File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/94/a94-1022_metho.xml
Size: 20,046 bytes
Last Modified: 2025-10-06 14:13:35
<?xml version="1.0" standalone="yes"?> <Paper uid="A94-1022"> <Title>The Delphi Natural Language Understanding System</Title> <Section position="4" start_page="132" end_page="133" type="metho"> <SectionTitle> 2 Grammar And The Syntax/Semantics </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="132" end_page="133" type="sub_section"> <SectionTitle> Interface </SectionTitle> <Paragraph position="0"> The Delphi grammar is a broad coverage, domain independent grammar of English written in a version of the Definite Clause Grammar formalism (Pereira and Warren, 1980) that has been extended to include labeling of right-hand side elements with the grammatical relations they bear to the head of the construction. An example is:</Paragraph> <Paragraph position="2"> subject: (NP ?arg ?mood etc.) head: (VP ?agr ?mood etc.) In this rule, there is a head VP and an NP which bears the SUBJECT relation to it. Other grammatical relations include the familar DIRECT-OBJECT and INDIRECT-OBJECT as well as the prepositions, such as TO, FROM, WITH and so on.</Paragraph> <Paragraph position="3"> Annotating sub-constituents with grammatical relations regularizes the syntactic structure with respect to particular grammatical rules, and allows a &quot;relation-to-relation&quot; form of compositionality, as opposed to the more traditional &quot;rule-to-rule&quot; version that is exemplified by such systems as Gemini (Dowding et al, 1993) and the Core Language Engine (Alshawi, 1992). In relation-to-relation compositionality, each grammatical relation in the syntactic structure corresponds to a semantic relation in a parallel semantic structure we call a &quot;semantic graph&quot;. The terminal nodes of the semantic graph are the word meanings, corresponding to the lexical heads of syntactic structure.</Paragraph> <Paragraph position="4"> An example of a semantic graph, representing the meaning of &quot;What flights fly from Boston to Denver&quot;, may be seen in Figure 2. The semantic graph is not a fully quantified formula; rather it may be thought of as a form of predicate-argument representation, with quantifiers in place, from which a fully quantified formula can be generated. The allowed class and relation labels come from the domain model.</Paragraph> <Paragraph position="5"> This view of the syntax/semantics interface has marked advantages. For one thing, because the syntactic/semantic structure is built up one argument at a</Paragraph> <Paragraph position="7"> time, it becomes much easier to accomodate such phenomena as order-variation and optionality of arguments that are difficult for other approaches The importance of this feature may be seen in the examples of argument order-variation and optionality that abound in real data. Consider the following from the ATIS domain, in which complements can vary freely in order: What flights fly from Boston to Denver? What flights fly to Denver from Boston? or be separated from the head by a modifier typically regarded as an adjunct: What flights fly at 3 pm from Boston to Denver? In some cases, modifiers can be omitted, as in: What flights fly from Boston ? What flights fly to Denver? and sometimes the omission of an argument can have anaphoric consequences, as in: What restrictions apply? which cannot be felicitously uttered except in a context where there is something in the discourse that a restriction could &quot;apply&quot; to.</Paragraph> <Paragraph position="8"> Conventional approaches to subcategorization, such as Definite Clause Grammar (Pereira and Warren, 1980), Categorial Grammar (Ades and Steedman, 1982), PATR-II (Shieber, 1986), and lexicalized TAG (Schabes et al, 1988) all deal with complementation by including in one form or another a notion of &quot;subcategorization frame&quot; that specifies a sequence of complement phrases and constraints on them. Handling all the possible variations in complement distribution in such formalisms inevitably leads to an explosion in the number of such frames, and a correspondingly more difficult task in porting to a new domain.</Paragraph> <Paragraph position="9"> In our approach, on the other hand, it becomes possible to view subcategorization of a lexical item as a set of constraints on the outgoing arcs of its semantic graph node. Different types of constraints - order of arguments, optionality of arguments, semantic-class constraints and semantic effects of arguments - can all be represented separately, instead of enumerating all possible argument sequences in a set of alternative subcategorization frames.