File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/91/w91-0209_metho.xml
Size: 25,882 bytes
Last Modified: 2025-10-06 14:12:48
<?xml version="1.0" standalone="yes"?> <Paper uid="W91-0209"> <Title>Lexical Operations in a Unification-based Framework</Title> <Section position="3" start_page="90" end_page="96" type="metho"> <SectionTitle> 2 The Lexical Representation Language </SectionTitle> <Paragraph position="0"> Our lexical representation language is unification-based, allowing complex interconnections between syntactic and semantic information to be defined, and making a tight interface possible between the lexicon and a parser/interpreter. It supports a restricted range of operations; (default) unification, (default) inheritance and lexical rule application. It does not support arbitrary inference. The language is based on tile use of typed feature structures similar to those described in Carpenter (1990). Feature structures must be well-formed with respect to types and particular features will only be appropriate to specified types and their subtypes. Types are hierarchically ordered; tile association of constraints with types allows non-default inheritance. We augment this with a restricted concept of default inheritance (allowing only 'orthogonal' multiple inheritance (Touretzky 1986)); default inheritance is formalised in terms of default unification of feature structures ordered by an inheritance hierarchy. The type system constrains both default inheritance and lexical rule application. This representation language is described in detail in Copestake et al (1991); the following sections are an informal description illustrated with relevant examples.</Paragraph> <Section position="1" start_page="90" end_page="93" type="sub_section"> <SectionTitle> 2.1 The type system </SectionTitle> <Paragraph position="0"> The type hierarchy defines a partial ordering (notated K) on the types and specifies which types are consistent. Only feature structures with mutually consistent types call be unified -- two types which are unordered in the hierarchy are assumed to be inconsistent unless the user explicitly specifies a common subtype. Every consistent set of types S C TYPE has a unique greatest lower bound or meet (notation r'lS). This condition allows feature structures to be typed deterministically -- if two feature structures of types a and b are unified the type of the result will be a t-1 b, which must be unique if it exists. If a I-1 b does not exist unification fails. Thus in the fragment of a type hierarchy shown in Figure 1 artifact and physobj are consistent; artifact I-1 physobj = artifact_obj.</Paragraph> <Paragraph position="1"> Our system differs somewhat from that described by Carpenter (1990) in that we adopt a different notion of well-formedness of typed feature structures. In our system every type must have exactly one associated feature structure which acts as a constraint on all feature structures of that type; by subsuming all well-formed feature structures of that type. The constraint also defines which features are appropriate for a particular type; a well formed feature structure may only contain appropriate features. Constraints are inherited by all subtypes of a type, but a subtype may introduce new features (which will be inherited as appropriate features by all its subtypes). A constraint on a type is a well-formed feature structure of that type; all constraints must therefore be mutually consistent. Constraints can be seen as extending the PATR-II notion of templates (eg. Shieber, 1986) in that the inheritance of constraints allows concise definitions of all feature structures, not just lexical entries; but in an untyped system, such as PATR.-II, there is no restriction on the features that can occur in a feature structure.</Paragraph> <Paragraph position="2"> For example the constraints associated with the types artifact and physobj might be: artifact 'ro,:t,t(: = formula\]</Paragraph> <Paragraph position="4"> Bold case indicates types; thus, for instance formula is a type and any feature structure of type artifact must have a feature structure of type formula as the value for its TELIC (purpose) feature, formula is intended to represent a formula in predicate logic, it therefore has a complex constraint itself:</Paragraph> <Paragraph position="6"> In contrast solid is an atomic type, it has no appropriate features and its constraint is simply the atomic feature structure \[solid\].</Paragraph> <Paragraph position="7"> The constraint on artifact_obj will contain information inherited from both parents, thus:</Paragraph> <Paragraph position="9"> Further examples of constraints and features which we will use in examples in this paper are:</Paragraph> <Paragraph position="11"> \] noun mass-noun_C lex-sign SYNTAX= \[COUNT=--1 The feature structure below is well-formed since it contains all the appropriate features and no inappropriate ones, it is subsumed by the constraints on its type and all its substructures are well-formed.