOMPLETE class from which CONTROL inherits), and alters the control relationship (inherited from SLASH-COMP and ultimately from CONTROL) so that the PP-For phrase rather than the subject of easy is interpreted as the subject of the VP complement. These are two further examples of the way in which default overwriting is employed; note that the latter represents a subregularity within a subregularity (cf. SLASH-COMP).</Paragraph> <Paragraph position="13"> With reasonable assumptions about the definitions of other relevant classes in the hierarchy, along with an explicit definition of the class ADJECTIVE, provided here for clarity in (16-17), we can introduce the (sparse) lexical entries for the two variants of easy employed in (la,b), as given in (17,18): The structure of word classes directly involved in the definition of complex adjectival lexical entries.</Paragraph> <Paragraph position="14"> Pairs of sparse lexical entries like those in (17,18) are related by a lexical rule we label LR-EASY, which simply states that for each member of the class IT-EASY there exists a corresponding lexical entry belonging to the class SLASH-EASY, with everything but the Superclasses property identical in the two (sparse) entries.</Paragraph> <Paragraph position="16"> Once each of (17) and (18) are fleshed out to include all of their inherited properties, they will of course be quite distinct, as needed to ensure the differences in distribution that we have described. Figure 2 summarizes the inheritance relationships thus far.</Paragraph> </Section> <Section position="6" start_page="277" end_page="282" type="metho"> <SectionTitle> 4. An Example Analysis </SectionTitle> <Paragraph position="0"> The purpose of this section is primarily illustrative--we would like to demonstrate the effect of the lexical specifications suggested on more familiar elements of grammatical analysis, viz. phrases, parse trees, and predicate logic representations.</Paragraph> <Paragraph position="1"> The semantics of the easy-SLASH construction, which treats easy as a relation between an individual and a state of affairs, is treated as a normal case of lexically in- null Dan Flickinger and John Nerbonne Inheritance and Complementation herited semantics, i.e., one in which the relation denoted has an argument place for the denotations of each of the role-playing complements, in this case the PP-FOR phrase and the XCOMP. This class of adjectives also has a SUBJECT among its complements, but it bears no role (as word class SLASH-EASY specifies), because this is a raising construction. For this reason, there is no argument place reserved in the semantics of easy-SLASH adjectives for the subject's denotation. To conserve space in the diagrams below, relations will be specified not using the keyword coding shown in word class and lexical entry specifications (above), but rather in the more familiar order coding. In order to make not only the semantics but also the syntax somewhat clearer in its intended effect, we include here somewhat elaborate analytical sketches of the complex adjectival phrase easy to get Mary to hire in (19): (19) Tom is easy to get Mary to hire.</Paragraph> <Paragraph position="2"> To begin, we note that the sparse lexical entry for the SLASH-EASY version of easy may be filled out to a much richer structure if inherited properties are noted explicitly. The features noted above were specified by the lexical entry together with the classes ADJECTIVE, SLASH-EASY, SLASH-COMP, FOR-EXPERIENCER, and CONTROL. Further subject properties would be inherited from INCOMPLETE, but for brevity these are not listed. (Of course many other properties, including e.g., gradation properties and the applicability of lexical rules, have likewise been suppressed in the interest of clarity in presentation.) This lexical description translates fairly directly (with some further simplifications and abbreviations) into a feature structure of the sort used by HPSG grammars.</Paragraph> <Paragraph position="3"> Computational Linguistics Volume 18, Number 3 stem: easy syn.loc.head: Adj sem: easy( \[~,^ \[~) syn.loc.subcat: sub j: \[ syn: NP-nom \] sem: ~\] pp_for: \[ sY n: PP-fdegr \] sem: \[\] syn: VP-inf sere: ~\] xcomp: subject.sem: \[~ slash: \[ 2eYnm:~ -acc \] We would like to draw attention to two semantic coindexings in the structure, which are lexically specified and which simplify subsequent (grammatical) processing. The coindexing of the xcomp's subject with the pp-for is effected in the SLASH-EASY word class, and the semantic coindexing seen above is just a consequence of that. The coindexing of the xcomp's slash's semantics value with the subject's semantics, on the other hand, derives ultimately from SLASH-COMP.</Paragraph> <Paragraph position="4"> In Figure 3 we examine the combination of a token from this class of easy adjectives and a VP/NP. The very sparse specification of the mother phrase's features is, in fact, solely for purposes of legibility--all of the information specified on the mother node may be derived from general HPSG principles, so that nothing is specified, e.g., on the rule that licenses head-complement combinations. The fact that the semantics attribute is identified with the subcategorizer's semantics follows from the HPSG Semantics Principle, which states that the semantics of a phrasal node is always to be identified with the semantics of a head in a head complement combination. The fact that the slash value of the mother structure is empty follows from the Binding Inheritance Principle, which states that slash values are collected going up a tree--unless a head subcategorizes for an element containing a slash value, in which case the slash satisfies the subcategorization requirement. The identification of the feature structure labeled V~, which is just the representation of the phrasal node dominating to get Mary to hire, with one of the adjective's subcategorization specifications, that labeled ~\], is just a condition for the applicability of the head-complement rule, not an additional specification. Of course, the phrasal node is massively underspecified here, but the suppressed information is predictable, not merely hidden.