File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/87/j87-3008_metho.xml
Size: 51,792 bytes
Last Modified: 2025-10-06 14:11:59
<?xml version="1.0" standalone="yes"?> <Paper uid="J87-3008"> <Title>A COMPUTATIONAL FRAMEWORK FOR LEXICAL DESCRIPTION</Title> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 4. WORD GRAMMAR </SectionTitle> <Paragraph position="0"> The morphological rules concerned with word-structure can be viewed as a &quot;Word Grammar&quot;, characterising derivationai and inflectional morphology in abstraction from the details of the actual character strings involved.</Paragraph> <Paragraph position="1"> These rules describe what constitutes an allowable sequence of morphemes, stating which concatenations are valid, and the syntactic class of the overall word formed by several morphemes. For example, happy+ness is a valid noun, but arrive+ness is not a valid word.</Paragraph> <Paragraph position="2"> The word grammar is based on the concept of features and values. Any constituent (morpheme, word, word-part, etc.) can be represented by a set of features and values, called a category. Our model of English morphology is based heavily on the GPSG model of syntactic features (cf. Gazdar et al. (1985), chap. 2) (although it could be used simply as a very general feature-grammar by anyone who did not wish to adopt the more esoteric aspects of GPSG theory). For example the category of a plural noun can be represented as:</Paragraph> <Paragraph position="4"> All features used in the word grammar (and lexical entries) must be declared to the analyser system. There are two types of features, atomic-valued and categoryvalued. Atomic-valued features must be declared with 294 Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 Graeme D. Ritchie, Stephen G. Pulman, Alan W. Black, and Graham J. Russell A Framework for Lexical Description an enumerated set of atomic values. Category-valued features can take any valid category as their value.</Paragraph> <Paragraph position="5"> These are declared using the keyword category, e.g.</Paragraph> <Paragraph position="6"> Feature N {+,-} Feature BAR {-1,0,1,2} Feature AGR category Although our sample English description uses particular feature names, there is no need for the linguist to copy such conventions. There is only one restriction on the features declared. If a feature of the name STEM is declared, it must be a category-valued feature. This feature is used by the WSister Convention (see below) and should not be used in any other way.</Paragraph> <Paragraph position="7"> The word grammar is a feature unification grammar with rules of the form: mother-> daughterl, daughter2 ..... daughterN where mother, daughterl, daughter2, etc. are categories made up of features. Rules may have one or more daughters. In addition to simple categories the grammar may also contain variables and aliases (see below).</Paragraph> <Paragraph position="8"> Aliases are a short-hand for writing categories (and parts of categories). They allow an atomic name to be associated with a category, and hence then be used to represent that category in a rule. For example the aliases Noun and Verb might be declared as: Alias Noun = ((BAR 0) (N +) (V -)) Alias Verb = ((BAR 0) (N -) (V +)) There are two types of variables allowed within the categories in the grammar; &quot;rule-category variables&quot; and &quot;feature value variables&quot;. Rule-category variables range over specific categories, and are a short-hand for writing similar grammar rules. They are declared with a range of possible values that must be stated as a list of aliases. Rule-category variables can be used to capture generalisations in rules. For example, in French both nouns and adjectives can take a plural morpheme s (which can be represented by the category ((PLU +)) ).</Paragraph> <Paragraph position="9"> This phenomenon could be described using the following alias statements and rules: Rule-category variables are &quot;compiled out&quot; during grammar compilation, and are thus actually used to collapse a number of rules.</Paragraph> <Paragraph position="10"> Feature value variables, on the other hand, can best be thought of as &quot;holes&quot; that are filled in during parsing (although theoretically they have equivalent semantics to rule-category variables, if we overlook the distinction between abbreviations for finite sets and for infinite sets). There are two types of feature value variables b atomic-valued and category-valued (category-valued variables are not the same as rule-category variables). The distinction is analogous to that between the atomic-valued features and category-valued features described above. Atomic-valued variables are declared with an enumerated set of values, while category-valued variables are declared with the keyword category: Variable ALPHA = {+,-} Variable ?AGR = category Feature value variables are not compiled out at grammar compile time but are instantiated during parsing. The ranges of feature value variables can be used to restrict the scope of rules. They can also be used to &quot;copy&quot; values of features up (and down) the parse tree. For example, a compound noun can be said to inherit its plural feature marking from the rightmost daughter.</Paragraph> <Paragraph position="11"> Using feature value variables we can write a rule that ensures that the compound noun will have the same PLU marking as its rightmost daughter:</Paragraph> <Paragraph position="13"> Note that although atomic-valued variables can be thought of as a short-hand for a number of rules, one for each value in the range of the variable, category-valued variables cannot. This is because there is potentially an infinite number of categories that could be the value of a category-valued feature.</Paragraph> <Paragraph position="14"> There are no typographical conventions built-in for specifying variables; the rule-writer, however, may wish to adopt some convention such as starting all variables with underscore or question mark. This does make rules easier to read but is in no way mandatory.</Paragraph> <Paragraph position="15"> In addition to the use of variables for &quot;passing&quot; features around during parsing there are some built-in feature-passing conventions (see below for more details).</Paragraph> <Paragraph position="16"> Before a description of what constitutes a valid analysis can be given two definitions are required.