<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-1058">
  <Title>COORDINATION IN RECONNAISSANCE-ATTACK PARSING*</Title>
  <Section position="4" start_page="0" end_page="285" type="metho">
    <SectionTitle>
1. Theoretical Background
</SectionTitle>
    <Paragraph position="0"> In a Reconnaissance-Attack parser, no structure-building is attempted until after an initial 'overflight' of the entire sentence has been made, directed at obtaining information, provided by low-level structural cues, which can then be exploited in narrowing the range of available options at a later point. (We assume here that the cues used are present in a minimally analyzed string, by which we mean one about which the only sU'uctural information available concerns the relative order and category membership of the individual words.) It is of the utmost importance to bear in mind that in this approach, ira given case cannot be resolved at a given point in tire parse, there is no guessing as to which type of coordination might obtain and hence no need to backlrack for the purpose of undoing the effects of erroneous hypotheses; rather, the parser simply defers the decision to a later phase at which more structural information is available. Note as well that this is not 'bottom-up' parsing in the usual sense either, since where more than one possibility is logically available, the parser makes no attempt to represent them all and cull out tlte false positives later on; there is a strict principle of 'altruism avoidance' (that is, never undertaking computational effort without a guaranteed payoff) which compels the parser to give no answer at all during lThe approach described in Sampson 1986, while quite different in its actual character, is nonetheless similar in spirit to what we are proposing.</Paragraph>
    <Paragraph position="1">  a particular phase if more than one answer is possible in principle given the information available to that point. (If, at the end of the process, unresolved indeterminacies remain, ambiguity is predicted.) Intuitively, the difference between Reconnaissance and Attack is that Reconnaissance constitutes the gathering of information while Attack constitutes anything which involves decision-making. More formally, Reconnaissance can be viewed as a series of parameter-setting operations each of which is done independently of any of the others while Attack requires simultaneous access to all parameters.</Paragraph>
    <Paragraph position="2"> It is worth noting that there does not appear to be any reason to exclude in principle the possibility of hybrid models in which principles of the sort we shall develop below are invoked prior to the application of a parser along the lines of those described in e.g. Dahl and McCord 1983 or Fong and Berwick 1985. Our principal contention is that whatever choices are made about how to go about 'parsing proper' (that is, actually building a syntactic representation for an input sentence), there is an advantage to having certain global structural information already available rather than starting 'blind'.</Paragraph>
    <Paragraph position="3"> Following Kac 1978 and 1985, we subsume under a single rubric of 'predicate coordination' the coordination of verbs, VP's, and S's on the rationale that common to all three types is that they have the effect of rendering predicates 'equiordinate' (that is, so related that neither is sub- or superordinate to the other). In e.g.</Paragraph>
    <Paragraph position="4"> (2) I believe that John likes Mary and Han'y admires Sue.</Paragraph>
    <Paragraph position="5"> the verbs likes and admires are both subordinate to believe but neither is subordinate to the other. Similarly, in a sentence like (ib) above, hits and attacks are both 'topmost' in the ordination scheme. (For a more detailed development of the theory of ordination relations, see Rindflesch forthcoming.) In this approach a distinction is made between STRICT and LOOSE coordination (two coordinate expressions are strictly so if separated by at most a conjunction, loosely coordinate otherwise, as in e.g. John and Mary. ran vs. John ran, and Mary (too)) and also between PRIMARY and SECONDARY coordination. The primary coordinates in a coordinate structure are the largest coordinate expressions (e.g. the S's in sentential coordination), while the secondary coordinates are smaller expressions contained in the primary ones taken (by the theory) to be coordinate by virtue of the coordination of the containing expressions; for example, the predicates of coordinate sentences (both VP's and V's) are secondary coordinates in a sentential coordination.</Paragraph>
    <Paragraph position="6"> For purposes of parsing, we assume that the first task is to coordinate WORDS rather than the larger expressions containing them; that is, secondary coordinates are sought first, and the primary coordinates in which they appear are identified later.</Paragraph>
    <Paragraph position="7"> This is consistent with the overall theoretical approach, described in more detail in Rindflesch op. cit., which is much more akin to dependency syntax than to phrase structure analysis. (See also Kac and Manaster-Ramer 1986.)</Paragraph>