</Paragraph> <Paragraph position="13"> Given the type system introduced above, alexicalentry, suchas: haddock 1 count-noun <rqs> = animal.</Paragraph> <Paragraph position="14"> would be expanded outinto such a ~ature structure I .</Paragraph> </Section> <Section position="2" start_page="93" end_page="94" type="sub_section"> <SectionTitle> 2.2 Default inheritance </SectionTitle> <Paragraph position="0"> To allow default inheritance we introduce the concept of psor~; a feature structure from which another feature structure inherits information, by default. The hierarchical ordering on psorts (which must be consistent with the type hierarchy) provides an order on defaults. Default inheritance is implemented by a version of default unification. Only orthogonal multiple inheritance (Touretzky 1986) is allowed; information inherited from multiple parents must not be contradictory. (A default inheritance hierarchy which connects semantic parts of lexical entries can be derived semi-automatically from taxonomies extracted from conventional dictionaries, see Copestake, 1990a). We refer to this particular case of the psort hierarchy as an IS_A hierarchy. Values of features can be associated either manually or semi-automatically with psorts in the IS_A hierarchy; the more specific word senses then inherit them, by default. (Defaults may also be useful in the representation of syntactic information in the lexicon (e.g. Flickinger, 1987).) Since the type system constrains the psort system it also constrains multiple default inheritance. If the value of the FOOD-TEMPERATURE feature for &quot;drink 2 (1)&quot; is low then this information would be inherited by the entry for &quot;beer&quot; which is below &quot;drink 2 (1)&quot; in the IS_A hierarchy. However inherited information may be overridden by associating other values with psorts lower in the hierarchy; for example although &quot;tea&quot; is under &quot;drink Types and features thus provide an organisation on the information which is necessary for interaction with lexical and syntactic rules. The IS_A hierarchy is motivated by defining its semantics in terms of the real world entities corresponding to the word senses and demonstrating that default inheritance of attributes in the lexicon correlates with default reasoning about properties of the entities. Copestake (1990a) outlines a preliminary attempt to formalise the relationship between this aspect of lexical semantics and world knowledge.</Paragraph> </Section> <Section position="3" start_page="94" end_page="96" type="sub_section"> <SectionTitle> 2.3 Lexical rules </SectionTitle> <Paragraph position="0"> A lexical rule is a feature structure of type lexical-rule. The expanded constraint for the</Paragraph> <Paragraph position="2"> thus all lexical rules have to have the features 0 and I which must both have values which are of type lex_sign.</Paragraph> <Paragraph position="3"> New lexical signs may be generated by unifying a copy of the lexical entry with the feature structure at the end of the path <I> in a copy of the lexical rule -- the feature structure at the end of the path <0> is then the new lexical sign. Lexical rules are indexed by the type of their &quot;input&quot; and &quot;output&quot; feature structures, so they will only be applied to entries of the appropriate type and will only create well-typed entries.</Paragraph> <Paragraph position="4"> A number of productive or quasi-productive phenomena, such as deverbal nominalisation, 'grinding', and so forth, can be represented as lexical rules which generate further lexical entries. A general type for grinding lexical rules could be specified in our system as follows:</Paragraph> <Paragraph position="6"> The effect of the iexical rule is to transform a count noun with the 'relativised qualia structure' (RQS, Calzolari, 1991) properties appropriate to an individuated physical object ind_obj into a mass noun with properties appropriate for a substance substance. Thus the core component of grinding is a linguistic, syntactic operation which affects syntactic realisation, such as the ability to appear without a determiner, correlated with an abstract and underspecified semantic operation. We would claim that specific predicational and syntactic contexts will result in coercion (application of the lexical rule) and that this much, at least, of the 'grinding' family of sense extensions must be seen as a non-default and essentially linguistic process.</Paragraph> <Paragraph position="7"> We specialise the grinding rule to allow for cases such as the animal/meat regular sense extension explicitly. The typed framework provides us with a natural method of characterising the subparts of the lexicon to which such rules should apply. The lexical rules can, in effect, be parameterised by inheritance in the type system. As our theory of lexical organisation allows us to make fine distinctions between classes of lexemes, in terms of both syntactic and semantic properties encoded in the type system, we expect that many processes which have been characterised as partially productive or semantic specialisations of productive processes will be characterisable as fully productive rules of sense extension applying to smaller semantically coherent subsets of the lexicon.