</Paragraph> <Paragraph position="5"> This is an intriguing aspect of HPSG, but we dwell on it here for self-serving purposes. If the properties of the phrasal combination of this fairly intricate syntactic structure require no further comment, that is largely because the lexicon has provided a wealth of richly structured representation. This would hardly be feasible in the absence of efficient and sophisticated lexical representation mechanisms. Dan Flickinger and John Nerbonne Inheritance and Complementation Adj easy slash: 0 AdjP to get Mary to hire t * stem: easy sem: \[~ easy( \[-i-\], ^ ~\] ) s bj: \[som\[\] \] pp or: \[ sem:\[ \] comps: subj: \[ som:\[ \] xcomp:\[\] sem:\[\] L slash.sem: \[\] sere: get(x,m, 1 \[\] &quot; hire(m, ~\])) slash.sere: \[\] Figure 3 The combination of complex adjective and slashed VP complement. To complete this illustration, we spell out the effects of unification on the structure above in Figure 4. Note in particular that because the slash semantics on the VP phrase is identified with the slash semantics on the subcategorized-for VP, which in turn is identified with the semantics of the subject for easy, the resultant phrase will bind its subject to the deeply embedded object argument position of the verb hire. This takes place even though the subject plays no role in the easy relation itself. This is exactly what is wanted semantically of a raising construction.</Paragraph> </Section> <Section position="7" start_page="282" end_page="293" type="metho"> <SectionTitle> 5. Extensions and Lexical Maintenance </SectionTitle> <Paragraph position="0"> The structured lexicon aims ideally at a redundancy-free specification of all lexical properties, and indeed, it achieves this largely through the use of inheritance. While we do see scientific parsimony as an end in itself, we see two further advantages in the employment of the structured lexicon, one scientific and one practical. The scientific advantage of the structured lexicon is that it identifies significant classes in the language. In a feature system with approximately 30 atomic features (including semantics), each of which ranges over approximately 10 values, it is certainly striking that we never see the need to distinguish 1030 classes of items. In fact we distinguish approximately 300 lexical classes in HP-NL, a large system with very broad grammatical coverage (see Nerbonne and Proudian 1987).</Paragraph> <Paragraph position="1"> Computational Linguistics Volume 18, Number 3 stem: easy sem.logic: \[\] easy( z, ^ get (z, m, &quot; hire(m, T,~\] U ~\]) ) comps:\[ subj: \[sem:\[~\]U\[~\] \] \] to get Mary to hire t Figure 4 The result of combining complex adjective and slashed VP complement. Note that the subject of easy is still semantically coindexed with the missing VP object. But the practical advantage of the structured lexicon may ultimately also be of scientific value, and that is because a structured lexicon is more easily maintained and extended than a nonstructured one. This advantage derives immediately from the characteristic that lexical properties are normally specified only once. Modifications tend then to be minimal and extensions less frightening. The ultimate scientific benefit this may bring derives from the fact that it is then easier in systems with structured lexicons to experiment with grammatical description.</Paragraph> <Paragraph position="2"> The following section is an attempt to buttress the claim that structured lexicons are easily extended. We examine therefore extensions to the analysis above of adjectives that govern VP complements--to nouns with similar subcategorizations, to the adjectival specifiers too and enough, and to adjectives that govern S complements rather than VP complements.</Paragraph> <Section position="1" start_page="283" end_page="287" type="sub_section"> <SectionTitle> 5.1 Pleasure Nouns </SectionTitle> <Paragraph position="0"> Adjectives like easy have been the most widely studied group of lexical types that populate the classes introduced in the analysis above, but they do not have exclusive claim to those classes. Lasnik and Fiengo (1974) observed that the English lexicon also contains a group of nouns with similar properties, as illustrated in (20-21), (20) a. Nureyev is a pleasure to watch.</Paragraph> <Paragraph position="1"> b. This course is a breeze to pass.</Paragraph> <Paragraph position="2"> c. Venice is a delight to visit.</Paragraph> <Paragraph position="3"> (21) a. It is a pleasure to watch Nureyev.</Paragraph> <Paragraph position="4"> b. It is a breeze to pass this course.</Paragraph> <Paragraph position="5"> c. It is a delight to visit Venice.</Paragraph> <Paragraph position="6"> Like the adjectives discussed above, nouns such as pleasure have two variants, one that appears with an ordinary NP subject and an infinitival complement containing an Dan Flickinger and John Nerbonne Inheritance and Complementation NP gap; and one that selects an expletive it subject and an infinitival complement with no gap. Given the word class definitions developed on the strength of the adjectival examples, an obvious analysis of the nominal examples suggests itself: pleasure, like pleasant, has one lexical entry belonging to the SLASH-EASY class, and a second entry that inherits from the IT-EASY class. The (sparse) descriptions of both entries are given in (22-23), parallel to those for easy given in (17-18) above, the salient difference being that the noun entries inherit from the class Common-Noun where the adjective entries inherited from the Adjective class. 