</Paragraph> <Paragraph position="17"> (a) A feature-value (either atomic or a category) is an extension of any variable of an appropriate type.</Paragraph> <Paragraph position="18"> Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 295 Graeme D. Ritchie, Stephen G. Pulman, Alan W. Black, and Graham J. Russell A Framework for Lexical Description (b) An atomic feature-value is an extension of itself.</Paragraph> <Paragraph position="19"> (c) Category A is an extension of category B iff for any feature f in category B , there is a value of f in A which is an extension of the value of f in category B.</Paragraph> <Paragraph position="20"> Unification The unification of two categories is the smallest category that is an extension of both of them if such category exists. It is possible that no such category exists, and in that case unification is undefined.</Paragraph> <Paragraph position="21"> Intuitively, extension and unification can be though of as the set relation superset and the set operation union, respectively, with the extra refinement of allowing at most one entry for each feature within a category. The creation of the unification of two (or more) categories is referred to as &quot;unifying&quot; the categories.</Paragraph> <Paragraph position="22"> The morphological analyser uses the rules in the Word Grammar to find all possible structures for a given word. Each structure is a tree in which each node is EITHER: the keyword ENTRY and a lexical entry OR: a local tree of the form N -> cl c2 . . . cn, where N is a category and ci is a constituent. This tree must match the following constraints a. there must be a rule in the word grammar of the form A -> dl d2 . . . dn, where category N is an extension of A and ci is an extension ofdi for each i from 1 to n .</Paragraph> <Paragraph position="23"> b. the local tree must be valid with respect to the feature-passing conventions.</Paragraph> <Paragraph position="24"> The analyser returns all constituents that span the given string and have a category that is an extension of the category which the linguist has defined to be the distinguished category. That is, in the same way that a traditional context-free grammar has a single distinguished symbol which is used to define complete derivations, our morphological model has a distinguished category.</Paragraph> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> 5. FEATURE-PASSING CONVENTIONS </SectionTitle> <Paragraph position="0"> Feature-passing conventions can be thought of as a way of extracting various patterns which occur in the word-grammar rules and stating them separately. The effect of this is to diminish the amount of explicit information that needs to be stated in the word-grammar rules, reducing both the size of the word-grammar (the number of rules) and the complexity of the individual rules. These regularities can be expressed as feature-passing conventions which can be thought of as rules for passing information UP the analysis tree (from terminal morphemes to the final word), or for passing information DOWN the analysis tree (from word to constituent morphemes). The way of stating these conventions is based on the mechanisms employed by Generalised Phrase Structure Grammar at the level of the sentence (Gazdar et all. (1985)), but the morphological generalisations embodied in them are essentially those of Selkirk(1982).</Paragraph> <Paragraph position="1"> There are three conventions built into the system at present. Notice that the definitions of the feature-passing conventions themselves are not under the control of the lexicon-writer, although the features that are affected by the conventions may be modified. The conventions act on certain specific features or feature-, classes, so the linguist can make use of these conventions by defining certain features to lie within these named classes. The system will then automatically apply the conventions to these features.</Paragraph> <Paragraph position="2"> All three feature conventions act on what is called within GPSG terminology a &quot;local tree&quot; -- a set of one mother node and its immediate daughters. The conventions were originally designed for binary branching rules (introducing exactly two daughters), but they apply to all rules. They are written in terms of &quot;mother&quot; , &quot;left daughter&quot; (i.e the leftmost daughter in a local tree) and &quot;right daughter&quot; (the rightmost daughter). In unary rules, those with just one category on the right-hand side, the left and right daughters are the same category.</Paragraph> </Section> <Section position="7" start_page="0" end_page="0" type="metho"> <SectionTitle> THE WORD-HEAD CONVENTION </SectionTitle> <Paragraph position="0"> The WHead feature-values in the mother should be the same as the WHead feature-values of the right daughter.</Paragraph> <Paragraph position="1"> In the word parser, this is achieved, roughly speaking, by unifying the WHead features of the right daughter and those of the mother when the daughter is attached. From a linguistic point of view, the WHead features are typically those that will be relevant to sentence-level syntax, and hence those that will be of particular use to a sentence-parser which uses the dictionary. This convention is a straightforward analogue of the simplest case of the Head Feature Convention in (Gazdar et al.</Paragraph> <Paragraph position="2"> (1985)). Its effect is to enforce identity of the relevant feature values between mother and the head daughter.</Paragraph> <Paragraph position="3"> Note that in the current system there is no formal definition of &quot;head&quot; to which the lexicon-writer has access (despite the name given to this convention), since the right daughter always acts in this head-like fashion within our treatment of English morphology.</Paragraph> <Paragraph position="4"> Other analyses may deviate from this pattern, of course; different views of &quot;head&quot; may be implemented using variables and unification.</Paragraph> <Paragraph position="5"> Assuming the set of WHead features includes N, V, PLU, and VFORM, the Word-Head Convention would allow the following trees:</Paragraph> <Paragraph position="7"> Graeme D. Ritchie, Stephen G. Pulman, Alan W. Black, and Graham J. Russell A Framework for Lexical Description but not (after all unification has occurred) trees of the form:</Paragraph> <Paragraph position="9"> since one of the trees has a clash in the V value for mother and right daughter, and the other lacks a VFORM marking on the mother to match that on the right daughter.