</Paragraph> <Paragraph position="8"> For example given the type hierarchy shown in Figure 1 we can give rules which inherit information from grinding such as animal_grinding:</Paragraph> <Paragraph position="10"> Thus given the lexical entry for &quot;haddock&quot; shown above we can apply the lexical rule to generate a sense meaning &quot;haddock-flesh&quot; (partially represented as):</Paragraph> <Paragraph position="12"> (where the specification of tile value eat for tile relic role arises from the constraint on the type food_substance, inherited from food, and the type mass-noun arises from grinding.) It would not be possible to apply this lexical rule to &quot;book&quot; and get a sense denoting &quot;book-flesh&quot; because &quot;book&quot; has the type inanimate_obj which is incompatible with animal. It would still be possible to apply the general lexical rule for grinding and to get a mass use of &quot;book&quot; but the denotation of the mass sense would be underspecified.</Paragraph> <Paragraph position="13"> We would expect the context to provide a more specific interpretation of the mass sense in such a (less-conventionalised) case.</Paragraph> <Paragraph position="14"> Furthermore, our approach provides a natural mechanism for dealing with semantic specialisation or restriction. We can represent lexicalised items as inheriting information by default from a psort which is the result; of applying an appropriate lexical rule to the base form. Such items may have specific information associated with them. In cases where the lexical rule predicts the extended sense exactly, the specific information will duplicate information already present. If tile rule is correct, but incomplete, the specific information will augment the inherited information. If it is partially incorrect, the more specific information will override that inherited from the result of lexical rule application.</Paragraph> <Paragraph position="15"> In an untyped system this representation would not constrain the structures associated with lexicalised derived forms, since tile entire feature structure output by the lexical rule might be overridden. Itowever default inheritance is constrained by the type system so that information may only be inherited from a structure of the same or higher type, and thus this treatment predicts that a derived form can never have a type which is incompatible with that determined by the lexical rule.</Paragraph> <Paragraph position="16"> We can illustrate the manner in which this type of semi-productivity might be dealt with if we assume that we are attempting to construct a lexicon semi-automatically from a conventional dictionary (e.g. Copestake, 1990a). If the result of applying a lexical rule to a sense is notated as sense+rule-name (eg lamb_l+animal_grinding) then the representation of the sense lamb (2) (&quot;the meat&quot;, from the Longman Dictionary off Contemporary English LDOCE) might be: lamb 2 < lamb_l+animal_grinding.</Paragraph> <Paragraph position="17"> In this case no extra information need be added. In contrast the entry for lamb (3) (&quot;a young gentle person&quot; LDOCE) might augment the information inherited from the lexical rule: lamb 3 < lamb_l+animal_metaphor < rqs : age > = low.</Paragraph> <Paragraph position="18"> In the case of &quot;haddock&quot;, where no LDOCE entry is found, the structure derived from the lexieal rule alone would be used.</Paragraph> <Paragraph position="19"> We can regard morphological rules as a particular type of lexical rule where the orthography of the output is not equal to that of the input (we assume that the regular spelling changes involved in affixation will be dealt with by a separate system, e.g. Cahill, 1990). In this case we could represent irregular forms as having an orthographic form which overrides that produced by rule application. In principle, multiple lexieal rules may be applied in sequence. For example the resultative senses of &quot;replacement&quot; and &quot;purchase&quot; mentioned in the introduction would be the result of applying a metonymic sense extension rule to the result of the nominalisation process. All outputs of lexical rules must be potentially valid lexieal entries. In the case of conversion or zero-derivational processes we wish to restrict the set of lexical rules so that application may not be circular -- that is if there is a lexical rule which could generate the set of feature structures F2 from the set F1, no other lexieal rule or sequence of lexical rules may be specified which could generate any member of F1, or a feature structure subsuming any member of F1, starting from any member of the set F2, since lexical rule application would not then terminate. (We can check for such potential circularities relatively efficiently by looking at the type of the feature structure that a lexical rule generates rather than the entire feature structure.) However, this condition is overrestrictive in general because some types of derivationai rule can apply to their own output iteratively (&quot;meta-meta-theory&quot;, &quot;anti-anti-missile&quot;, &quot;great-great-grandmother&quot;, &quot;re-re-program&quot;). This observation suggests that we need to distinguish types of lexical rule, such as at least derivational rules and processes of conversion, and associate slightly different constraints with them.