</Paragraph> <Paragraph position="10"> Subcategorization constraints in Delphi are encoded in lexical entries using a structure called a &quot;map&quot; (Stallard and Bobrow, 1991). Below is part of the lexical entry for &quot;fly&quot; in the ATIS domain:</Paragraph> <Paragraph position="12"> Map entries have &quot;translation&quot;, &quot;realization&quot; and &quot;completion&quot; components. The translation part of this entry specifies that the lexical head &quot;fly&quot; is to correspond to a semantic-graph node labeled with event-class FLY. The realization part of the entry specifies what grammatical relations the lexical item takes, and what semantic relations these correspond to, or &quot;realize&quot;, in the semantic graph. Here, the entry specifies that &quot;fly&quot; takes SUBJECT, TO, and FROM complements, and that these grammatical relations correspond to the semantic relations FLIGHT-OF, DEST-OF, and ORIG-OF respectively. Semantic selectional restrictions in these argument positions - that the filler of DEST-OF be a city, for example - are implicit from the declarations of the relations in the domain model.</Paragraph> <Paragraph position="13"> The &quot;completion&quot; part of the entry specifies what out-going arcs are required for the node. Here, the entry requires that the FLIGHT-OF role be filled, and that either the DEST-OF or ORIG-OF roles be filled (forbidding the intransitive &quot;the flight flies&quot;). More complex optionality cases are encoded with other completion predicates. For example, the case where an anaphor must be present (&quot;What restrictions apply&quot;) is encoded by the predicate FILLED-OR-ANAPHOR.</Paragraph> <Paragraph position="14"> Some realization rules are tied to semantic classes rather than lexical translations, and require for their application only that semantic class restrictions implicit from the domain and range of the realized relation be satisfied. Typical examples are the rules governing noun modifier meanings, such as &quot;Delta flights&quot;, &quot;Delta's flights&quot;, &quot;the flights on/aboard Delta&quot;. These would all be handled by the global realization rule:</Paragraph> </Section> </Section> <Section position="5" start_page="133" end_page="133" type="metho"> <SectionTitle> {NOM-COMP POSS ABOARD ON ...} </SectionTitle> <Paragraph position="0"/> <Paragraph position="2"> Determining what semantic relation a given grammatical relation instance corresponds to is most generally viewed as a form of goal-solving in Delphi, in which a chain of rules can be invoked. For example, syntactic constructions such as &quot;X with Y&quot;, &quot;X has Y&quot; and &quot;X's Y&quot; are interpreted by first appealing to a rule mapping them to a pseudo-relation called GENERALIZED-POSSESSION, and then seeking a realization for it that is compatible with the classes of X and Y. This avoids having to write three different versions of the same realization rule.</Paragraph> <Paragraph position="3"> An important advantage of the realization rule formulation, apart from its its power and flexibility, is its simplicity. Realization rules are very simple to write, and make maximal use both of knowledge about the domain and general knowledge of language.</Paragraph> </Section> <Section position="6" start_page="133" end_page="134" type="metho"> <SectionTitle> 3 Ill-Formedness Handling: The </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="133" end_page="134" type="sub_section"> <SectionTitle> Semantic Linker </SectionTitle> <Paragraph position="0"> When an utterance cannot be parsed with Delphi's best-first parser (Bobrow, 1991) - either because it is illformed, mis-recognized by the speech system, or simply because it is outside the coverage of the grammar it can still be partially understood by the system, often well enough to give the correct response. The component responsible for partial understanding in the Delphi system is called the Semantic Linker (Stallard and Bobrow, 1993).</Paragraph> <Paragraph position="1"> After a parse fails there is a set of fragmentary constituents left over in the chart, corresponding to a set of semantic graphs. The Semantic Linker seeks to connect these sub-graphs into a single connected one by adding links between nodes in the different sub-graphs.</Paragraph> <Paragraph position="2"> At top-level, this is the same thing that the parser and grammar do. The difference is that the parser and grammar have an idea of what the grammatical relationship between constituents is, based on requirements of their proximity in the string and other syntactic evidence. The Semantic Linker does not have these requirements, being a looser form of combination that can ignore fragment order and skip over intervening, unanalyzable material with ease.</Paragraph> <Paragraph position="3"> Although it is a very different algorithm, the Semantic Linker uses the same set of realization rules that drives the regular parser. Using the realization rules, the Linker determines for each pair of nodes in different semantic graphs the set of all links which can connect them. It then uses an A* search to find the most plausible set of links which produce a complete graph.