8 Having declared nouns like pleasure to have entries that are members of SLASH-EASY and IT-EASY, nothing more needs to be said in order to capture the syntactic relationship between these two forms of pleasure. The lexical rule we proposed earlier to link pairs of adjectives like the two variants of easy is defined as a regularity holding between the two classes SLASH-EASY and IT-EASY, making no mention of the class ADJECTIVE in its formulation. Hence it also serves to link the pair of noun entries in (22-23).</Paragraph> <Paragraph position="7"> Some further explanation needs to be provided about the semantics of this class of nouns, since the nouns do seem semantically anomalous even if we shall maintain that all of the apparent anomaly ultimately stems from their having a subject--and thus being available for control (by be and other raising verbs). In general a common noun is interpreted as a relation between a theme argument and the denotation of its complements, if there are any. For example, friend is interpreted as a relation between a theme argument and the denotation of the complement PP-OF phrase. We refer to the theme argument of the relation denoted by the common noun as its denotation. An apparent peculiarity of nouns such as pleasure is that there appears to be no denotation of the noun in the usual sense, e.g., in (20a). At issue is whether there is any theme argument position for the &quot;pleasure&quot; in the relation denoted by pleasure. That is, does pleasure denote the same two-place relation between individuals and states of affairs that pleasant does, or is there a third argument position in pleasure that is occupied by an (abstract) &quot;pleasure&quot; individual? The suspicion that no denotation is involved likely stems from our intuition that we do not seem to refer to an object that is a pleasure in uttering either (20a) or (21a), Computational Linguistics Volume 18, Number 3 Now this suggests that the noun (phrase) is used predicatively, much as many noun phrases are after the verb be. Compare Tom is a linguist.</Paragraph> <Paragraph position="8"> This does not help a great deal, however. Even though the analysis of predicative NPs is an old topic semantically (cf. the definition of be in Montague 1973, p. 261), there has been essentially no successful attempt to treat predicative nouns as if they had no denotation. Any attempt to do so seems to run afoul of the standard (if limited) determination and adjectival modification found in phrases such as no great pleasure to watch; at least such examples point out the inevitable duplication a semantic analysis would incur if predicative nominals had no denotation.</Paragraph> <Paragraph position="9"> We therefore interpret pleasure as a three-place relation pleasure (theme : e~ for : x~ soa : s) which obtains just in case e is the pleasure x has in case s. It should of course also turn out that this relation for some e holds iff pleasant (for : x~ soa : y), but we will not be concerned with showing that here. e provides a denotation that is subject to determination (no) and (intersective) adjectival modification (great). Under this analysis, a pleasure to watch and no pleasure to watch denote quantifiers, i.e., in each case a set of properties of pleasures (e's from above). Of course, a quantifier does not by itself represent a proposition, something that could be true or false---for that it must be paired with a property. In these cases, the relevant property is always the universal (existence) property; i.e., utterances of sentences such as (20a) are true just in case there is a pleasure of the relevant kind (and mutatis mutandis for the negative existentials). We therefore postulate that the predicate be in these sentences denotes the universal property. 9 What is striking about this proposal is that it assigns the common noun pleasure exactly the semantics the general scheme predicts--a relation between a theme and the denotations of other complements. For this reason, the word classes for pleasure nouns make no special stipulations about semantics.</Paragraph> <Paragraph position="10"> We therefore derive feature structures such as the following, which are used in the syntax and semantics processing of the word pleasure. The first structure represents the member of the SLASH-EASY class, and the second the member of the IT-EASY class. (We have simplified the structures to highlight the semantically relevant parts.) stem: &quot;pleasure&quot; sem.logic: pleasure (e, ~-~ \[\] ) subcat: subj: I sere:\[\] \] pp-for: I sem:\[\] \] \[ subj: \[\[sem: \[\] \]1 xcomp: logic: 2~ sem: slash: 9 In fact, we do not stipulate a peculiar semantics for the raising verbs (such as be) that are involved here. Instead, we allow be to denote the identity relation, which holds of a single argument just in case there is some way of filling in the missing argument--i.e., in case the first exists. This follows from the general treatment of unsaturated relations in Situation Theory (cf. Section 2 under subcategorization). Note, however, the one exceptional aspect, i.e., that the subject of the verb be is not linked to any argument position in the relation denoted by the controlled complement (in this case, pleasure). Dan Flickinger and John Nerbonne Inheritance and Complementation stem: &quot;pleasure&quot; sere.logic: pleasure (e, ~-~,^ ~-\] ) comps: subj: NP-it pp-for: \[ sem: \[~ \[\] xcomp:Isubject: sem:~\] \] \] sere: \[~ On the other hand, the noun classes are exceptional in that the nouns involved have subjects--a property they inherent finally from INCOMPLETE, in the one case through CONTROL, IT-SUBJ, and IT-EASY; and in the other from CONTROL, SLASH-COMP, and SLASH-EASY. It is this property, shared by the NPs they give rise to, that explains (i) their ability to be controlled, e.g., by the verb be--only unsaturated phrases are subject to control; (ii) their inability to function in normal NPs, e.g., in the subject position of any intransitive verb; and finally (iii) the fact that they can stand in construction with the main verb be without being asserted to be identical to its subject. We turn now to further points on the syntax of the pleasure nouns. The two definitions of entries for &quot;pleasure&quot; also predict the grammaticality judgments seen in (24), analogous to the examples given above for adjectives, and based on the definitions given for the IT-EASY and SLASH-EASY word classes. 