</Paragraph> </Section> <Section position="8" start_page="0" end_page="0" type="metho"> <SectionTitle> THE WORD-DAUGHTER CONVENTION </SectionTitle> <Paragraph position="0"> (a) If any WDaughter features exist on the right daughter then the WDaughter features on the mother should be the same as the WDaughter features on the right daughter.</Paragraph> <Paragraph position="1"> (b) If no WDaughter features exist on the right daughter then the WDaughter features on the mother should be the same as the WDaughter features on the left daughter.</Paragraph> <Paragraph position="2"> Again, this is ensured by carrying out unification of the appropriate feature markings during parsing. This convention is designed to capture the fact that the subcategorization class of a word (in English) is not affected by inflectional affixation, although it may be affected by derivation.</Paragraph> <Paragraph position="3"> Assuming the feature SUBCAT to be the only WDaughter feature, this convention allows trees such In the first example the right daughter is specified for a SUBCAT value, and the mother has the same specification; in the second example, the right daughter has no specification for SUBCAT and so the second clause of the WDaughter convention applies. The third example is illegal because the values of SUBCAT on the right daughter and mother differ, and the fourth is illegal because, under clause (b) of the convention, the left daughter and mother WDaughter features must be identical when there are no WDaughter features in the right daughter.</Paragraph> </Section> <Section position="9" start_page="0" end_page="0" type="metho"> <SectionTitle> THE WORD-SISTER CONVENTION </SectionTitle> <Paragraph position="0"> When one daughter (either left or right) has the feature STEM, the category of the other daughter must be an extension (superset) of the category value of STEM.</Paragraph> <Paragraph position="1"> This third convention enables affixes to be subcategorized for the type of stem to which they attach. Notice that this convention is not defined in terms of any feature-classes, but is defined using just one &quot;built-in&quot; feature (STEM). Hence, the way that the lexicon-writer makes use of this convention is not by declaring the extent of feature-classes (as for the other two conventions), but by adding STEM specifications to the features in morphemes in the lexicon, thereby indicating the combination possibilities for each affix. The following examples follow the convention</Paragraph> <Paragraph position="3"/> </Section> <Section position="10" start_page="0" end_page="0" type="metho"> <SectionTitle> 6. FEATURE DEFAULTS </SectionTitle> <Paragraph position="0"> Feature Defaults are similar in concept to the Feature Specification Defaults of Gazdar et al. (1985). They are statements which define values for particular features in circumstances where no value has been entered by other mechanisms (i.e. the original morpheme entries, the action of the lexical rules, or the feature-passing conventions). That is, they state what the value of a feature should be if there is no information to indicate any other value for it. The defaults are applied to all new constituents (words or parts of words) built during morphological analysis. (In terms of the active chart implementation of the parsing mechanism, the default checking is done whenever a complete (inactive) edge is entered into the chart).</Paragraph> <Paragraph position="1"> At present, only very simple defaults are available, compared to the various kinds of defaults proposed (for sentence-level grammar) by Gazdar et al. (1985). All the linguist can do is define the default value for a given feature (either a category-valued feature or an atomic-valued one). For example, the statement Defaults BAR 0, AGR Inf declares default values for two features (BAR and AGR), where &quot;Inf&quot; could be an alias for some category. However, Completion Rules have arbitrary descriptive power, and can be used to achieve complex insertion of Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 297 Graeme D. Ritchie, Stephen G. Pulman, Alan W. Black, and Graham J. Russell A Framework for Lexicai Description default feature values, providing that the default-insertion can be performed adequately on individual morphemes (not categories formed by combining morphemes), since Completion Rules have their effect prior to morphological analysis. (See discussion in Section 8 below) 7. LEXICAL RULES All three types of rule (Completion Rules, Multiplication Rules, and Consistency Checks) have the same basic form: < pre-condition > < operator > < action > Although the < operator > and < action > are different in each type of rule, the syntax of the < pre-condition > is the same. Pre-conditions are specified as conjunctions of (possibly negated) patterns, describing lexical entries. Variables are denoted by atoms starting with an underscore e.g. fred, fix etc. and are bound during matching so that t--hey ca--ffbe used later in a match or in a rule action. There is a special variable consisting of only an underscore ( .... ), which never gets bound but can be used to match--anything (cf. Prolog). All other variables have a consistent interpretation throughout a rule. Matching is done from left to right (which is significant in the matching of syntactic fields). The entry being matched does not have to have the features in the same order as the pattern.</Paragraph> <Paragraph position="2"> The following examples of syntactic category matching illustrate some of the above points:</Paragraph> <Paragraph position="4"> with rest bound to ((PLU -) (INFL +))</Paragraph> <Paragraph position="6"> with rest bound to an empty list of features. The pattern ((N +) junk (V -)) does not match any syntactic category because the variable junk will match all remaining features in the category being checked.</Paragraph> <Paragraph position="7"> When negation is used no bindings that are made within a negative pattern are passed on through the match (agalin cf. Prolog), although bindings can be passed into negations.