<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-1011">
  <Title>PROCESSING COMPLEX NOUN PHRASES IN A NATURAL LANGUAGE INTERFACE TO A STATISTICAL DATABASE</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> A prototype natural language front-end to statistical databases is being developed as part of an Executive Information System for Rogers Cablesystems, a Canadian cable television company. The initial target database is the Rogers Technical Operations Database, a relational database containing statistical data describing aspects of the company's business related to customer service.</Paragraph>
    <Paragraph position="1"> The front-end employs an HPSG chart parser. There axe numerous variations of HPSG; we have chosen \[PS87\] since it is the most familiar and widely published. Our results can be extended to other variations. In the spirit of HPSG, we have avoided a proliferation of grammar roles and kept them highly schematic.</Paragraph>
    <Paragraph position="2"> In developing the grammar for the queries in our corpus, we encountered a selection of interesting noun phrase constructions which caused us to examine the treatment of adjunct modification of nominals within HPSG. This has resulted in a proposal which should be of interest to other researchers developing natural language interfaces.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Complex NPs in Queries
</SectionTitle>
    <Paragraph position="0"> We began the project by collecting a corpus of 68 English language queries from three senior executives at Rogers. Our corpus contains constructions paradigmatic of a wide selection of natural language queries that the executives would like to pose to their database.</Paragraph>
    <Paragraph position="1"> A selection of these queries are shown in (1-6).</Paragraph>
    <Paragraph position="2">  (1) Give me the we.stem region outage log summary.</Paragraph>
    <Paragraph position="3"> (2) Give me the system reliability performance.</Paragraph>
    <Paragraph position="4"> (3) Compare the basic service problem statistics per thousand customers.</Paragraph>
    <Paragraph position="5"> (4) Compare the terminal equipment problems.</Paragraph>
    <Paragraph position="6">  The sentences contain complex NP constructions and there is a large amount of variation with respect to the location and ordering of the modifiers. For example, most pre-nominal modifiers may also appear as post-nominal modifiers.</Paragraph>
    <Paragraph position="7">  (5) Vancouver system reliability performance (6) system reliability performance for Vancouver  Prepositional phrases like for Vancouvercan be viewed as an abbreviated form of the prepositional phrase for the Vancouver division.</Paragraph>
    <Paragraph position="8"> The NPs within these sentences contain a great deal of syntactic ambiguity. Consider the complex NP in (1). The adjective western can either modify region or outage or log or summary. Similarly, region could modify any of the nominals appearing to its right. However, much of this syntactic ambiguity does not ACRES DE COLING-92. NANTES, 23-28 AOt3&amp;quot;r 1992 4 6 I'ROC. OI: COLING-92, NANTES. AUO. 23-28, 1992 have a semantic interpretation in the database semantics. For example, (1) has only a single interpretation although there are numerous syntactic analyses.</Paragraph>
    <Paragraph position="9"> We have gone into detail about the corpus to show the rich structure of noun phrases and to motivate the reasons for the design choices in our semantics and grammar.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Complex NPsin HPSG
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Overview of HPSG
</SectionTitle>
      <Paragraph position="0"> HPSG is one of the best known uni fication-based grammar formalisms. It employs attribute value matrices (called signs) to represent lexical entries, grammar rules and principles. HPSG borrows freely from other formalisms. For example, the treatment of syntactic categories, syntactic features, and sonre of the principles are from generalized phrase structure grammar (GPSG) \[GKPS85\]. The main syntactic categories in HPSG are heads (the head constituents of phrases), adjuncts (traditionally called modifiers) and complements (traditionally called arguments). The principles of HPSG include the Constituent Order Principle, Sub-categorization Principle, Head Feature Principle, and Semantics Principle.</Paragraph>
      <Paragraph position="1"> HPSG contains three grammar rules for combining heads with complements.</Paragraph>