</Paragraph> </Section> </Section> <Section position="4" start_page="96" end_page="99" type="metho"> <SectionTitle> 3 Grinding </SectionTitle> <Paragraph position="0"> In this section we justify our treatment of grinding as a lexical rule and show why relatively complex semantic information is needed to adequately account for this sense extension.</Paragraph> <Paragraph position="1"> Tile first point to consider is that grinding processes appear to be genuinely productive.</Paragraph> <Paragraph position="2"> Thus we find: Badger hams are a delicacy in China while mole is eaten in many parts of Africa.</Paragraph> <Paragraph position="3"> in the Lancaster-Bergen/Oslo (LOB) corpus. We therefore cannot assume that the ground senses are necessarily lexicalised, even ill the relatively conventionalised uses to mean meat, fur etc.</Paragraph> <Paragraph position="4"> One approach which allows for this productivity is to treat all nouns as being initially underspecified with respect to the count/mass distinction. Thus it is possible to produce a grammar where nouns are initially undefined with respect to a syntactic count feature and where lamb ~ is, in effect, taken as denoting both animals and meat and so on (see the &quot;p-theory&quot; in Pelletier and Schubert 1986 and also Copestake 1990b). In contexts * where one interpretation is forced (&quot;a piece of lamb&quot; vs &quot;two lambs&quot;) the predicate can be restricted to denote either count or the mass senses (in this case either the animal or the meat senses). However this seems to predict that NPs such as &quot;the lamb&quot; are vague rather than ambiguous between count and mass readings. Thus the peculiarity of sentences such as: ? John fed and carved the lamb.</Paragraph> <Paragraph position="5"> is not accounted for (see also the introduction). It is perhaps significant that in most dictionaries the mass sense is specified as well as a count sense for the conventionalised grinding examples we have been considering. Lexicographers are sometimes aware of the regularity of the extension (Atkins 1990) but have no way of representing this in a conventional dictionary. We thus regard a noun like &quot;lamb&quot; as ambiguous between mass and count senses rather than vague, and the mass (meat) sense as an extension of the count (animal) sense, specified by lexical rule. (There are cases where it is reasonable to claim that a nominal should be underspecified with respect to the count/mass distinction; it is frequently unclear whether individuation is occurring with nouns like &quot;data&quot;, however in the grinding examples there is a clear change in meaning.) Our current treatment thus has similarities to the &quot;s-theory&quot; of Pelletier and Schubert (1986) where &quot;lexical extension rules&quot; are used to produce mass nouns from count nouns. However these rules merely change the value of the syntactic count feature, apply a predicate operator which is tile same for all cases of grinding, and mark the mass sense resulting as &quot;+EXT&quot; which is supposed to suggest that it is in some way abnormal. This is clearly inadequate: John carved the lamb.</Paragraph> <Paragraph position="6"> would be marked &quot;+EXT&quot; for the reading where lamb was used in a mass sense and not for the count sense reading. By having an inheritance ordering on lexical rules we can express the conventionalised processes that apply to semantically specified parts of the lexicon and account for the possibility of multiple distinct mass senses being possible; for example &quot;rabbit&quot; is given distinct senses in LDOCE for the meat and the fur, and (in context) an underspecified sense is available: After several lorries had run over the body, there was rabbit splattered all over the road.</Paragraph> <Paragraph position="7"> Although the denotation of the count sense and the mass sense are distinct there clearly is some relationship between them. A full account of sense extension must be able to represent relationships between the senses' denotations. For grinding in general the most specific claim that can apparently be made is that the ground sense denotes some &quot;stuff&quot; which was at some past time part of one or more individuals denoted by the count sense. (We can formally specify this relationship between the ground sense G and the base sense B as Vx, t\[G(x, t) ~ 3y, t'\[*B(y, t') A t' < t A x Eo y\]\] using the formalisation developed in Copestake (1990a) following Krifka (1987) where nominal predicates are taken as being true of quantities of matter at some time index, where *B denotes a potentially plural entity, and where Eo