1deg (24) a. Nureyev is a pleasure for us to watch.</Paragraph> <Paragraph position="11"> b. It is a pleasure for us to watch Nureyev.</Paragraph> <Paragraph position="12"> c. For us, Nureyev is a real pleasure to watch.</Paragraph> <Paragraph position="13"> d. *For us, Nureyev is a real pleasure for our parents to watch.</Paragraph> <Paragraph position="14"> e. For us, it is a real pleasure for our parents to watch Nureyev.</Paragraph> <Paragraph position="15"> f. It is a real pleasure for us for our parents to watch Nureyev.</Paragraph> <Paragraph position="16"> g. *Nureyev is a pleasure watching.</Paragraph> <Paragraph position="17"> h. It is a pleasure watching Nureyev.</Paragraph> <Paragraph position="18"> 10 Nothing we have said so far captures the fact that some pairs of members of these two classes, such as &quot;pleasant&quot; and &quot;pleasure,&quot; are morphologically related. We do not offer here a proposal for capturing nonproductive regularities of this kind, though some extension of the lexical rule mechanism might serve, an extension that would depend heavily on the ability to specify negative exceptions to lexical rules, given examples like the following.</Paragraph> <Paragraph position="19"> (i) It is difficult to hire Bill.</Paragraph> <Paragraph position="20"> (ii) *It is a difficulty to hire Bill.</Paragraph> <Paragraph position="21"> (iii) *Bill is a difficulty to hire.</Paragraph> <Paragraph position="22"> (iv) It is impossible to work with Bill.</Paragraph> <Paragraph position="23"> (v) *It is an impossibility to work with Bill (vi) *Bill is an impossibility to work with.</Paragraph> <Paragraph position="24"> Computational Linguistics Volume 18, Number 3 Recalling further that the adjectives we looked at above fell into not two but four distinct classes, we might expect to find nouns as well that belong to the other two classes, IT-SUBJ and SLASH-COMP. Such instances are found in English, as illustrated for IT-SUBJ nouns by the examples in (25), and for SLASH-COMP nouns by those in (26), drawn from Lasnik and Fiengo. 11 (25) a. It would be a mistake to fire Bill.</Paragraph> <Paragraph position="25"> b. It was a shock to find Bill here.</Paragraph> <Paragraph position="26"> c. *Bill would be a mistake to fire.</Paragraph> <Paragraph position="27"> d. *Bill was a shock to find here.</Paragraph> <Paragraph position="28"> (26) a. This room is a pigsty to behold.</Paragraph> <Paragraph position="29"> b. Nureyev is a marvel to watch.</Paragraph> <Paragraph position="30"> c. *It is a pigsty to behold this room.</Paragraph> <Paragraph position="31"> d. *It is a marvel to watch Nureyev.</Paragraph> <Paragraph position="32"> The noun mistake and the adjective possible have in common just those properties specified by the IT-SUBJ class (together with its superclasses); and like the differences between pleasure and easy, their differences result from mistake being a member of the COMMON-NOUN class while possible inherits from the ADJECTIVE class. Since the lexical rule relating the two variants of pleasure (and the two variants of easy) is defined to link members of the two classes SLASH-EASY and IT-EASY, the rule correctly does not predict the existence of similar alternate entries for nouns like mistake and pigsty. Interaction with lexical rules. Given that the domain of lexical rules is always one or more word classes, and that the LR-Intraposition rule is defined on the IT-SUBJ class, we predict the grammaticality of the following examples with pleasure nouns, since they also have entries belonging to the IT-SUBJ class, and should be expected to conform to the LR-Intraposition rule. Here again, the combined devices of inheritance and lexical rule produce the desired results for nouns without requiring that anything be added to the analysis motivated from data on adjectives and verbs. (27) a. (For me) to stay another day would be a real pleasure.</Paragraph> <Paragraph position="33"> b. It would be a real pleasure (for me) to stay another day.</Paragraph> <Paragraph position="34"> c. To visit Venice now might be a disappointment for you.</Paragraph> <Paragraph position="35"> d. It might be a disappointment for you to visit Venice now.</Paragraph> </Section> <Section position="2" start_page="287" end_page="293" type="sub_section"> <SectionTitle> 5.2 Too and Enough </SectionTitle> <Paragraph position="0"> To drive home our central point about the expressive and predictive power of inheritance in lexical representation, we turn to a third, small class of lexical entries that show complementation properties like those we have already seen. Jackendoff (1972) noticed that the words too and enough also appear in constructions with an infinitival complement that contains an NP gap, as illustrated in (28) with examples drawn from 11 Additional IT-SUBI nouns include battle, disgrace, error, honor, relief, shock, and surprise. Other SLASH-COMP nouns include beauty and terror.</Paragraph> <Paragraph position="1"> Dan Flickinger and John Nerbonne Inheritance and Complementation Lasnik and Fiengo (1974). 12 (28) a. The mattress is thin.</Paragraph> <Paragraph position="2"> b. *The mattress is thin to sleep on.</Paragraph> <Paragraph position="3"> c. The mattress is too thin to sleep on.</Paragraph> <Paragraph position="4"> d. The football is soft.</Paragraph> <Paragraph position="5"> f. *The football is soft to kick.</Paragraph> <Paragraph position="6"> g. The football is soft enough to kick.</Paragraph> <Paragraph position="7"> In particular, the examples in (29) suggest that these adverbs select for complements that are the same as adjectives like pretty, entries that are not related via lexical rule to variants that license an expletive it subject.