File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/concl/87/p87-1026_concl.xml
Size: 18,422 bytes
Last Modified: 2025-10-06 13:56:15
<?xml version="1.0" standalone="yes"?> <Paper uid="P87-1026"> <Title>FLUSH: A Flexible Lexicon Design</Title> <Section position="5" start_page="186" end_page="189" type="concl"> <SectionTitle> 4. Lingulstic/Conceptual Relations </SectionTitle> <Paragraph position="0"> These are expressions that cannot be easily handled as exclusively linguistic constructs, such as &quot;giving permission&quot;, &quot;getting permission&quot;, and &quot;having permission&quot;. These expressions can be represented as an abstract possession concept where the possessed is ':noun-permission&quot;, thus combining a class of concepts with a lexical category.</Paragraph> <Paragraph position="1"> These compound lexemes have the unique characteristic of allowing linguistic relations to have explicit conceptual constraints. In the phrase &quot;give a hug&quot; there is an abstract relationship between the concept of giving and the simple lexeme noun.hug that implies the concept of hugging. Figure 3 shows the representation of this linguistic/conceptual relation. This kind of compound lexeme is invoked by the semantic interpreter, rather than by the parser, during a process called concretion--making concepts more concrete. The scope of this paper does not permit a discussion of concretion, but refer to \[Jacobs, 1986b\] for more information.</Paragraph> <Paragraph position="2"> The descriptions in this section illustrate how FLUSH is able to represent a wide range of lexical phenomena in a hierarchical and uniform manner. The four classes of compound lexemes that are described encompass many of the usually problematic expressions in natural language, yet they are represented in a way that supports extension and adaptation. The next section describes how these representations are accessed by FLUSH.</Paragraph> <Paragraph position="3"> Although the compound lexeme representations illustrated in the previous section differ, FLUSH is able to employ a fairly flexible algorithm for accessing them. When the parser encounters a relation that may constitute a compound lexeme, it passes the name of the relation and the constituents that fill the appropriate roles to FLUSH. If FLUSH finds a compound lexeme that satisfies the constraints, it passes the lexeme back to the parser.</Paragraph> <Paragraph position="4"> For example, if TRUMP is working on the sentence &quot;:John picked up the book&quot;, it encounters a possible verb-particle relationship between the verb &quot;picked&quot; and the preposition &quot;up&quot;. When this relationship is apparent to the parser, FLUSH is called with the verb-part relation with the constituents of pt-verb.pick as the verb and prep-up as the particle:</Paragraph> <Paragraph position="6"> In this example, the compound lexeme verb.part-pickup is found by FLUSH and is returned to the parser. If instead the sentence is &quot;:John meditated up the hill&quot;, the parser takes the same action, but no compound lexeme is found by FLUSH because &quot;meditated up&quot; has no special meaning.</Paragraph> <Paragraph position="7"> FLUSH uses a two step procedure to locate specific compound lexemes. First, entries below the given relation in the hierarchy are checked to see if any of them satisfy the given constraints. If a compound lexeme exists, it is usually found during this step. There are some cases, however, in which the desired compound lexeme exists as a subcategory of an ancestor of the given relation. This situation was seen in the description of the modifying relation (rood-tel), verb-adjunct (va), and noun-post-modifier (npm) in the previous section (see Figure 2). In this case, a second step in the search process looks at the sibling categories. This process continues until either the top of the compound.lexeme hierarchy is reached (which happens immediately for most relations) or until a suitable compound lexeme is found.</Paragraph> <Paragraph position="8"> The process of finding a compound lexeme below the given relation is a matching problem. In response to the example call to find-compound above, the lexicon proceeds to look at the defined categories underneath verb-part, which include verb.part-C/ZZoUp, verb-part-C/xzout, verb-part-z~zx-off, etc., to see which one(s) satisfies the constraints, verb-part.zzz-up is found as a possibility, resulting in the same function being called recursively with the remaining constraints to find an appropriate category below it: (f ind-eompound verb-part-xxx-up (v-verb-part p~-verb-pick) ) This process is repeated until one of two conditions occurs: either the given constraints are exhausted, in which case a category that satisfies all of them has been found; or there are no more categories to search but there are still constraints left, in which case no match has been found and it may be appropriate to search the ancestors' subcategories. In this example, the verb-part-pick-up category is found and returned on the second recursion, therefore, there is no need to search the hierarchy at a higher level. If instead the parser is working in the sentence &quot;The man arrived from New York&quot;, it encounters a possible verb-adjunct (va) relation between the verb &quot;arrived&quot; and the prepositional phrase &quot;from New York&quot;. The lexicon is called with the va relation, but the first step in the search process (i.e., looking below the given relation) does not yield a compound lexeme because mod-rel-zxx-from is defined in terms of the rood.tel relation rather than in terms of the va relation (see Figure 2). So even though the relation that the parser encounters in the pattern is a verb-adjunct relation, the lexicon is flexible enough that it can apply more general knowledge to the retrieval problem.