File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/90/h90-1047_metho.xml
Size: 17,113 bytes
Last Modified: 2025-10-06 14:12:30
<?xml version="1.0" standalone="yes"?> <Paper uid="H90-1047"> <Title>Syntactic and Semantic Knowledge in the DELPHI Unification Grammar</Title> <Section position="2" start_page="231" end_page="233" type="metho"> <SectionTitle> PP Interpretation </SectionTitle> <Paragraph position="0"> Prepositional phrases, in both post-copular and post-nominal positions, are very common in the ATIS domain (and most other domains as well). Some examples: Which flights are on Delta Airlines Which flights are on Thursday Which flights are after 4 pm The role in our grammar which generates a post-copular</Paragraph> <Paragraph position="2"> PREDICATIVE-PP is the constraint relation in the role. It is responsible for specifying the formula meaning of the VP in terms of the translation of the PP (:PP) and the translation of the subject passed down from the clause (:SUB J).</Paragraph> <Paragraph position="3"> The PREDICATIVE-PP solution for the &quot;flight-on-airline&quot; sense is as follows: The first occurences of the variables :NP and :SUBJ above are paired with semantic types AIRLINE and FLIGHT; this is shorthand for me actual state of affairs in which a term representing a package of information (including encoding of semantic type) appears in the slots of the role these variables occur in. This term carries quantification and semantic type information, as below:</Paragraph> <Paragraph position="5"> This structure is so constructed as to not unify with another such structure if its semantic type is disjoint, using a method for encoding semantic types as terms described in \[20\].</Paragraph> <Paragraph position="6"> The PP translation is also a package with the functor PP-SEM, containing the preposition and the translation of the NP object of the PP. No local attempt is made to translate the preposition.</Paragraph> <Paragraph position="7"> When parsing with the above predicate PP role, the system searches through a database of PREDICATIVE-PP solutions like the above, much as a PROLOG-based system would.</Paragraph> <Paragraph position="8"> If a solution succesfully unifies, the formula is passed up as the translation of the VP. Recursion is allowed, as in: This rule says that any PP relating a flight to a time should be translated as if it related the departure time of the flight to that time. In this way, a common system of PREDICATIVE-PP solutions stipulating the meanings of various prepositional comparisons between times (&quot;after&quot;, &quot;before&quot;, &quot;on&quot;, &quot;during&quot;) can be specitied just once, and used in multiple contexts.</Paragraph> <Paragraph position="9"> Such a method for interpreting PPs can be compared to one that uses unification over features of constituent elements instead of constraint relations, such as proposed in \[13\] and \[8\]. There, the multiple meanings of a preposition are enumerated locally as a semantic feature of the preposition, using a form of full-scale disjunction. Type constraints filter out those which are not meaningful in the given context, just as in our work.</Paragraph> <Paragraph position="10"> We claim, however, that the constraint relation method is superior exactly in view of its recursive power as described above. In order to handle such constructions as &quot;flight after 4pm&quot;, &quot;flight before 4 pm&quot; etc. a strictly unification-based approach would have to compile out all the possibilities for the type of the clause subject (TIME, FLIGHT etc.) and store them as disjunctive values of the translation feature on the given preposition. Thus in the ATIS domain there would be (at least) two senses of the preposition &quot;after&quot;, both expressing the &quot;time-after&quot; relation. Clearly such an approach would not capture the general relations on times that these prepositions express.</Paragraph> <Paragraph position="11"> Subcategorization One of the devices used by the DELPHI grammar for encoding subcategofization is derived from GPSG \[7\]. This is to place a feature for subcategofization on the rule for a given verb, and key all VP rules off this feature:</Paragraph> <Paragraph position="13"> The semantics of the subject of the sentence is passed down through the :SUBJ variable to the V, along with the semantics of the complements. The V in turn passed back up the formula representing the semantics of the whole sentence, through the :WFF variable.</Paragraph> <Paragraph position="14"> Of course this mechanism requires one VP rule for every subcategorization frame one wants to handle, and this can mn to many rules (about 60 in the DELPHI grammar). A more serious problem, however, arises in the case of optional complements to verbs, as seen in the following actual ATIS training sentences that use the verb &quot;arrive&quot;: Show me all flights from Boston to Denver that arrive before 5 PM Show me flights ... that arrive in Baltimore before noon Show me all the nonstop flights arriving from Dallas to Boston by 10 PM ...