</Paragraph> <Paragraph position="8"> (29) a. *It is too thin to sleep on this mattress.</Paragraph> <Paragraph position="9"> b. *It is soft enough to kick this football: Informally, it seems that when too or enough combines with an ordinary adjective, the resulting phrase (too thin and soft enough) exhibit complementation properties very much like those of pretty adjectives. By defining the lexical entries for these two adverbial specifiers as members of the SLASH-COMP class, we begin to provide an account for examples (28c,g) as well as those in (29). The entry for too is given in (30), inheriting both from the ADVERB class and from the SLASH-COMP class; the entry for enough is similar, leaving out of the present discussion an account of the linear order difference between the two adverbs with respect to the adjective they modify. With the inclusion of this class of adverbs, our lexical subhierarchy involving complementation of slashed VPs has grown to a point where it surely demonstrates the virtues of the structured lexicon approach. Figure 5 illustrates the more complete structure. It is a curious fact that the number of lexical classes does not grow enormously even while fairly detailed analyses involving very different grammatical areas are undertaken. In several years of development at Hewlett-Packard Laboratories involving detailed analyses of dozens of constructions, the number of word classes never exceeded 400. This must be due finally, not to the lexical analysis tool, but rather to the tendency of language to reuse significant classes.</Paragraph> <Paragraph position="10"> This analysis of these two unusual adverbs has left begging an important issue about how the complementation specifications provided by too are propagated up to the phrase too thin. 13 We have said little here about how lexically supplied 12 Baltin (1987) presents a more recent analysis of these &quot;degree complements.&quot; 13 One might be tempted to try a lexical rule approach that would treat too thin as a derived lexical item that selects for a VP complement. But slightly more complicated examples quickly render this approach untenable. Cf. This country is too thinly populated to worry about (where we take the scope of the specifier too to be thinly populated). Here, the lexicalized form that selects for a VP complement would have to be The lexical subhierarchy involving elements that govern &quot;slashed&quot; verb phrases. Note that the original hierarchy needed very little modification, merely addition. We speculate that this is due to the fact that significant classes are being identified in detailed grammatical description. There is also a version of too that inherits from ADVERB and IT-EASY that is not shown (since it was not discussed). The asymmetry is only apparent.</Paragraph> <Paragraph position="11"> subcategorization information is employed in parsing, referring the reader to full accounts given in Pollard and Sag (1987) and related references. Yet it is clear that something more must be said about this construction, given that in HPSG it is the syntactic head of a phrase that imposes constraints on its complements; and we assume that thin, not too, is the head of the phrase too thin to sleep on. To motivate the necessary elaboration of our analysis for these two adverbs, we turn to one more set of data involving gappy infinitival complements, one that has received little study to date. Excursus on subcategorization transfer. As the example in (31) shows, adjectives such as easy appear not only in predicative constructions like those illustrated above, but also as nominal modifiers.</Paragraph> <Paragraph position="12"> (31) John is an easy man to talk to.</Paragraph> <Paragraph position="13"> too thinly populated, a consequence we regard as unacceptable.</Paragraph> <Paragraph position="14"> Dan Flickinger and John Nerbonne Inheritance and Complementation While the example in (31) is good, employing the easy that belongs to the EASY-SLASH class, the examples in (32-33) are ungrammatical. The analysis we have provided thus far does not yet explain the grammaticality of (31) and the ungrammati- null cality of (32,33).</Paragraph> <Paragraph position="15"> (32) a. *John is an easy to talk to man.</Paragraph> <Paragraph position="16"> b. *John is an easy man.</Paragraph> <Paragraph position="17"> (33) a. *John is an easy man to talk to Bill.</Paragraph> <Paragraph position="18"> b. *John is a man easy to talk to Bill.</Paragraph> <Paragraph position="19"> We will focus on explaining the grammaticality of (31), assuming that the right syntactic structure for the sentence is the binary-branching structure given in (34), where easy forms a constituent with man, and where to talk to is sister to the phrase easy man. We adopt the binary structure largely because it will simplify the exposition here; it might be equally defensible to hold that easy, man, and to talk to are all sisters of a single phrase24 What is awkward about this structure is that the head noun man does not by itself subcategorize for the VP/NP. ~5 Rather, it seems that when easy combines with man, the resulting phrase has a subcategorization list that contains not only the optional and obligatory complements that man started out with, but also the obligatory VP/NP complement and the optional For-PP controller required by easy. No mechanism presented so far provides for an adjunct combining with its head to affect the subcategorization of that head or of the resulting phrase. Yet if the phrase structure proposed in (33) is correct, some kind of merging of subcat information between adjunct and head must be provided for. 16 14 And it is worth noting that the alternative constituent structure would not modify the head relationship, and therefore would not substantially alter the analytic problem--that of explaining how a complement to talk to can be licensed by a nonhead.