</Paragraph> <Paragraph position="8"> The above examples concern only the syntactic field, but pre-conditions match against entire entries. For example: -(be ) and</Paragraph> <Paragraph position="10"> would match all entries that do not have the citation form be and are marked with the features (N -) and</Paragraph> <Paragraph position="12"/> </Section> <Section position="11" start_page="0" end_page="0" type="metho"> <SectionTitle> COMPLETION RULES </SectionTitle> <Paragraph position="0"> Completion Rules are designed to be used to add values to tile entries that are specified by the linguist, and are applied in order to the entries (after aliases have been expanded). Accordingly, the order of the Completion Rules is significant. A Completion Rule is of the form < pre-condition > = > < entry skeleton > Ifa pre-condition matches an entry the entry is replaced with the newly constructed one described by the entry skeleton. A entry skeleton is of the same general form as a lexical entry, but various parts of it may contain the ampersand symbol (&), to mean &quot;the same as in the original entry&quot;, or variables which have appeared in the pre-condition (and hence would have been bound in the matching process).</Paragraph> <Paragraph position="1"> For example the rules:</Paragraph> <Paragraph position="3"> have the action of adding (BAR -1) to entries containing the feature FIX, adding (BAR 0) to all entries that do not have a BAR marking and lastly adding (INFL +) to all values of STEM that do not already have a marking for INFL. Note that the ordering of the first two rules is significant. If the first two rules were in the reverse order, the FIX rule would not apply to any entries, as all entries would by that time have had (BAR 0) added.</Paragraph> </Section> <Section position="12" start_page="0" end_page="0" type="metho"> <SectionTitle> MULTIPLICATION RULES </SectionTitle> <Paragraph position="0"> Multiplication Rules construct additional entries (as opposed to the replacement of entries performed by 298 Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 Graeme D. Ritchie, Stephen G. Pulman, Alan W. Black, and Graham J. Russell A Framework for Lexical Description Completion Rules). These are typically used to generate similar entries with slightly varying feature markings; for example, in English these rules can be used to generate the first person, second person and plural of verbs from the base form (as an alternative to designing the morphological rules to handle the verb paradigm).</Paragraph> <Paragraph position="1"> The syntax of these rules is very similar to that of the completion rules: < pre-condition > =>> ( < list of entry skeletons > ) The syntax of the entry skeletons is exactly the same as above. The ordering of the rules is not significant as newly created entries are not re-tested against the Multiplication Rules. This is to avoid possible infinite application of rules.</Paragraph> <Paragraph position="2"> A Multiplication Rule to generate the first and second person singular and plural of a base verb could be: tion of a category that contains a V marking but no N marking, but the linguist may wish to specify that such a category is invalid. Consistency Checks are statements of the form: < pre-condition > demands < post-condition > The <post-condition> has the same syntax as the pre-conditions. The interpretation is: If an entry_matches the pre-condition it must also match the post-condition.</Paragraph> <Paragraph position="3"> For instance, if ali entries that are marked for V must also be marked for N and vice versa then this condition can be written as the following two Consistency Checks:</Paragraph> <Paragraph position="5"> The rules are applied in the following order: Multiplication Rules (order is not significant), Completion Rules (in order of specification), and finally the Consistency Checks are applied to each entry created by the previous rule applications.</Paragraph> </Section> <Section position="13" start_page="0" end_page="0" type="metho"> <SectionTitle> 8. DESCRIPTION OF ENGLISH </SectionTitle> <Paragraph position="0"> Note that the entry being tested is not replaced but remains in the lexicon (assuming the Consistency Checks are passed -- see below). So, given the entry</Paragraph> <Paragraph position="2"> four entries would exist after the application of the Multiplication Rule, having the form:</Paragraph> <Paragraph position="4"/> </Section> <Section position="14" start_page="0" end_page="0" type="metho"> <SectionTitle> CONSISTENCY CHECKS </SectionTitle> <Paragraph position="0"> Consistency Checks are applied to every entry (including newly created ones) after the above two sets of rules have applied. Any entry that does not pass these tests is not included in the lexicon. The only formal requirement on lexical entries is that entries are quintuples and that the syntactic field is a set of feature pairs with values as declared. These Consistency Checks allow the linguist to check linguistic dependencies within entries; for example, there is no built-in prohibi-The mechanisms outlined in the preceding sections could be used to construct almost any description of English lexical facts. Here we sketch one such description, which we have developed using the mechanisms described here. It is worth observing in passing that the features used in the description can be broadly grouped into the following (overlapping) classes: Purely sentential. These features are included as part of the grammatical description of sentence structure, and do not have any particular import within morphological rules. For example, the feature SUBCAT is used to indicate the subcategorisation of verbs, etc. but does not affect spelling or word-structure.</Paragraph> <Paragraph position="1"> Features which do not affect morphology may still be manipulated by the morphological rules, since the feature-passing conventions will cause various features to be passed from morphemes to words. Thus whole words will inherit features from their component morphemes, without the rules mentioning the features explicitly.</Paragraph> <Paragraph position="2"> Sentential with morphological effects. The features V and N (for classifying nouns, verbs, adjectives and prepositions, in the GPSG style), although obviously motivated by the syntactic form of sentences, also affect various affixing processes.