</Paragraph> <Paragraph position="15"> Show me flights departing Atlanta arriving San Francisco by 5 PM Show me flights arriving into Atlanta by 10 PM from Dallas We see here that, in addition to the temporal PP that always accompanies it, &quot;arrive&quot; can be followed by (1) nothing else, (2) a PP with &quot;in&quot;, (3) a &quot;from&quot;-PP and a &quot;to&quot;-PP, (4) a bare noun phrase, or (5) an &quot;into&quot;-PP and a &quot;from&quot;-PP. The principle pattern that emerges is one of complete optionality and independence of order. Indeed, in the fifth example, the temporal PP, which might be more traditionally regarded as an adjunct rather than a complement, and thus as one of the siblings of the VP rather than one of its constituents, is instead interposed between two PPs complements, making the adjunct analysis rather problematic, s The only way the subcategorization scheme presented above could deal with these variations would be to enumerate them all in separate roles. But this would clearly be infeasible. The solution we have adopted constructs a right-branching tree of verbal complements, where the particular constituents admitted to this tree are controlled by constraint relations keying off the lexical head of the verb. There are The OPTCOMP-PP is the constraint relation; it keys off the lexical head of the verb (the variable :LEX) and combines the subject, object, and PP complement translations to produce the contribution of the PP complement to the final formula that represents the sentence meaning. An arbitrary number of PP complements are provided for by the recursion of the first role above, which bottoms out in the case of the second role when there are no more complements. Phrasal types other than PPs are accomodated by similar mles.</Paragraph> <Paragraph position="16"> The solution for PP complements to &quot;arrive&quot; such as &quot;in Atlanta&quot;, &quot;into Baltimore .... at Denver&quot; etc. follows: 5To see that these PPs are truly associated with the verb rather than somehow modifying the subject flight-NP, one need only replace the subject with the pronoun &quot;it&quot;.</Paragraph> <Paragraph position="18"> This rule says that for a flight to &quot;arrive&quot; INTO, IN or AT a city means that the city equals the value of the flight's DESTINATION-CITY attribute. Semantic type information is here notated with a shorthand keyword &quot;type&quot;; in the actual system a partially-specified term that packages semantic type information in a specific slot is unified into such variables as :SUBJ, :OBJ and :NP. Note also the use of disjunction (the :OR) to combine different prepositions together.</Paragraph> <Paragraph position="19"> Other devices for encoding subcategorization have been proposed. One, made use of in PATR-II and described in \[19\] encodes subcategorization infomaation as a list of the complement patterns that are required to follow the verb, and generates a right-branching tree of VPs, each absorbing one of the complements in the list until no more are left.</Paragraph> <Paragraph position="20"> The difference between the PATR-II scheme and the one way presented here is that the complements in PATR-II are required and must follow one another in a predetermined order. The only obvious way it could handle the optionality and order independence seen in the &quot;arrive&quot; examples is just the same brute-force method of enumerating possiblities, using lists of differing order and length.</Paragraph> <Paragraph position="21"> Using Unification to Encode Semantic Constraints beyond Semantic Type Not all constraints on meaningfulness are strictly reflections of the semantic type of phrase denotations. Consider the lexical item &quot;L&quot; in the ATIS training set. Seen in a number of different database fields, it can variously denote limousine availablity, lunch service, or other classes of service available on a flight. Yet in the following example it is clear that its usage is relevant to just the first of these: What is transport code L? Our claim is that not just the referent of &quot;L&quot;--limolasine service for ground transportation--that plays a role here, but also the means by which it gets to that referent: namely, by being an abbreviation or code rather than a name. That is, &quot;transport code L&quot; is a meaningful compound, while &quot;transport code limousine&quot; would not be. The lexical entry for abbreviation terms like &quot;L&quot; reflects this by taking the form of an inverse function application: the referent of the lexical item &quot;L&quot; is the ground transportation type that has the string &quot;L&quot; as its abbreviation:</Paragraph> </Section> <Section position="3" start_page="233" end_page="234" type="metho"> <SectionTitle> (INVERSE-VAL* (ABBREV-OF) &quot;L&quot; (GROUND -TRANSPORTATION)) </SectionTitle> <Paragraph position="0"> While this has the same referent as the lexical entry for &quot;limousine&quot; it has a different form, one which the role analyzing the construction above makes use of.</Paragraph> <Paragraph position="1"> Nominal compounds in the DELPHI system are generated by the following rule: in which the constraint relation NOM-COMP-READING computes the semantics of the whole conslruction from the semantics of the head houn and the nominal modifier. NOM-COMP-READING has different solutions for different semantic types of noun translation. The relevant one here is: The first slot of the NOM terms above encodes the argument-taking property of relational nouns such as &quot;code&quot;, &quot;salary&quot; or &quot;speed&quot;, and has been described in an earlier paper \[20\]. The rule states that an inverted attribute reference (here, &quot;L&quot;) preceded by a relational noun (here, &quot;code&quot;) for that same attribute (here, &quot;ABBREV-OF') simply refers to that inverted attribute reference.