</Paragraph> <Paragraph position="20"> 15 At least not with the intended reading. There is a suspiciously similar construction, illustrated in (i), which might be expected to shed some light on the proper analysis of (31), but which has a restricted enough interpretation to suggest that it should be treated separately, probably inheriting a specification from the more general construction exhibited in (31).</Paragraph> <Paragraph position="21"> a. John is a man to admire.</Paragraph> <Paragraph position="22"> (i) b. Mary is a woman to emulate.</Paragraph> <Paragraph position="23"> c. This is a word to keep on the tip of your tongue.</Paragraph> <Paragraph position="24"> These examples seem to mean something like John is a good man to admire or Mary is a good woman to emulate, where the semantic contribution of good has been incorporated into the N-VP/NP construction in (i). To test this, consider the examples in (ii), where the good reading should lead to an anomalous interpretation, and does (cf. the corresponding examples in (iii)).</Paragraph> <Paragraph position="25"> (ii) a. ?Mary is a person to underestimate.</Paragraph> <Paragraph position="26"> b. ?Sharks are animals to tame.</Paragraph> <Paragraph position="27"> (iii) a. Mary is an easy person to underestimate.</Paragraph> <Paragraph position="28"> b. Sharks are difficult animals to tame.</Paragraph> <Paragraph position="29"> Given the constrained interpretations of examples like those in (i-ii), it does not seem defensible to treat easy man to talk to as simply the modifier easy combining with man to talk to. Any such attempt would be strained in accounting for (ii); in addition, such an analysis would leave unexplained the ungrammaticality of *John is an easy man.</Paragraph> <Paragraph position="30"> 16 It is probably worth noting that extraposition seems unlikely to be generalizable to all cases involving transferable subcats, at least if extraposition is to be bounded uniformly: (i) An easy man to talk to arrived yesterday.</Paragraph> <Paragraph position="31"> (ii) *An easy man arrived yesterday to talk to.</Paragraph> <Paragraph position="32"> Complex adjectivals &quot;wrapped&quot; around a modified noun. Note that the N head of the I~I constituent in construction with the complex adjectival has not licensed it. Subcategorization transfer has taken place.</Paragraph> <Paragraph position="33"> The examples in (34-35) illustrate that the flow of information from an adjunct's list of subcats to the head's must be quite restricted; it would not do to simply merge the Complements list of any adjunct with that of the head in every case.</Paragraph> <Paragraph position="34"> (34) a. *an eager man to please b. *a fearful man of snakes c. *a frightened man by snakes d. *an angry man at John (35) a. a man eager to please b. a man fearful of snakes c. a man frightened by snakes d. a man angry at John Dan Flickinger and John Nerbonne Inheritance and Complementation The above examples might suggest that what distinguishes easy from these other adjectives is that the VP/NP complement of easy is obligatory, while the PP complements of the above adjectives are optional. While there are not many adjectives against which to test this hypothesis, the one clear case of an adjective that takes an obligatory complement counts against the idea: (36) a. a man fond of snakes b. *a fond man c. *a fond man of snakes.</Paragraph> <Paragraph position="35"> The analysis we propose localizes in lexical entries the ability of a subcat to be transferred from adjunct to head. Just as subcats can be marked for the obligatory/optional distinction in a class definition or in a lexical entry, so can they be marked for a distinction we term transferable. While as a default subcats will be nontransferable, those subcats that are identified by a class or lexical entry as transferable will be subject to the following informally stated principle.</Paragraph> <Paragraph position="36"> Transferable Subcat Principle. When a transferable subcat on a daughter in a local subtree is not associated with some sister in that subtree, the subcat becomes part of the corresponding subcat list of the head daughter in that subtree.</Paragraph> <Paragraph position="37"> In the constructions studied here, this principle applies in cases where the lexical entry or phrase with a transferable subcat serves as an adjunct (easy) or a specifier (too), so that the word or phrase's subcat list is not used directly. The intent of the principle in such cases is to make the transferable subcat a part of the head, so the subcategorization principle will ensure that the information is propagated to the mother node. This is intended as a modification of the subcategorization principle---note that it has the effect of licensing a kind of &quot;discontinuous constituent. &quot;~7 Having introduced this additional property of subcats, that they can be specified as transferable, we note that the default value for this property must be negative, since in general subcats from adjuncts and specifiers do not pass to heads, as seen in (33) and (34) above. This default value will be overridden for the VP/NP and the For- null PP subcats in the SLASH-COMP class, to reflect the grammaticality of both examples in (37).</Paragraph> <Paragraph position="38"> (37) a. That was a melodious sonata to listen to.</Paragraph> <Paragraph position="39"> b. John is an easy man to please.</Paragraph> <Paragraph position="40"> Members of the SLASH-COMP class, including the relevant lexical entry for easy, will inherit this nondefault transferable property for both the XComp and the For-PP, so when easy combines as an adjunct with the head noun man, these two subcats will become part of the subcategorization of man, by the principle above, and will then become part of the subcategorization of the node easy man, accounting for the grammaticality of (31) above.