</Paragraph> <Paragraph position="4"> Suppose for example, we have the three fragments &quot;to Boston&quot;, &quot;Denver&quot; and &quot;Delta flights on Monday&quot;. Then the three corresponding sub-graphs are those shown in Figure 3 where a PP is treated as its NP object with the preposition as a tag. For this set of fragmentary sub-graphs, the possible links are: la. FLIGHTS1--- DEST-OF -> BOSTON:TO</Paragraph> <Paragraph position="6"> where the links are grouped together in a ordered list according to the fragment-pairs they connect.</Paragraph> <Paragraph position="7"> The plausibility of a given link is a function of a number of different features, including penalities from assumptions made in its computation (e.g. that a given preposition can be ignored or assumed) and empirically determined probabilities for the given link (e.g. that given an AIRLINE and a FLIGHT they are most probably linked by the relation AIRLINE-OF).</Paragraph> <Paragraph position="8"> The semantic linker may also &quot;hallucinate&quot; a new node to bridge two fragments between whom no links can otherwise be computed. For example, for the utterance &quot;from Boston to Denver&quot;, which has no explicit FLIGHT-object, a FLIGHT node can be inserted between the fragments to make sense of the utterance.</Paragraph> <Paragraph position="9"> Because the Semantic Linker uses the same set of realization rules as the rest of the system, when the system is ported to a new domain the Semantic Linker can be used immediately - a distinct advantage over some other approaches to fallback understanding, such as (Stallard and Bobrow, 1992) or (Jackson et al, 1991).</Paragraph> <Paragraph position="10"> In formal experiments (as we discuss subsequently) the Semantic Linker has been show to dramatically improve Delphi's performance.</Paragraph> </Section> </Section> <Section position="7" start_page="134" end_page="134" type="metho"> <SectionTitle> 4 Quantification </SectionTitle> <Paragraph position="0"> The quantifier scoping module in Delphi takes a semantic graph and produces a fully-scoped expression in the logical language FMRL. The basic strategy for quantifier scoping is a descendant of that used in the LUNAR system (Woods et al, 1978). This is made possible by the use of the semantic graph as a common underlying representation for both the grammatical and ill-formod parts of fragmentary utterances. Delphi's scoping module traps quantifiers from relative clauses, makes the quantifiers from PPs etc. outscope the NP quantifier, and resolves the scope of quantifiers from parallel constituents in terms of left-to-right order in the input. These general rules are modified to take into account differing strengths of quantifiers such as EACH.</Paragraph> <Paragraph position="1"> Left-to-right ordering and syntactic structure for grammatical portions of the utterance are recovered from the semantic graph by backpointers to the lexical items and grammatical relations from which the graph was produced. Links established by the semantic linker are treated by the quantification mechanism as if the constituency is indeterminate, so that only left-to-right scoping rules and individual quantifier preferences take effect.</Paragraph> <Paragraph position="2"> The resulting mechanism is robust, and quantificational scoping has been an insignificant source of error in the official ARPA blind-test evaluations of the ATIS system. More complex strategies have been proposed and implemented in the last two decades, and could in principle be modifed to work with ill-formed input, but the simple and robust LUNAR approach handles essentially all the phenomena seen in the tens of thousands of sentences of ATIS training collected during experiments with non-linguist users.</Paragraph> </Section> <Section position="8" start_page="134" end_page="135" type="metho"> <SectionTitle> 5 Discourse </SectionTitle> <Paragraph position="0"> The discourse mechanism of Delphi consists of several components: resolution of local ambiguities, pronominal and deictic antecedent resolution, ellipsis handling and discourse constraint propagation.</Paragraph> <Paragraph position="1"> The most common case of local ambiguity in the ATIS domain involves temporal phrases as in &quot;the nine o'clock flight&quot;. The resolution mechanism searches both for linguistic information in the current and previous sentences, as well as properties of entities in previous answers, to resolve whether &quot;nine o'clock&quot; is AM or PM.</Paragraph> <Paragraph position="2"> The pronourddeictic resolution mechanism used in Delphi makes use of locally expressed or implied semantic constraints to search through a set of candidate antecedents. The current mechanism ignores syntactic number as a cue, because empirically in the ATIS corpus (and we suspect in other spontaneous speech applications) it is often in error. A simple-minded focus component is used, primarily based on recency, and secondarily based on grammatical relations within an utterance. Because of the strength of semantic cues and the prevalence of ill-formed input, the use of syntactic cues for focus is limited.