  </Section>
  <Section position="5" start_page="285" end_page="285" type="metho">
    <SectionTitle>
2. A Sketch of the Parsing Strategy
</SectionTitle>
    <Paragraph position="0"> In this paper, our focus will be on determining, from a minimally analyzed string, whether or not a given instance of and or or enters into a predicate coordination as defined above. (A longer paper giving full details of the approach is in preparation.) In the earliest stages of parsing a given sentence containing a coordinating conjunction, each conjunction is identified as either (a) definitely involved in a predicate coordination, (b) as definitely not involved in such a coordination, by virtue of falling certain necessary conditions for being so involved, or (c) as of indeterminate status which must be resolved (if possible) in a later phase of the parse. The following principles are invoked for this purpose: Applied early in Attack:  (3) LIMITS CONSTRAINT (Rindflesch forthcoming) The number of predicate-coordinating conjunctions in a sentence must be smaller than the number of verbs.</Paragraph>
    <Paragraph position="1"> (4) POSITION CONSTRAINT (Kac 1978, 1985) If a coordinating conjunction conjoins expressions X and Y, it lies somewhere between X and Y.</Paragraph>
    <Paragraph position="2"> Applied late in Attack: (5) MAIN PREDICATE CONSTRAINT There is at least one predicate in every sentence which is not subordinate to any other predicate in that sentence.</Paragraph>
    <Paragraph position="3"> (6) EQUIORDINATION CONSTRAINT  If two predicates are coordinate then they are also equiordinate.</Paragraph>
    <Paragraph position="4"> 28g The principles (3-6) are all rather straightforward, even common-~sensical; it is nonetheless not entirely uninteresting to learn that they Ibrm the basis for an extremely effective parsing strategy.</Paragraph>
    <Paragraph position="5"> Reconnaissance involves a single pass through the currant string, the first steps being lexical lookup and counting and indexing all categories. The information gained from this counting a0d indexing is then used to eliminate impossible structures, via a check for compatibility with the principles (3-6) above.</Paragraph>
    <Paragraph position="6"> In order to deal with coordination two ancillary lists, called POTlr.NTIAL COORDINATION LISTS, are associated during Reconnaissance with each conjunction which occurs in the input siring. One of these, PCL-L, contains words which occur to the left of the ronjunction with which the list is associated; each of these word.~ could thus potentially serve as the left-hand member of a coordination effected by that conjunction. The other list, PCL-R, se,:ves a similar %nction for words which occur to the right of the conjunction. Two elements can be coordinated only if one occurs in PCL-I, for a given conjunction and the other occurs in PCL-I~ for that conjunction Ttw constraints which apply early in Attack presuppose no information beyond what is gathered during Reconnaissance and are used to eliminate words in the input string as candidates for inclusion in these lists (on the assumption that it is best to elinfinate as much as possible as early as possible on the basis of the least possible amount of information and thus enhance the efficiency of the parser). The remaining constraints remove words from the lists. In the early stages of the parse, each of these lists may be quite long, but as the pm'se proceeds, elements are deleted by the invocation of the Attack principles, until, for well formed input strings, each list contains only elements which, on some adnfissible reading of the input, can enter into a coordination effected by the associated conjunction. (In ambiguous cases such as John believes the boys and the girls believe Fred, each list would have more than one member.) In unambiguous cases, it can be determined that a conjunction is definitely involved in predicate coordination if both its PCL-L and its PCL-R contain exactly one predicate and no other word, and a conjunction is definitely not involved in predicate coordination if either of its PCL's does not contain any verb at all. The coordination status of a conjunction is indeterminate with regard to predicate coordination when, although both PCL's contain a verb, one (or both) of them contains at least one additional word. A natural question to ask at this point is whether the strategy just described is not just bottom-up parsing of the familiar sort. The answer is no, for at least two reasons. First, the PCL's do not hold fully specified analyses of substrings of the input; they contain only words which, on the basis of ilfformation so far available, cannot be excluded from consideration as potential coordinates of the conjunction associated with a given pair of lists. Nor do the lists hold potential conjunct pairs. (Suppose, for example, that PCL.