</Paragraph> <Paragraph position="41"> Example (32a) will be ruled out because of an independent constraint that restricts pre-head adjuncts to those that are (Head-Final +). Example (32b) is excluded because easy has an obligatory VP/NP complement, which must be included as an 17 The ability to transfer a subcategorization requirement from a modifier to a mother (or to a head) is perhaps a bit similar in effect to FUNCTION COMPOSITION in categorial grammar (Bach 1988). But in HPSG the possibility may be lexically constrained.</Paragraph> <Paragraph position="42"> Computational Linguistics Volume 18, Number 3 obligatory complement of the phrase easy man, due to the convention adopted above about merging of subcat information between a head and its sister. Finally, (33a) is excluded because the easy that requires an unslashed VP complement will not pass on its XComp subcat to the noun it modifies, since that XComp is, like most subcats, nontransferrable. So easy man to please Bill will be excluded for the same reason that eager man to please is excluded: nothing licenses the postnominal infinitival VP. Example (33b) is probably best excluded on semantic grounds, since the subject of easy to please Bill is an expletive pronoun, the wrong sort to unify with the head noun being modified. On the assumption that a noun must serve semantically as the subject of adjectival adjuncts, those adjuncts must specify some thematic role for the noun to play. Thus any adjective that requires an expletive subject should give rise to a semantically ill-formed expression when it appears as an adjunct to a noun. What prevents the IT-EASY easy from serving as an adjunct to problem is the fact that this easy requires an expletive it subject.</Paragraph> <Paragraph position="43"> Given this transferable complement mechanism, we may straightforwardly complete the analysis of the earlier too/enough examples: the lexical entries for these two adverbs simply specify that their gappy infinitival complement is transferable. TM</Paragraph> </Section> </Section> <Section position="8" start_page="293" end_page="295" type="metho"> <SectionTitle> 6. Conclusions and Future Directions </SectionTitle> <Paragraph position="0"> The study of inheritance and, more generally, the study of structured lexical representations is an exciting and promising field. We would like to use this section to summarize how we view this work and to suggest directions in which we feel it should move.</Paragraph> <Section position="1" start_page="293" end_page="294" type="sub_section"> <SectionTitle> 6.1 Conclusions </SectionTitle> <Paragraph position="0"> We have presented a treatment of complementation that uses nonmonotonic lexical inheritance. The lexical specifications are quite compact and therefore both readily extendible and easily modified. We pointed out cases where nonmonotonic, default specification seems most natural, and the entire treatment turns on the possibility of there being genuine multiple inheritance of a &quot;complements&quot; attribute.</Paragraph> <Paragraph position="1"> We adopt a skeptical approach to inheritance conflict. If there are inheritance conflicts in the system presented here, nothing is inherited. Mechanisms that warn users about such conflicts are useful, but we are wary of attempts to decide conflicts &quot;intelligently.&quot; They seem likely to us to lead to cases where minor changes may have remote consequences, which would detract from maintainability.</Paragraph> <Paragraph position="2"> We do not feel that we have overstated the case for structured lexicons by choosing a particularly messy or poorly understood area. To insist on this point somewhat, let us note that we omitted significant aspects of the grammar of the &quot;raising nouns,&quot; e.g., their complements, specifiers, and adjuncts. 19 Grammar abounds in poorly understood areas, including comparatives, superlatives, adverbials, internal NP syntax, and the &quot;specialized grammars&quot; found in dates, places, and technical vocabulary. All of these areas can benefit from the application of a tool for complex lexical description.</Paragraph> <Paragraph position="3"> 18 This leaves much to be said about the lexical properties of too and enough, but more detailed analysis at this point would take us too far afield; it is clear enough that, whatever their other properties, these two adverbs share complementation properties with the adjectives and nouns studied here.</Paragraph> <Paragraph position="4"> 19 For example, for &quot;pleasure&quot; nouns, some adjectives are okay, but not all (a real/*competent pleasure to work with; relative clauses are impossible: Sally is a pleasure *\[that is real\] to work with; and some nominal complements are fine: Sally was a pleasure of the rarest kind to work with.</Paragraph> <Paragraph position="5"> Dan Flickinger and John Nerbonne Inheritance and Cornplernentation</Paragraph> </Section> <Section position="2" start_page="294" end_page="295" type="sub_section"> <SectionTitle> 6.2 Hypotheses or Tools? </SectionTitle> <Paragraph position="0"> The conclusions above may be read as a plea for the employment of an important tool in computational linguistics, and, indeed, we see the primary significance of the use of structured lexicons not in new expressive power which they bring to natural language processing or description (there is perhaps none), but rather in the increased ease and reliability with which they allow old hypotheses to be formulated and put to use.</Paragraph> <Paragraph position="1"> Brachman (1983, p. 35) summarizes the dominant view of inheritance in knowledge representation: Even though much has been made in the past of the significance of inheritance in semantic nets, no one has been able to show that it makes any difference in the expressive power of the system that advertises it... It is strictly implementational.</Paragraph> <Paragraph position="2"> Given this authority on the technical side, it may be surprising to hear more application-oriented users of inheritance mechanisms hedging at all on whether there is any scientific significance to the proposal here. But there is at least a potential candidate: lexical rules may distinguish inherited from specified information.