</Paragraph> <Paragraph position="9"> The meanings of compound lexemes are represented and accessed using a reference pointer that links the linguistic category to a conceptual structure. Some of the conceptual reference pointers for compound lexemes are more complicated than simple lexical access because often there are several components that need to be mapped, but they are still defined in terms of the ref association \[Jacobs, 1986a\]. The example form below defines a reference from the compound lexeme mod-rel-zxz-from to the transfer-event concept: (ref transfer-event <-> mod-rel-xxx-from (source <-> m-mod-rel-xxx-from)) This reference establishes that the modifying relation mod-rel-zzx-from should invoke the transfer-event concept, and the modifier part of mod-rel-zzx-from, namely m-modrel-zxz-from, should fill the role of source in this transferevent. In the sentence &quot;The man arrived from New York&quot;, the prepositional phrase &quot;from New York&quot; invokes rood. rel-zxx-from. In turn, the transfer-event concept is invoked with &quot;New York&quot; as the source of the transfer.</Paragraph> <Paragraph position="10"> The explanations above illustrate that FLUSH is capable of representing and accessing most of the different types of lexical knowledge that natural language processing systems need to have. They also show how FLUSH can do most of it in a general manner, making extensions fairly straightforward. FLUSH is equipped also with a mechanism for automatic acquisition of new lexemes, described in \[Besemer, 1986\]. The discussion that follows concentrates on the application of the hierarchical lexicon to semantic interpretation in TRUMP.</Paragraph> <Paragraph position="11"> III. Semantic Interpretation using FLUSH Section II. described the organization of the FLUSH lexicon, distinguishing several classes of lexical knowledge and showing the use of a hierarchical knowledge representation in representing examples of each class. One goal of this hierarchical organization is parsimony: because categories of compound lexemes inherit their constraints from more general categories, the number of linguistic constraints encoded explicitly can be reduced. A second function of the hierarchical representation, perhaps more important, is to facilitate the interpretation of the meaning of a compound lexeme.</Paragraph> <Paragraph position="12"> Semantic interpretation is facilitated by each of the classes of compound lexemes discussed in section II.. The simple example of word sequences allows the semantic interpreter to set aside the meanings of the individual words to interpret phrases such as &quot;by and large&quot; and '~C/ick the bucket&quot; correctly. Lexical relations, such as &quot;pick up&quot; and &quot;working directory&quot;, permit the association of specialized meanings as well as the contribution of certain flexible lexical classes to the meaning of a phrase. For example, the phrase &quot;branch manager&quot; is interpreted using knowledge that it belongs to a lexical category common with &quot;lab manager&quot; and &quot;program manager&quot;. Linguistic relations such as mod-rel-~zx-fram permit general lexical knowledge to apply to the filling of conceptual roles. Linguistic/conceptual relations such as let-give-hug permit the specialized interpretation of expressions such as &quot;give a hug&quot; in a broad range of surface forms.</Paragraph> <Paragraph position="13"> The following examples illustrate the operation of the TRUMP semantic interpreter and its use of the FLUSH lexicon. null</Paragraph> <Section position="1" start_page="189" end_page="189" type="sub_section"> <SectionTitle> Example 1: </SectionTitle> <Paragraph position="0"> Send the laser printer characteristics to the branch manager.</Paragraph> <Paragraph position="1"> Processing the above sentence stimulates a steady flow of information between TRUMP'S parser and semantic interpreter and the FLUSH lexical access mechanism. The lexical analyzer recognizes &quot;laser&quot;, &quot;printer&quot; and &quot;characteristics&quot; as nouns, but the search for compound lexical entries is activated only as the parser recognizes that the nouns form a compound. The specific entry for &quot;laser printer&quot; in the FLUSH lexicon, returned using the compound access method described in the previous section, provides two important pieces of information to TRUMP: First, it gives the semantic interpreter the correct meaning of the phrase, permitting TRUMP to forbear consideration of interpretations such as &quot;a printer that prints lasers&quot;. Second, it enables the parser to favor the grouping \[\[laser printer\] characteristics\] over \[laser \[printer characteristics\]\] and thus come up with a viable meaning for the entire phrase.</Paragraph> <Paragraph position="2"> The handling of the relationship between &quot;characteristics&quot; and &quot;laser printer&quot; makes use of the middle-level category en-~xx.characteristic, much like the verbpar~icle.~-up category described in section II. The cn-XZXocharac~eris~ic category, representing compound nominals whose second noun is &quot;characteristic&quot;, is associated with its meaning via a I%EF link in the following way: (ref characteristic <->. cn-xxx-charac~eristic (manifes~er <-> In-cn-xxx-charac~eris~ic)) The above association, in which ln.cn.~:zz-charac~er~stic denotes the first noun of a particular nominal compound, suggests the interpretation &quot;characteristics of the laser printer&quot;. The treatment of this association as a middle-level node in the hierarchical lexicon, rather than as an independent lexical entry, has two features: First, it is often overridden by a more specific entry, as in &quot;performance characteristics&quot;. Second, it may cooperate with more specific lexical or conceptual information. For example, the conceptual role manifesIer is a general one that, when applied to a more specific category, can lead to a specific interpretation without requiring a separate conceptual entry. This would happen with &quot;laser printer performance characteristics&quot;. The phrase &quot;branch manager&quot;, like &quot;laser printer characteristics&quot;, is interpreted using an intermediate entry en.zzx-manager. While FLUSH has the capability, like PHRAN \[Wilensky and Arens, 1980b\], to constrain this category with the semantic constraint that the first noun must describe a bureaucratic unit, it is at present left to the semantic interpreter to determine whether the preceding noun can play such an organizational role.