</Paragraph> <Paragraph position="3"> Purely morphological. Certain features have been postulated in our description solely to distinguish classes of morpheme that have different behaviour morphologically. For example, the feature FIX (with possible values PRE and SUF) indicates an affix, and the Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 299 Graeme D. Ritchie, Stephen G. Pulman, Alan W. Black, and Graham J. Russell A Framework for Lexicai Description feature INFL (possible values + and -) indicates whether a word or morpheme is capable of further inflection.</Paragraph> <Paragraph position="4"> Notice that this is not a formal distinction, and does not correspond to any sub-divisions in our mechanisms i it is merely an observation about our description of English that certain features are not motivated by morphological considerations. In a sense, they could still be said to &quot;affect&quot; the morphological processing, since if a feature is mentioned in the STEM value of a morpheme, it will restrict possible morpheme combinations. The sentential features (i.e. the first two classes above) have been devised in collaboration with the writers of a medium-sized grammar of English (Briscoe et al. 1986), but we shall not discuss here the justifications for the decisions regarding sentential grammar. Appendix I outlines the usage of the more morphologically or lexically significant features.</Paragraph> <Paragraph position="5"> The Word Grammar describing inflectional and derivational morphology is not large; the complete set is given below. Each rule is preceded by a mnemonic name, and VAL is a variable ranging over + and -.</Paragraph> <Paragraph position="6"> Since the feature-marking (BAR 0) indicates a item which constitutes a whole word, the PREFIXING rule can be summarised as &quot;Any word can be made up of a prefix followed by another valid word&quot;. (Properties of the prefix and stem determine the full features of the word by the Feature-Passing Conventions -- see below).</Paragraph> <Paragraph position="8"> The SUFFIXING rule can be phrased &quot;Any noun or adjective can be made up of a noun or adjective stem followed by a suffix&quot;. Notice that this rule does not stipulate that the stem must be of the same major category (noun or adjective) as the overall word, and hence it covers derivational morphology (where the category is altered by affixation) as well as noun inflections. The restriction to nouns or adjectives (i.e. entities marked as (N +)) is necessary as verbs require the slightly more detailed rules V-SUFFIXING and NON-V-SUFFIXING (and prepositions -- ((N -) (V -)) -- do not take affixes at all).</Paragraph> <Paragraph position="9"> The V-SUFFIXING rule states &quot;that a verb can be made up of a verbal stem of the same auxiliary marking followed by a verbal suffix&quot;. This is to cover general verb inflection, for both auxiliaries (AUX +) and main verbs (AUX -).</Paragraph> <Paragraph position="10"> The NON-V-SUFFIXING rule is to cover those cases of derivational morphology where a noun or adjective (N +) stem becomes a verb through suffixation -- &quot;any noun or adjective which forms a whole word can form a whole word verb by the addition of a verbal suffix&quot;.</Paragraph> <Paragraph position="11"> In all these cases, the rules may seem to be rather sketchy and lacking in feature specifications. For example, the PREFIXING rule does not stipulate much about the relationship between stem and composite word, and therefore seems to omit the generalisation that prefixation does not alter the grammatical features of the word (in particular, the major category is the same). However, these highly economical grammar rules are made possible by the assumption that the various feature-passing conventions (and feature defaults) will ensure that features are correct. Hence it is crucial that the word grammar be assessed in conjunction with the feature-passing conventions defined in Section 5 above, and the following definitions of feature classes: WHead Features: N V INFL PAST AFORM VFORM BARE-ADJ</Paragraph> </Section> <Section position="15" start_page="0" end_page="0" type="metho"> <SectionTitle> ADV AGR PLU POSS FIN WDaughter Features: SUBCAT </SectionTitle> <Paragraph position="0"> The features in the WHead list will be forced to have the same values on the right-daughter and the mother; hence these feature-values when specified on a suffix will percolate on to the main word, and will also remain on the main word when a prefix is added. Similarly, the WDaughter feature will be inherited from the appropriate part of the word.</Paragraph> <Paragraph position="1"> There are also two feature-defaults:</Paragraph> </Section> <Section position="16" start_page="0" end_page="0" type="metho"> <SectionTitle> BAR 0 LAT + </SectionTitle> <Paragraph position="0"> These ensure that any morpheme, word, or part of a word which does not have a value for BAR will be marked as a potential whole word, and that any item not marked for being &quot;Latinate&quot; will be assumed to be so. There are two main types of lexical rules I Completion Rules and Multiplication Rules. The third type (Consistency Checks) are desirable for disciplined lexicon-writing, but they do not insert any features or entries, and will not be discussed here.</Paragraph> <Paragraph position="1"> Although it is not obvious from the outline of the formalism given here, Completion Rules can be used in several ways to control the content of lexical entries.</Paragraph> <Paragraph position="2"> 300 Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 Graeme D. Ritchie, Stephen G. Pulman, Alan W. Black, and Graham J. Russell A Framework for Lexical Description This is not to say that the notation has several interpretations, but rather that the lexicon writer can choose to employ the facilities of these rules in different styles, in much the same way that a programmer can use a programming language in different manners. We will term the three main usages overwriting, obligatory insertion, and optional defaulting. The overwriting use is fairly straightforward -- in this, a rule is used to alter the values of one or more specified features. For example, a rule like the following would change all adjectives to nouns:</Paragraph> <Paragraph position="4"> This is achieved by mentioning the relevant features explicitly in the pre-condition (pattern), and supplying a new form (right-hand side) which has an explicitly stated revised form of them; all other features are carried over unchanged by the variable rest. Leaving aside the rather absurd content of the ~-ove example, we have made no use of this effect in our description of English, as there seemed no point in putting in entries which were to be systematically altered by rule (notice that this is different from having the entries filtered out by some rule such as a Consistency Check).