      <Paragraph position="2">  (7) \[SUBCAT (\[ \])\] --, H\[LEX+, INV-\], C&amp;quot; (8) \[SUBCAT( )\] ~ a\[LEX-\], C (9) \[SUBCAT( )\] --, H\[LEX+, INV +\], C*  One rule (7) combines a lexical head with everything but its final complement. This rule can also be used to convert a lexical head requiring only a single complement into a non-lexical constituent still requiting a single complement. Another rule (8) combines a non-lexical head with its final complements. Yet another rule (9) works for inverted constructions: those involving a lexical head that is marked for inversion. As in GPSG, generalizations about the relative order of sister constituents is factored out of the phrase structure rules and expressed in independent linear precedence (LP) constraints. The LP constraints are used by the Constituent Order Principle. HPSG roles are immediate dominance (ID) rules. Consequently, a single ID rule of the form X --, HA could describe a head constituent H either preceded or followed by an adjunct A -- the relative ordering of H and A is determined by the LP constraints.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.2 Issues in the Treatment of Adjuncts
</SectionTitle>
      <Paragraph position="0"> Nominal modification is treated ill HPSG by having heads that contain a set valued feature called ADJUNCTS \[PS87\]. Each element of this set is a sign which describes a potential adjunct. For instance, the ADJUNCTS feature for a noun will contain an entry for adjectives, one for nouns, one for prepositional phrases and one for verb phrases.</Paragraph>
      <Paragraph position="1"> An alternative, which was also discussed in \[PS87\] and has been adopted in other grammar formalisms (e.g., \[Usz86, CKZ88\]) and some variations of HPSG \[Coo90, Po191\], is to allow adjuncts to select their heads, t The head feature called HEADS contains a set of descriptions, one for each construction that can be modified by the adjunct. For example, the HEADS feature for an adjective will contain a sign for a noun. In our corpus, a head has more possible classes of modifiers than modifiers have classes of possible heads. For example, the set of modifiers for NPs and Ns (i.e., NPs lacking determiners) includes adjectives, nominals, PPs and even VPs (relative clauses). In SS3.4 we shall see that each of these modifiers can have only one or two possible heads. Furthermore, the task of reducing the size of the HEADS or ADJUNC'rS set, by discovering common semantic features for which a constituent can select, meets with greater success if modifiers select their heads. That is, one is more likely to find commonality among the constituents which an adjunct can modify than among the modifiers which a head can take. Selections of heads by adjuncts permits a greater range of subcategorization to be specified through default inheritance rather than explicit specification. null Some aspects of adjunct semantics arc impossible if adjuncts are selected by heads rather than heads selected by adjuncts. Predicates, both adjectives and verbs, have argument structure which coerces their arguments into thematic roles. For exanlple, the adjective modern imposes on its argument the thematic role of Theme. ~ It is not obvious how the nonrinal argument of the adjective receives its thematic role unless it is the adjective which selects the nominal, parallel to the assignment of thematic roles by verbs to their NP arguments. If modern selects its head, then the thedeg matic role of the head may be specified in the HEADS I Cooper ICoo90, Ch.3, SS6\] looks in some detail at the arguments in favour of adjuncts selecting their heads.</Paragraph>
      <Paragraph position="2"> 2In \[Po191, SS1.3\], Pollard and Sag introduce semantic features like AGENT, GOAL andTHEME within the feature structure containing the semantic CONTENT.</Paragraph>
      <Paragraph position="3"> ACl'ES DE COL1NG-92, NANTI.:S, 23-28 AOt~T 1992 4 7 PROC. OF COLING-92, NA~rn~S, Auo. 23-28, 1992 attribute and inherited by the head when it unifies with the HEADS attribute. If instead, heads subcategorize for their adjuncts, this information must be inherited in some other fashion, perhaps through structure sharing from the adjuncts list.</Paragraph>
      <Paragraph position="4"> The problem and its solution are evident when derivational morphology are considered. The verb read imposes the thematic role of Agent (Ag) on its subject and the thematic role of Theme (Th) on its object. When this verb is coerced into an adjective by the derivational suffix -able, the resulting adjective assigns the thematic role of Theme to its argument. If adjectives select their heads, then the derivational rule is evident.</Paragraph>