</Paragraph> <Paragraph position="3"> The interpretation of later sentences often must include information from previous sentences, without explicit linguistic cues. This is especially true in &quot;design dialogues&quot;, where the goal is to find a description of a set of objects that will meet some set of implicit or explicit constraints. Consider for example the following discourse from the ATIS domain.</Paragraph> <Paragraph position="4"> Show Delta fights from Boston to Dallas tomorrow.</Paragraph> <Paragraph position="5"> Can I leave in the morning? Is there a nonstop flight? Show me the American flights.</Paragraph> <Paragraph position="6"> I want to go from Dallas to Chicago on Wednesday Note that the constraints of prior sentences (such as on airline, origin, destination etc.) are implicit for subsequent sentences unless contradicted by information in the current sentence (e.g. &quot;American&quot; overrides the &quot;Delta&quot; from the first sentence) or until there is evidence that a new problem is being solved (the new origin and destination in the last sentence indicates that all previous constraints can be dropped). Delphi has a &quot;context tracker&quot; that maintains a stack of the constraints from previous utterances, and has a set of rules for when constraints are to be modified or deleted before being merged with the current sentence.</Paragraph> <Paragraph position="7"> Finally, we handle ellipsis as a special case of semantic linking. If we have the two utterances: Show me the meals on the morning flight.</Paragraph> <Paragraph position="8"> on American at 12:30 We can treat these as if they were one run-on ill-formed input and link &quot;American&quot; to &quot;flight&quot;, and replace &quot;morning&quot; with &quot;12:30&quot;, using a minor variant of the Semantic Linker linker which allows for later constraints to overwrite earlier ones of the same type. This strategy has been very effective, and covers a large class of elliptical constructions.</Paragraph> </Section> <Section position="9" start_page="135" end_page="135" type="metho"> <SectionTitle> 6 Backend Mapping </SectionTitle> <Paragraph position="0"> In order to get a response to a user query, the complete FMRL interpretation of an utterance must be translated to an expression of a target query language which can be evaluated directly against the tabular database to retrieve the answer.</Paragraph> <Paragraph position="1"> A key step is bridging the gap in conceptual vocabulary between the two representations. For example, the FMRL interpretation of the query &quot;How many flights on Delta serve meals&quot; has one-place predicates like FLIGHT and AIRLINE, and two-place predicates like AIRLINE-OF and MEAL-OF. The database for the ATIS domain, on the other hand, only has a single table FLIGHT with fields containing airline and meal information. Delphi bridges this gap between representations with a system of local mapping rules which translate the one- and two-place predicates of the FMRL into expressions of a relational algebra target language which retrieve the extensions of these predicates.</Paragraph> <Paragraph position="2"> Sometimes, however, some combination of FMRL predicates has a correspondence in the database but the individual predicates themselves do not. For example, in the database for the SPLINT domain a table relating aircraft-types to their physical characteristics has a field for the number of engines the aircraft has, but no representation for the engines themselves. If we now ask &quot;How many engines does an F-16 have?&quot;, there is no local translation of the FMRL predicate ENGINE.</Paragraph> <Paragraph position="3"> To deal with this, Delphi has a system of global transformations that are applied first, rewriting subsets of the FMRL clauses to a form that can be handled with local translation. The rule that handles this example is:</Paragraph> </Section> <Section position="10" start_page="135" end_page="136" type="metho"> <SectionTitle> 7 Interface To A Speech Recognizer </SectionTitle> <Paragraph position="0"> In spoken language applications, Delphi is interfaced to the output of the Byblos speech recognition system (Bates et al, 1993). The N-best paradigm is used, in which the recognizer outputs in order its top N guesses at the transcription of the sentence, for some value of N (usually 5). Delphi then runs over these transcriptions in the order they have been ranked, first with the Semantic Linker disabled so that only grammatical utterances are allowed, and if none is found, runs over them again with the Semantic Linker enabled.</Paragraph> </Section> class="xml-element"></Paper>