</Paragraph> <Paragraph position="5"> The notion of obligatory insertion is more subtle.</Paragraph> <Paragraph position="6"> This involves writing rules which will insert a feature value if it is not there already, but will result in the lexical entry being discarded if that feature is already specified in it. For example, consider a rule which adds (INFL -) to all entries marked with an AFORM value:</Paragraph> <Paragraph position="8"> The Completion Rule mechanism is not defined to check for the presence of features it is attempting to add, so this rule will attempt to add (INFL -) even if the entry already has a value for INFL (whether + or -). If there is no previous marking for INFL, this insertion will be successful, and the rule will have effectively added a default marking. If there is some previous marking, the insertion will fail, since the mechanism is not defined to overwrite entries, and the lexical entry will be discarded. null Optional defaulting is a slightly more circumspect way of inserting default values. Consider the following rule, which also inserts (INFL -) as a default marking on entries specified for AFORM.</Paragraph> <Paragraph position="9"> In this version, the pre-condition has an explicit check for the absence of an INFL marking (since the tilde sign indicates negation of the immediately following condition). This rule will apply only to entries which are unspecified for INFL, and will have no effect (i.e. leave in the lexicon unaltered) entries for which this pre-condition is not true.</Paragraph> <Paragraph position="10"> All the Completion Rules in our description of English are written to act in a &quot;default&quot; manner; in fact, to use the terminology introduced immediately above, they are written in the &quot;optional default&quot; style, in that each rule checks for the absence of a feature-value before inserting the value. The full set of Completion Rules is given in Appendix 3.</Paragraph> <Paragraph position="11"> There are two Multiplication Rules. The first is to account for the fact that any noun or adjective which is subcategorised for a complement (e.g. critic takes a prepositional phrase with 039 can also occur with no complement:</Paragraph> <Paragraph position="13"> The second clause of the pre-condition is necessary simply because the pattern-matching mechanism does not permit the use of negated feature values (e.g.</Paragraph> <Paragraph position="14"> (SUBCAT &quot;NULL)) within a pattern. The second Multiplication Rule expands the present tense verb paradigm for all verbs apart from be, by adding three further lexical entries per verb: &quot;(be ) and</Paragraph> <Paragraph position="16"> It might seem that this regularity would be more naturally handled in the morphological analyser, rather than as lexical redundancy. However, that would necessitate the introduction of a morpheme whose entire surface form was null (with suitable syntactic features). The complications this would introduce into the morphographemic segmentation and word-grammar parsing are Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 301 Graeme D. Ritchie, Stephen G. Pulman, Alan W. Black, and Graham J. Russell A Framework for Lexicai Description regarding as wholly unacceptable computationally, and so the Multiplication Rules have been used to capture this generalisation. (This is a very obvious example of the methodological issue mentioned in our opening section, concerning the need for a viable procedural interpretation of the whole set of mechanisms).</Paragraph> <Paragraph position="17"> There are 15 spelling rules in our description. Appendix 2 contains an annotated list of them. In addition to the actual rules the spelling rule mechanism requires the definition of the lexical and surface alphabets. The Surface Alphabet contains all normal alphabetic letters, space, hyphen and apostrophe (for simplicity, we shall ignore the issue of upper and lower case here). The Lexical Alphabet contains exactly the same symbols together with the plus sign (+) which we use to mark morpheme boundaries. The null symbol (0) is a part of the formalism, and hence is not regarded as part of either alphabet (but may occur in rules anywhere that a normal alphabet symbol might occur). In addition to standard identity pairs made from the intersection of the lexical and surface alphabets, three default pairs are declared, so that these pairs are valid in any context during matching.</Paragraph> <Paragraph position="18"> +:0 a morpheme boundary symbol may be deleted on the surface.</Paragraph> <Paragraph position="19"> -:0 hyphens in a citation form (e.g. data-base) may be absent from the surface string (e.g. database).</Paragraph> <Paragraph position="20"> .... :- where a lexical form has a space (e.g. data base ), the surface strings may optionally contain a hyphen instead ( data-base ).</Paragraph> </Section> <Section position="17" start_page="0" end_page="0" type="metho"> <SectionTitle> 9. AN EXAMPLE </SectionTitle> <Paragraph position="0"> The above word grammar, and various other parts of the rules and definitions, can be illustrated with a simple example u the analysis of the word applications. The Spelling Rule interpreter will segment this (using the C-Insertion rule, and the Default Pair definition that pairs morpheme boundaries with null) into three morphemes ---apply, +ation, and +s. In the original lexical entries, these morphemes are listed thus, with +s having two entries: However, various Completion Rules will have acted upon these basic entries at the pre-compilation stage of the lexicon, resulting in the following more detailed entries for the three morphemes we are interested in here (ignoring the other entry for +s): Seven rules achieve this -- one adds the marking (BAR -1) to entries marked as affixes (i.e. specified for FIX), another adds (BAR 0) to all entries which are specified for V and N but lack a BAR value, the third adds (INFL -) to all morphemes marked with (PLU +), the fourth adds (INFL +) to all (BAR 0) entries which lack an INFL value, the fifth adds (AUX -) to all verbs (but not verbal affixes), the sixth adds (LAT +) to any entry with a V marking, and the seventh adds (AT +) to all entries with a (LAT +) marking. Notice that the ordering of the Completion Rules in the description is crucial, for example the third of these rules affects the fourth.