      <Paragraph position="5"> (lO) V\[SUBCAT (NPTh, NPAg)\] Adj q-&amp;quot;able&amp;quot;\[HEADS {NTh}\] Given that adjuncts will select their heads, a grammar role for adjuncts can be stated most concisely if we combine a head with a single adjunct at a time. Thus, our constituent structures will contain an ADJUNCT-DTR feature which will take the adjunct as its value, rather than a list-valued ADJUNCT-DTRS feature which would take a list of adjuncts as its value. A head that is modified by more than one adjunct will require more than one application of the grammar rule.</Paragraph>
      <Paragraph position="6"> One disadvantage of this approach is that a complex nominallike system reliability for Vancouverwill have two analyses: one where the PP for Vancouver modifies the head noun reliability and another where it modifies the head nominal system reliability. If the adjuncts rule combined a head with all of its adjuncts at the same time, there would be only one analysis.</Paragraph>
      <Paragraph position="7"> However, one could argue that there should be two interpretations for the phrase and that both should be reflected in the grammar. Pollard and Sag note that &amp;quot;there is evidence that noun-noun and adjective-noun structures share some syntactic properties with lexical nouns as opposed to typical common noun phrases, e.g. they can occur themselves as modifiers in noun-noun structures&amp;quot; \[PS87, p.73\]. They propose analyzing noun-noun and adjective-noun constructions as \[LEX +\] even though they have internal structure. By adopting this treatment of complex noun phrases, we can prevent analyses for ungrammatical constructions like system for Vancouver reliability, plus we can prevent ambiguity in the analysis of phrases like system reliability for Vancouver. In our grammar we introduce two rules for adjuncts, which are designed to give wide coverage and to avoid spurious ambiguities.</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.3 Two Rules for Adjuncts
</SectionTitle>
      <Paragraph position="0"> One adjunct grammar rule is required for combining saturated lexical adjuncts with their heads. That is, for lexical adjuncts which have empty subcategorization lists, like adjectives, proper nouns (specifically, the proper nouns corresponding to months and cities) and adverbs. The rule will be restricted so that it will apply to phrases with unsaturated heads. Heads that fall into this category are Ns, PPs, 3 VPs, and APs. The specific pairing of adjuncts to heads is determined by the HEADS feature of the adjunct (SS3.4). Additionally, if the head modified by the adjunct is marked \[LEX +\] then the resulting constituent will also be \[LEX +}, thus implementing the analysis of adj-noun and noun-noun constructions discussed in the previous section. Using the schematic notation for grammar rules introduced in \[PS87\], we can present the rule as shown in (11).</Paragraph>
      <Paragraph position="2"> Note that the two appearances of \[\] in (11) indicate that the head and the resulting constituent share the same value for their LEX features. The Subcategorization Principle will ensure that the head and the resulting constituent will have the same value for their SUBCAT features. Since the grammar rule is an ID rule, it does not place any restriction on the linear ordering of the head (H) and adjunct (A). This rule is designed so that it applies before a head is combined with its final complement (8). It can be viewed as the HPSG counterpart to the adjunct rule from X-bar theory \[Cho82\] shown below, where the ADJUNCT is required to be lexical and not subcategoriz, e for any arguments.</Paragraph>
      <Paragraph position="4"> in order for heads to be modified by unsaturated adjuncts, we propose a second grammar rule.</Paragraph>
      <Paragraph position="6"> ~Like \[PS87, p.70\], we propose that propositions have two elements on their subcategorization list, the first being the prepositional object and the second its subject. A PP is obtained by combining a preposition with its object NE We do not propose lexical entries for prepositions having only the object NP on its SUBCAT list since this would complicate the LP roles (~3.5) and grammar rules (7) and (8).</Paragraph>
      <Paragraph position="7"> AO'ES DE COLING-92, NANTES, 23-28 AO(rl&amp;quot; 1992 4 8 PROC. OF COLING-92. NAI'rrEs, AUG. 23-28, 1992 Rule (13) requires the adjunct to have a single element in its SUBCAT list, thus allowing PR VP and modiliers to modify PPs, VPs and Ns. Of course, the contents of the HEADS feature will restrict the applicabillty of this role (fi3.4). Unlike rule (11) which allowed a lexical adjunct to modify either a lexical or non-lexical head, rule (13) requires the head, adjunct and resulting constituent to possess the same values for their LEX features, as reflected by the coindexing wilh \[j_-\]. With this role, a &amp;quot;lexical&amp;quot; compound noun can modify a lexical noun to yield a &amp;quot;lexical&amp;quot; compound noun (e.g., N -~ N, N), or a (non-lexical) PP can nmdify a non-lexical nominal m yield a non-lexical nonrinal (N -~ N, PP).</Paragraph>