represents a relationship of material constituency. In the lexicon we actually use a feature ORIGIN which call be taken as an abbreviation for the relationship specified above.) A good theory of sense extension should give some treatment of blocking, which appears to occur with some cases of regular sense extension in a way that seems similar to derivational morphology. For example the use of &quot;pig&quot; to denote the meat seems to be blocked by the existence of &quot;pork&quot; -- &quot;pig&quot; can be used in the extended sense but such a use is marked, suggesting for example that the meat is distinctly inferior. To account for this we need to be able to recognise that &quot;pork&quot; is equivalent to the sense obtained from &quot;pig&quot; using the lexical rule. Although there are many problems with this (what do we mean by equivalence, why does this apparently not apply to metaphorical sense extension) in order to do it at all we clearly need a rich representation which indicates information such as &quot;origin&quot;.</Paragraph> <Paragraph position="8"> Bauer (1983) distinguishes two types of non-productivity (which he refers to as established senses) - lexicalisation and institutionalisation. Lexicalisation is defined as irregular and unpredictable modification of some or all of the semantic, syntactic or phonological properties of a derived form. Institutionalisation, by contrast, involves restriction, rather than modification along one of these dimensions; thus &quot;telephone box&quot; is institutionalised to mean telephone kiosk unambiguously, although the liberal rules of noun compounding predict other possibilities. In our approach, we can treat institutionalisation as a form of blocking in which forms such as &quot;telephone box&quot; would have separate entries equivalent to one productive meaning predicted by a putative lexical rule of compounding. This would predict a strong preference for this interpretation (except in a marked context). Bauer (1983:58) points out that some treatment along these lines will be required since the other meanings are not completely ruled out, and therefore simply listing them as independent entries will be inadequate. The more specific rules of grinding (as opposed to the most general rule) are instances of productive institutionalisation within sub-classes in that the specific interpretations they introduce can be overridden in marked contexts.</Paragraph> <Paragraph position="9"> Similarly, lexicalisation is usually a partial process which affects one aspect of a derived lexical entry, whilst the rest remains productive. Bauer gives the example of &quot;disbelieve&quot; which can be productively derived through a lexical rule which prefixes &quot;dis+&quot; to the verb &quot;believe&quot; with a predictable change of meaning, except that &quot;disbelieve&quot; does not inherit the syntactic properties of &quot;believe&quot; because it cannot take sentential or infinitival complements. The productive aspects of the relation between the two verbs can be expressed by a lexical rule for &quot;dis+&quot; prefixation, whilst the non-productive aspects can be captured naturally in this framework by positing an independent entry for &quot;disbelieve&quot; which overrides some of the information provided by the iexical rule.</Paragraph> <Paragraph position="10"> We think of lexical rules as defining the limits of coercion amongst lexemes and argue that lexical rule application, or selection of the derived entry (which is equivalent in many cases), will be forced when the type of the basic entry is incompatible with the syntactic or predicational context in which the lexeme occurs. Consider the following example, taken from the Lancaster-Bergen/Oslo (LOB) corpus: f More than 1,000 union men and their families arrived to play bowls, eat barbecued chicken and row on his fish-infested lake.</Paragraph> <Paragraph position="11"> The application of a grinding lexical rule is triggered by a combination of syntactic and ' semantic effects arising from the context (for example the predicate &quot;eat&quot; takes an object denoting food in preference to an animal, the bare NP &quot;mole&quot; in the earlier example must have a negative value for the syntactic feature count). By default, the most specific lexical rule applicable will be used, in this case &quot;animal.grinding&quot; as opposed to the general grinding rule and so the interpretation of &quot;chicken&quot; as &quot;chicken-flesh&quot; and &quot;mole&quot; as &quot;mole-flesh&quot; is possible, by default. More open-ended (non-lexical) reasoning might cause the default interpretation of the mass sense to be overridden ill some marked informationally-rich contexts. For example: John bit into the lamb.</Paragraph> <Paragraph position="12"> It kicked and struggled.</Paragraph> <Paragraph position="13"> We assume a similar account of the overriding of default interpretations with these types of example as we offered in the case of logical metonymies (Briscoe et al., 1990).</Paragraph> </Section> class="xml-element"></Paper>