</Paragraph> <Paragraph position="1"> The SUFFIXING rule in the Word Grammar combines the first two morphemes into a subtree with the lexical entries for apply and +ation as daughter nodes. The left-hand side of this rule assigns the following syntactic category to the mother node: markings on the right daughter (+ation). The marking (LAT +) follows directly from the Feature Default, since these are added to all constituents found by the word-parser, not just to individual morphemes. The (SUBCAT NP__PPTO) is a result of the WDaughter Convention, because there is no SUBCAT feature on the right daughter it must be the same as that on the left. The SUFFIXING rule operates again, to combine this word (application) with the plural morpheme +s , to form a tree whose daughter categories are: Notice that this combination will accord with the WSister convention, since the category of the leftdaughter is an extension of the value of STEM on the</Paragraph> <Paragraph position="3"> of the mother node includes, from the SUFFIXING rule, the following markings:</Paragraph> <Paragraph position="5"> Again, further feature markings are then computed, giving:</Paragraph> <Paragraph position="7"> The markings (V -), (PLU +) and (INFL -) result from the WHead Convention, since V, PLU and INFL are WHead features and so must have the same values as on the right daughter (+s). The marking (LAT +) follows directly from the Feature Default and the SUBCAT feature is a result of the WDaughter Convention. The overall structure for the word can then be viewed as a tree in which each node is annotated with either a syntactic category and a rule-name or the keyword ENTRY and a morphemic lexical entry, as shown below.</Paragraph> </Section> <Section position="18" start_page="0" end_page="0" type="metho"> <SectionTitle> 10. CONCLUSIONS </SectionTitle> <Paragraph position="0"> We have presented an integrated set of formalisms for describing various aspects of the lexicon in a computationally tractable manner, which have been used to create a non-trivial description of English lexical phenomena. All these facilities have been implemented (in Franz Lisp on a Sun 2/120) and are being used as part of collaboration between Edinburgh, Cambridge and Lancaster universities to develop a set of software tools for natural language processing, under the Alvey Programme. It should be borne in mind that all the rule-formalisms are highly experimental, and if they are to form a useful linguistic theory (as opposed a practical software package) a great deal of refinement is required.</Paragraph> <Paragraph position="1"> Not only are some of them perhaps too powerful (e.g.</Paragraph> <Paragraph position="2"> Completion Rules have arbitrary computational power), some of them may be too weak descriptively (e.g. it is not clear if the morphological mechanisms are adequate for all languages). The description is also still under development; the rules given here reflect the state of the system in summer 1986.</Paragraph> </Section> <Section position="19" start_page="0" end_page="0" type="metho"> <SectionTitle> APPENDIX 1--LEXICALLY SIGNIFICANT SYNTACTIC FEATURES </SectionTitle> <Paragraph position="0"> The following are brief explanations of the features which are involved in the Completion Rules in Appendix 3, the Word Grammar in Section 8, or which are particularly pertinent to the Feature Passing Conventions. Each feature name is followed by a list of its allowable values.</Paragraph> </Section> <Section position="20" start_page="0" end_page="0" type="metho"> <SectionTitle> AT(- +) </SectionTitle> <Paragraph position="0"> Stems to which the suffixes +ation and +ative may attach are marked as (AT +), while those taking the corresponding forms +ion and +ive are (AT -). This specification is referred to in the STEM feature of the suffixes in question, resulting in action and presentation, but not e.g. presention.</Paragraph> <Paragraph position="1"> The feature LAT distinguishes those stems traditionally analysed as latinate from others. Certain affixes may only attach to latinate stems; +an is one such, giving magician, but not artan.</Paragraph> <Paragraph position="3"> This is a feature allowing us to refer to two disjoint category sets. Certain suffixes (e.g. +ly) may attach either to the base form of regular, inflectable, adjectives (as in easily), or to non-inflectable adjectives (as in dangerously). They may not, however, attach to inflected forms (easiestly); BARE-ADJ distinguishes e.g. easy and dangerous ((BARE-ADJ +)) from easiest, whereas INFL (see below) does not.</Paragraph> </Section> <Section position="21" start_page="0" end_page="0" type="metho"> <SectionTitle> FIX (PRE SUF) </SectionTitle> <Paragraph position="0"> All affixes bear a specification for FIX; prefixes have the value PRE, and suffixes have the value SUF. The rules of the word grammar refer to the specifications, so that prefixes always precede their stems and suffixes always follow.</Paragraph> <Paragraph position="2"> The feature AGR is responsible for enforcing the necessary correspondence between categories in sentence structure. This is most common in the case of subject and verb; is is specified as (AGR SING3), and am as (AGR SING1).</Paragraph> </Section> <Section position="22" start_page="0" end_page="0" type="metho"> <SectionTitle> POSS (+ -) </SectionTitle> <Paragraph position="0"> Distinguishes possessive items from others. His, whose, and the possessive's are specified as (POSS +).</Paragraph> </Section> <Section position="23" start_page="0" end_page="0" type="metho"> <SectionTitle> INFL (- +) </SectionTitle> <Paragraph position="0"> INFL distinguishes those stems which may bear an additional suffix (e.g. walk) from those which cannot (e.g. walking).