</Paragraph> <Paragraph position="3"> Cancel the transmission to the printer.</Paragraph> <Paragraph position="4"> In this example, the lexical access mechanism must determine that &quot;to the printer&quot; invoked the mod-rel-~zzto linguistic relation, which can be attached either to the verb &quot;cancel&quot; or the nominal &quot;transmission&quot;. The semantic interpreter then finds the following association: (ref ~rans~er-even~ <-> mod-rel-xxx-~o</Paragraph> <Paragraph position="6"> The REF association above indicates that the object of the preposition &quot;to&quot; is related to the destination role of some generalized transfer event. Since &quot;cancel&quot; describes no such event, but &quot;transmission&quot; does, TRUMP correctly interprets &quot;printer&quot; as being the destination of the transmission. This allows the semantic interpreter to handle this example much in the same way as it would handle '`Transrnit the job to the printer n, because the rood-tel relation class includes both postnominal modifiers and adverbial prepositional phrases. As in the previous example, the semantic interpreter can make use of the interaction between this general interpretation rule and more specific knowledge; for example, &quot;the sale of the the book to Mar!f' invokes the same mod-rel.xxx-to relation, but the role of Mary is determined to be customer because that role is the conceptual specialization of the destination of a transfer. The process of correctly determining a conceptual role using linguistic relations is described in \[Jacobs, 1987\].</Paragraph> <Paragraph position="7"> Example 3: How many arguments does the command take? There are two major differences between this example and the previous two: First, the lexicon is driven by information passed from TRUMP~S semantic interpreter, not only from the parser. In the previous example, the parser recognizes a potential relationship between a verb or nominal and a prepositional phrase. In this case, the semantic interpreter must determine if the conceptual relationship between the concept of taking and the term &quot;arguments&quot; invokes any special lexical knowledge. Second, the interpretation of &quot;take arguments&quot; is not a specialization of an abstract concept such as transfer-event, but rather is a result of a metaphorical view mapping from this concept to the concept of command-execution.</Paragraph> <Paragraph position="8"> The interpretation of this sentence thus proceeds as follows: At the completion of the syntactic parse, the semantic interpreter produces an instantiation of the concept taking with the object arguments. The lexical access system of FLUSH, using the same discrimination process that determines a specialized linguistic relation, identifies Icr-transfer-arguments as a linguistic/conceptual relation invoked by the concept of a transfer with the lexical term &quot;argument&quot; attached to the conceptual object role. The same linguistic/conceptual relation is invoked by &quot;giving arguments&quot; or &quot;getting arguments&quot;. The semantic interpreter continues by determining the metaphorical mapping between the transfer-event concept and the command-execution concept, a mapping that derives from the same conceptual relationships as other similar metaphors such as &quot;The recipe takes three cups of sugar.&quot; In this way the amount of specialized information used for &quot;take arguments&quot; is kept to a minimum; effectively, FLUSH in this case is merely recognizing a linguistic/conceptual trigger for a general metaphor.</Paragraph> <Paragraph position="9"> This section has described the application of the FLUSH lexicon to the process of semantic interpretation in the TI~UMP system. The examples illustrate some characteristics of the flexible lexicon design that differ from other phrasal systems: (1) There are a broad range of categories to which specialized information may be associated. The treatment of &quot;branch manager&quot; and &quot;transmission to&quot; illustrates the use of compound lexical knowledge at a more abstract level than other programs such as PHRAN. (2) The hierarchical lexicon reduces the number of phrasal entries that would be required in a more rigid system. Expressions such as &quot;take arguments&quot; and &quot;get arguments&quot; share a common entry. (3) The quantity of information in each phrasal entry is minimized. Linguistic constraints are often inherited from general categories, and the amount of semantic information required for a specialized entry is controlled by the method of determining an appropriate conceptual role. The &quot;take arguments&quot; expression thus does not require explicit representation of the relationships between linguistic and conceptual roles.</Paragraph> <Paragraph position="10"> IV. Conclusion FLUSH is a flexible lexicon designed to represent linguistic constructs for natural language processing in an extensible manner. The hierarchical Organization of FLUSH, along with the provision for a number of types of phrasal constructs, makes it easy to use knowledge at various levels in the lexical hierarchy. This design has the advantage of handling specialized linguistic constructs without being too rigid to deal with the range of forms in which these constructs may appear, and facilitates the addition of new constructs to the lexicon. FLUSH permits the correct semantic interpretation of a broad range of expressions without excessive knowledge at the level of specific phrases.</Paragraph> </Section> </Section> class="xml-element"></Paper>