</Paragraph> <Paragraph position="3"> In expressing the relationships between members of two sets of lexical entries, we make crucial use of the distinction between idiosyncratically specified information (which appears in a sparse Inonredundant\] lexical entry) and inherited information. We have adopted here the restrictive hypothesis, proposed in Flickinger (1987), that lexical rules hold for minimally specified lexical entries, without having access to inherited, predictable information. Adopting this hypothesis imposes a constraint on the form and function of lexical rules that is strong, perhaps too strong, but one that allows a simpler formulation of rules by keeping to a minimum the amount of information to be managed. Only two kinds of information are relevant for a lexical rule: the word classes that each of the two related entries belong to, and any idiosyncratic properties specified by either lexical entry. We note that if lexical rules were insensitive to the distinction between idiosyncratic and predictable properties of lexical entries, the statement of even a simple rule like LR-PAST, given earlier, would be much more difficult. If the lexical rule for past tense verbs had to cope with fully specified entries that blurred this distinction, it would be difficult to express in the rule just which properties of the one entry had to match in the other, related entry.</Paragraph> <Paragraph position="4"> For example, the verb like idiosyncratically requires a verbal complement that is either infinitival or gerundive, while the verb enjoy does not allow the infinitival form, allowing only the gerundive form for its complement. Since all of the inflected forms of like allow the same choice of two permissible forms for the verbal complement, while all of the inflected forms of enjoy insist on the gerundive complement, the lexical rules like LR-PAST or the similar one for present third singular forms must preserve these idiosyncracies. Yet a fully specified entry for the base form enjoy stipulates not just the form of the complement, which would have to be identical in the present third singular entry enjoys; the fully specified base entry for enjoy also specifies that its subject be unmarked for number, an indifference that crucially must not be shared by the entry for enjoys. Short of tagging each attribute value in a fully specified entry as local or inherited, it is not clear how the lexical rule for present third singular forms could be constrained to ensure identity of the verbal complement's VFORM value while ignoring differences in the subject's AGREEMENT value for these two entries for enjoy. In sharp contrast, this difference in idiosyncratic vs. inherited information can be exploited by lexical rules without stipulation when they are constrained to apply only to minimally specified entries.</Paragraph> <Paragraph position="5"> Computational Linguistics Volume 18, Number 3 It may not be superfluous to add that, even if the argument above about distinguishing inherited and specified information is ultimately fallacious, so that the use of inheritance were seen purely as a tool, and not at all as a scientific hypothesis, it may nonetheless prove to be of great significance, just as many tools have advanced areas of science that nothing to do with their development. The development of lenses revolutionized astronomy, even though glass grinding embodied no astronomical hypotheses. null</Paragraph> </Section> <Section position="3" start_page="295" end_page="295" type="sub_section"> <SectionTitle> 6.3 Emergent Issues in Structured Lexicons </SectionTitle> <Paragraph position="0"> Perhaps more interesting are the many directions in which this research may be developed. We suggest some of these in the questions below.</Paragraph> <Paragraph position="1"> What are lexical classes and lexical entries? The careful reader noted in Section 5 above that our lexical specifications are translated into feature structures. Theoretically, we could dispense with the translation for nearly all of the information involved, and have the lexicon describe feature structures directly. But this does not correspond to our implementation, nor are we clear on how, e.g., information on lexical rules and their application ought to be rendered in features. Perhaps lexical entries must be structured so that one component of a lexical entry is a feature structure, while others are not. 2deg Can inheritance be exploited in the specification of inflectional variation? This appears to be a promising area of application, since in general, one can view inflected elements as further specifications of abstract lexemes (cf. Evans and Gazdar 1989 for an intriguing proposal).</Paragraph> <Paragraph position="2"> Can derivational lexical rules be treated more satisfactorily? For example, it is clear that at least some lexical rules relate not merely a pair of word classes, but rather entire lexical substructures (involving several classes) to one another. Can the techniques of inheritance be applied here, so that exceptional elements may be easily accommodated?</Paragraph> </Section> </Section> <Section position="9" start_page="295" end_page="295" type="metho"> <SectionTitle> Acknowledgments </SectionTitle> <Paragraph position="0"> We are indebted to Mark Gawron, Masayo Iida, Bill Ladusaw, Joachim Laubsch, Carl Pollard, and Tom Wasow for frequent conversations about this analysis. We are also grateful to Anthony Kroch, the participants at the Tilburg Workshop on Inheritance in Natural Language Processing, and three referees for further comments.</Paragraph> <Paragraph position="1"> This work was partially supported by a research grant, ITW 9002 0, from the German Bundesministerium ftir Forschung und Technologie to the DFKI DISCO project.</Paragraph> </Section> class="xml-element"></Paper>