</Paragraph> <Paragraph position="1"> STEM category The STEM feature controls the attachment of affixes to stems. The value of STEM in an affix category must (by the WSister feature passing convention) be included in the category of any stem that affix attaches to. In this way, +ing, for example, can be restricted to the base form of verbs.</Paragraph> <Paragraph position="2"> BAR(-10 1 2) The sentence grammar employs a three-level system of categories, various phrases being specified as (BAR 1) or (BAR 2), and preterminals as (BAR 0). In our analysis of word-structure, we extend this concept below the level of the complete word; stems are specified as (BAR 0), and affixes as (BAR -1).</Paragraph> <Paragraph position="3"> v(-+) N(-+) The major categories (nouns, verbs, adjectives, and prepositions) are classified by means of the features V and N. Verbs and adjectives, and their phrasal counterparts, are specified as (V +), while nouns and prepositions are specified as (V -). Nouns and adjectives, and their phrasal counterparts, are specified as (N +); verbs and prepositions are (N -).</Paragraph> <Paragraph position="5"> Determiners (articles like the, a and adjectives like all, three) are specified as (QUA +). Other adjectives are (QUA -).</Paragraph> </Section> <Section position="24" start_page="0" end_page="0" type="metho"> <SectionTitle> ADV (- +) </SectionTitle> <Paragraph position="0"> Adverbs derived from adjectives (quickly, easily) are analysed as adjectives bearing the specification</Paragraph> </Section> <Section position="25" start_page="0" end_page="0" type="metho"> <SectionTitle> (ADV +). AUX (- +) </SectionTitle> <Paragraph position="0"> Verbs are specified as (AUX +) if they are auxiliary verbs, and as (AUX -) otherwise.</Paragraph> <Paragraph position="2"> Verbs are specified as (FIN +) if they are finite (tensed), and as (FIN -) otherwise.</Paragraph> <Paragraph position="4"> Finite verbs are specified as (PAST +) if they are in the past tense, and as (PAST -) otherwise.</Paragraph> <Paragraph position="6"> NEG distinguishes negative words from others; aren't etc. are specified as (NEG +).</Paragraph> <Paragraph position="8"> PLU distinguishes plural nouns from others; men and cats both bear the specification (PLU +), and man and cat (PLU -).</Paragraph> <Paragraph position="10"> Determiners are specified for DEF; the is (DEF +) and a is (DEF -).</Paragraph> </Section> <Section position="26" start_page="0" end_page="0" type="metho"> <SectionTitle> AFORM (ER EST NONE) </SectionTitle> <Paragraph position="0"> AFORM encodes information concerning adjective morphology. Comparatives and superlatives are specified as (AFORM ER) and (AFORM EST) respectively, and non-inflectable adjectives are</Paragraph> </Section> <Section position="27" start_page="0" end_page="0" type="metho"> <SectionTitle> (AFORM NONE). NFORM (IT THERE NORM) </SectionTitle> <Paragraph position="0"> NFORM encodes the type of a noun phrase. The &quot;dummy subjects&quot; it and there are specified as (NFORM IT) and (NFORM THERE) respectively, while other NPs are (NFORM NORM). Certain verbs can then be associated with one type of NP by means of their AGR feature.</Paragraph> </Section> <Section position="28" start_page="0" end_page="0" type="metho"> <SectionTitle> VFORM (BSE EN ING TO) </SectionTitle> <Paragraph position="0"> VFORM encodes verb morphology. &quot;Bare infinitives&quot; are (VFORM BSE), passives and past participles are (VFORM EN), gerunds and present participles are (VFORM ING), and the infinitive to is SUBCAT encodes the subcategorization class of a word. Elapse, like other purely intransitive verbs is (SUBCAT NULL), and devour, like other transitives, is (SUBCAT NP). Many minor category items are their own subcategory; and has the specification</Paragraph> </Section> <Section position="29" start_page="0" end_page="0" type="metho"> <SectionTitle> (SUBCAT AND), etc. APPENDIX 2-----SPELLING RULES </SectionTitle> <Paragraph position="0"> The first five rules are based on those described in Karttunen and Wittenburg (1983). There are 15 rules in total: Epenthesis +:e <=> { < { s:s h:h > S:S y:i }--- s:s This allows the insertion of an e at a morpheme boundary, before an s and preceded by either sh, ch, one of s, x or z, or a y/i pair as in fly~flies. Gemination +:X <=> < C:C V:V =:X --- V:V where X in { b d f g 1 m n p r s t } This deals with doubling of consonants in words like bigger, travelling, etc.</Paragraph> <Paragraph position="2"> This rule deals with changing a lexical y to an i as in applies (Note that this requires both this rule and the Epenthesis rule above as there is an e insertion too).</Paragraph> <Paragraph position="3"> The c:t change is to cope with words like democratic (from democracy+ic). The second clause of the Y-to-I rule deals with words like application.</Paragraph> <Paragraph position="4"> This deals with e-deletion in words like moved (move+ed). The second clause deals with words ending in two vowels like agreed. The third clause deals with hard and soft g's and c's so that an e must be deleted in faced but not in advantageous (if it were the g would become hard).</Paragraph> <Paragraph position="5"> I-to-Y i:y <=> =: .... <e:0 +:0 i:i > This handles words like dying and lying.</Paragraph> <Paragraph position="6"> C-insertion +:c <=> y:i --- < a:a { t:t 1:1 b:b } > This is required for application. Note that this rule requires a clause in the Y-to-I rule so that the the y:i in this left context is allowed by the Y-to-I rule. K-insertion +:k <=> < V:V c:c > --- { < i:i n:n > e:e y:y } This caters for words ending in c that keep a hard c when a suffix is added e.g. panicky and picnicking while it does not require k insertion for words like criticise. This covers words like pronounciation, gracious, spacious etc. (The NB (&quot;not b&quot;) set is to stop the able suffix from being affected.</Paragraph> <Paragraph position="7"> I-insertion + :i < = > < C:C Vp:Vp NLR:NLR > --- <a:a { n:n 1:1 } > Examples of this are baronial, academician, civilian, dictatorial (This could be extended to cope with adverbial, etc).</Paragraph> <Paragraph position="8"> These above two rules deal with matching ability to able+ity as in probability, and similarly abilize to able+ize as in stabilize. These rules are an interesting example of how to deal with a change that happens over several characters. They deal With matching bilO to ble + where the il matches le.</Paragraph> </Section> class="xml-element"></Paper>