      <Paragraph position="8"> Direct consequences of our two adjuncts ndes are that prepositions and verbs are not allowed to modify anything (these have two or nrore elements in their SUBCAT lists), sentences or complex noun phrases cannot appear as adjuncts, and NPs, Ss, adjectives, verbs and prepositions cannot be modi fled by anything.</Paragraph>
      <Paragraph position="9"> Our grammar does not prevent nouns from being modified, since rule (7) can be applied to a lexical noun to yield a non-lexical nominal (essenti',dly, N ~ N). If we &amp;quot;allowed full NPs or Ss to be modified, the result would be a syntactic ambiguity which would not have any semantic relevance.</Paragraph>
    </Section>
    <Section position="4" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.4 The HEADS Feature
</SectionTitle>
      <Paragraph position="0"> The applicability of the two adjuncts grammar roles is restricted by the value of the HEADS feature of the adjunct. For prepositions (lexical entries with SYNILOCIHEADIMAJ = P), the value of the HEADS feature will be a set containing a sign for N constituents (N\[SUBCAT (\[\]), LEX-\]) and a sign for VP constituents. 4 Lexical entries for nouns and adjectives will have a single element in their HEADS set. It will contain a sign for lexical nouns, which inchtdes compound nouns (N\[SUBCAT (\[ \]), LEX +\]). We are proposing that pre-nominal lnodifiers, like adjectives and (compound) nouns, will be combined with their head nouns before post-nominal modifiers, like PPs.</Paragraph>
      <Paragraph position="1"> We adopted this decision because applying modifiers in different orders does not result in any difference in the resulting semantic interpretation. Specifically, the semantic representation associated with \[the \[lsystern reliability\] for Vancouverl\] is the same as that 4In our corpus PPs do not appear to nlodify any VPs, so we can actually simplify the HEADS feature so that it contains only the N entry.</Paragraph>
      <Paragraph position="2"> for \[\[the \[system mliabilityl\] lot Vancouverl and \[the Isystem \[reliability tbr Vancouverlll. With our proposal, we obtain only one analysis tot the phrase discussed alx~ve. Finally, in order to allow relative clauses (MAJ=V), we need only propose that they contain a sign for N in their HEADS set. Ttms, we effectively treat relative clauses like restrictive relative clauses.</Paragraph>
      <Paragraph position="3"> As was the case with PP adjtmcts, the same semantic representation is obtained regardless of whether the relative clause modifies an N (restrictive relative) or an NP (non-restrictive relative).</Paragraph>
    </Section>
    <Section position="5" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.5 Linear Precedence
</SectionTitle>
      <Paragraph position="0"> We adopt the same LP constraints for heads and complement danghters as proposed in \[PS87\]. Lexical heads are required to precede their cmnplement(s), while non-lexical heads tollow their complement(s).</Paragraph>
      <Paragraph position="1"> Sister cmnplements appear in the reverse order of their appearance in tim SUBCAT list of flmir head. The I,P constraints lot adjuncts require signs with MAJ-A or MAJ:N (+N categories in terms of the chtssification present in \[Cho821) to precede their beads, while adjuncts with MAJ=V or MAJ=P (-N categories) are required to follow their heads. Thus adjectives and nominal modifiers will precede the nouns they modify, while PPs and relative clauses will follow the constituenls they modify.</Paragraph>
    </Section>
    <Section position="6" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.6 Semantics
</SectionTitle>
      <Paragraph position="0"> Due to the close relationship between syntax and semantics in HPSG, we can avoid syntactic ambiguities which do not con'espond to distinct semantic analyses.</Paragraph>
      <Paragraph position="1"> Semantic infomlation, consisting n fTYPE and content (CONT), can be used to prevent ceIl.ain analyses. The TYPE of a complex constituent will be tbe san~e as that of its head. The Semantics Principle is responsible for creating the CONT of a complex constituent from that of its daughters (suhconsfituents) \[PS871. We adopt a version of this principle for building up semanlic information for database stntctures, which we call the Database (DB) Semantics Principle \[McE911.</Paragraph>