</Paragraph> <Paragraph position="2"> Our view is that the preceding nominal modifier essenfiaUy performs a function of disambiguation: it serves to distinguish the desired sense of the head from any other possible one. This is reinforced when the whole compound-&quot;transport code L&quot; is considered. Another NOM-COMP-READING role is responsible for combining &quot;code&quot; with This constrains the domain of the relational noun it modifies to be just the set that is the translation of the modifying noun. The semantic type of the modifying noun must be unifiable with the argument type of the head relational noun. After this unification fomas the translation of the compound &quot;transport code&quot;, the resulting type constraints serve to distinguish the correct sense of &quot;L&quot;--that of ground transportation-from the meal service and other senses.</Paragraph> <Paragraph position="3"> Our technique of allowing roles to impose restrictions on the forms of semantic translations themselves, rather than merely on the semantic types of this translations, bears some discussion because it differs from proposals made by others, such as \[14\]. In that work, the position is taken that inspecting or restricting the structure of a logical form is inadmissible on theoretical grounds, in that it violates or makes unenforceable the principle of compo@tionality. As far as theoretical matters go, we believe that the principle of compositionality is open to many interpretations (for example, see \[15\]) and that its most general interpretation does not exclude techniques such as ours.</Paragraph> <Paragraph position="4"> A more practical concem, and one which may well underly or justify the theoretical qualm, is that rules based on the structure of logical form may not succeed if that structure happens to be transformed (say through wrapping with another structure) into some syntactically different form which cannot be recognized as one which the role should allow.</Paragraph> <Paragraph position="5"> This is not a problem for the rules presented in this section, since the operation of nominal compounding is so tightly localized, operating in a function-argument fashion that gets to the noun meaning &quot;first&quot;, before other modifications such as postnominal adjuncts. Ultimately, though, we feel that the real solution must lie in a different approach to meaning representation, one in which non-denotational properties of the utterance meaning are encoded or highlighted in a way that stands above the variances of syntactic logical form. But this is a matter for future research.</Paragraph> <Paragraph position="6"> Can Constraint Relations be Compiled out of Rules? A natural question to ask is whether or not constraint relations can be compiled out of grammar rules, turning them into unifications like any other over the features of &quot;real&quot; constituent elements.</Paragraph> <Paragraph position="7"> One technique for this kind of compilation is certainly well-known in logic programming: the method of partial execution \[16\] in which the constraint relation is solved for at the time the rule is read in, not when it is used. The solution(s) so obtained are simply unified back into the remainder of the rule to create the compiled version. If the constraint relation is non-deterministic there will be more than one solution and hence more than one compiled version of the mle.</Paragraph> <Paragraph position="8"> This is certainly undesirable computationally, since multiple variants of the the same grammar rule will cause a parser to do redundant matching. On the other hand, if the unification method used permits full disjunction the variation can be kept &quot;factored&quot;, and the redundant matching avoided. Such a compilation will be possible, obviously, only if the constraint relation can be guaranteed to have finite number of solutions, given the degree of instantiation in which it appears in the role. But this is actually a rather strong condition to place on constraint relations, particularly on those which participate in semantic interpretation. Semantic strategies which defer part of the semantic computation for a constituent can have a problem with this condition, since there is no bound (given conjunction) on the size of most types of constituents. An example in our work would be the treatment of PP semantics outlined earlier, in which the PP is not fully translated at its own level, but instead passed up as a package of preposition and NP translation. The obvious extension for a conjoined PP would be as list of such structures. But since there is no limit to the number of PPs that can be conjoined, there is no limit to the length of the list, and thus there can be no limit to the number of solutions for the attachment constraint.</Paragraph> <Paragraph position="9"> Even ff a finite number of solutions can be guaranteed, however, there are still reasons why one might not want to compile constraint conditions out of the grammar. One reason is that leaving them in allows one to add solutions to a constraint condition without re-compiling all the grammar mles and other constraint rules which use it (a situation exactly analogous to the macro/function distinction in Lisp programming). Another is that distinguishing certain constraints from the rest of the unifications of the grammar mle enables us to intervene in the search with specific indexing methods (say on semantic type). This is discussed in \[3\].</Paragraph> <Paragraph position="10"> And a final benefit of distinguishing constraint conditions from other unifications is that relative frequency counts can more easily be made of their different solutions, just as can be made on semantic senses of lexical items. This is discussed in \[1\].</Paragraph> </Section> class="xml-element"></Paper>