      <Paragraph position="2"> We incorporate selectiomd restrictions based on a semantic type hierarchy which incoq~orates aspects of the database design. The Rogers Technical Opera~ tions Database is a statistical database; that is, each table in the database contains one or motx: category attributes (columns) whose values define sets of entities of a single type, and one or more statistic attributes (columns) whose values smnmarizc these sets. The AC'I~!S DP; COLING-92, NANrI~S. 23-28 AO\[TI' 1992 4 9 PROC. OF COLING 92, NANrI~S, At;(;. 23-28, 1992 stype sset ,/~ time Io(: . . .</Paragraph>
      <Paragraph position="3"> Figure 1 : Semantic Type ttierarchy complex noun phrases used in natural language queries to this database consist of nominals, or nominal modifiers which belong to five general classes: statistical type (stype), statistical set (sset), entity set (eset), modifier (mud) and pre-moditier (pmod). Each of these classes may be divided into subclasses using information from the conceptual database design. These five classes are arranged in a semantic type hierarchy as shown in Figure 1. Using this hierarchy, we can incorporate selectional restrictions into the HEADS feature of modifiers. Nouns like summary, sum, and ratio are used to refer to particular (sets of) statistics. Members of the sset class (e.g., log, performance, activity) may be used to modify stypes. Nouns from the sset class may be semantically vacuous, that is, we assume that all requests are forsome set of statistics and these nouns may not carry any information that can help identify the particular statistics sought by a user. We allow (compound) nouns within the eset class (e.g., problem, outage, call, reliability) to modify (compound) nouns of type star (i.e., sset or stype). Adjuncts of type rood may modify subclasses of eset. For example, a user can request either system reliability .statistics or service c~dls. The type proud may modify other modifiers and selected types of eset.</Paragraph>
      <Paragraph position="4"> The selectional restrictions distilled from our type hierarchy are by themselves not powerful enough to eliminate all of the &amp;quot;spurious&amp;quot; ,ambiguities. Just as we can use the TYPE feature from the semantics of the sign, we can also use the CONT to restrict possible analyses. To do this, we have modified the DB Semantics Principle with an Adjunct Contribution Constraint so that an adjunct is required to contribute semantic information to a head-adjunct constituent -- in particular, adjuncts must contribute references to database constructs -- hence the constraint disallows semantically vacuous adjuncts from combining with a head.</Paragraph>
      <Paragraph position="5"> A complex constituent like outage log summary, in which outage has semantic content but log makes no contribution of database information, would have only one analysis. The noun log would not be allowed to  modify summary, but outage could modify log, and then outage log could modify summary.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 Implementation
</SectionTitle>
    <Paragraph position="0"> Our treatment of complex NPs has been incorporated into the SX natural language interface \[MC90\]. The SX system uses grammar developed within the HPSG-PL grammar development system \[PV91a\]. The semantic representations built up by an HPSG parser are directed to a module which converts them into an SQL query. The query can then be directed to an Oracle database to obtain the requested information.</Paragraph>
    <Paragraph position="1"> SX makes use of chart parsing implementations of HPSG developed in LISP by McFetridge \[MC90\] and in Prulog by Popowich and Vogel \[PV91 b\]. Chart parsing is a type of parsing in which all syntactic structures which are built are placed on a single graph structure called a chart. Nodes in the chart correspond to positions in an input sentence, with edges between the nodes describing analyses ofsubstringsofthe input. A successful parse corresponds to an edge that spans the entire input sentence. The performance of the Prolog parser on sentences (1)-(4) are summarized in Table 1. For each sentence, the table shows the time in CPU seconds for obtaining the first parse (Parse) and for searching for all possible interpretations (Total). The table also contains the number of edges created by the chart parser while searching for these interpretations. To illustrate the effect of the Adjunct Contribution Constraint discussed in SS3.6, Table 1 also shows (in brackets) the number of edges and CPU times when this constraint is not used. The tests were performed on a SUN SPARCstation 1 running Quintus Prolog 3.0.</Paragraph>
  </Section>
class="xml-element"></Paper>