-L and PCL-R respectively hold words A, B and C and X and Y. There is an obvious difference between the two lists and the six conjunct pairs derivable from them, that is, &lt;A, X&gt;, &lt;A, Y&gt;, &lt;B, X&gt; ... ) Reconnaissance consists of a single pass throvgh the inpitt string, during which, after lexical lookup, each word is indexed, a count is kept of the number of tokens of each category which occurs in the input string, and the PCL's are crcated for each conjunction. After Reconnaissance, if there are any conjunctions, the PCL's are filled subject to the Limits Constraint and the Position Constraint. The IAmits Constraint is applied only when PCL-L is filled, and the Position Constraint is applied only when PCL-R is filled. PCL-L is filled first. A word is put into PCL-L if and only if its index is less than the index of the conjunction with which the PCL-L is associated and the number offwords of this category in the string is greater titan one (when this second condition is met the Limits Constraint is satisfied). Thus when hits is encountered while the parser is attempting to fill PCL-L for the conjunction in (la), hits is not put into PCL-L since there is only one verb in the string. It can accordingly be determined that the conjunction is not coordinating predicates in (1 a), since there will be no verb ill either of the PCL's.</Paragraph>
    <Paragraph position="7"> In order to satisfy the position constraint when PCL-R is filled, a word is put into PCL-R if and only if its index is greater than the index of the current conjunction and there is already a word in the PCL-L for the current conjunction which has the same category as the word being considered for inclusion in the PCL-R for this conjunction. For example, in processing (7) John and Martha know Fred likes Dora The parser does not put either know or likes into PCL-R because there are no verbs in PCL-L.</Paragraph>
    <Paragraph position="8"> As will be discussed below, in the vast majority of cases in at least one domain the type of coordination occurring in a sentence cart be determined solely on the basis of these straighb forward principles. In these eases, the structure encountered is similar to that seen in (1 a). In order to determine whether predicates are being coordinated in structures like those seen in (lb)  and (lc) it is necessary to have somewhat more information about the input string.</Paragraph>
    <Paragraph position="9"> The additional information required to deal with strings such as (lb) and (lc), only one of which involves predicate co-ordination despite the fact that the two are nearly identical, concerns the relationships which obtain between predicates in a complex sentence. These relationships are enforced by constraints (5-6) above, in conjunction with</Paragraph>
  </Section>
  <Section position="6" start_page="285" end_page="288" type="metho">
    <SectionTitle>
(8) MULTIPREDICATE CONSTRAINT
</SectionTitle>
    <Paragraph position="0"> Every predicate in a multipredicate sentence must be in an ordination relationship with another predicate in the same sentence.</Paragraph>
    <Paragraph position="1"> The task of the parser confronted with polypredicational examples of the type in which we are interested is to distinguish coordination of predicates, as in (lb), from sub-/superordination, as in (lc). During the Attack phase of the parse, we capitalize on the fact that it is possible to resolve certain indeterminacies about the structure of a sentence on the basis of only incomplete information about the ordination relations which obtain in.the sentence. This depends on the fact that ordination relations can exist only in the presence of ORDINATION RELATION SIGNALS (ORS's). While space does not permit a complete discussion of ORS's here, some examples are subordinators (e.g. complementizers and subordinating conjunctions) and the marking of verbs like know and believe as allowing predicational objects.</Paragraph>
    <Paragraph position="2"> Here we will concentrate on subordinators. Each subordinator in a sentence r0ust be associated with a verb in that sentence, and this association causes that verb to be necessarily subordinate to some other predicate. The fact which is of value in parsing coordinate structures is that this can be known even before the superordinate partner of the subordinate predicate has been identified. For example in (lc) even before anything else is known about the structure of the sentence, it can be determined that the subordinator when is associated with hits and that therefore hits will have to be subordinate to some other predicate in that sentence. null As noted above, the parsing principles applied during Attack remove words from the PCL's. In the parse of (lb), while there are nouns and verbs in both PCL's at the beginning of Attack, all the nouns are removed, as Attack proceeds, from both PCL's, leaving only the verbs to be coordinated. The way in which Attack accomplishes this is as follows.</Paragraph>
    <Paragraph position="3"> There is more than one predicate in (lb) and thus the predicates have to be in an ordination relation in order to satisfy the Multipredicate Constraint. This relation cannot be subordination, since no subordinating ORS is present; assuming co-ordination to be the only other possibility, and given that there is a coordinating conjunction between the two predicates, we conclude that the predicates are in fact coordinate. In order to satisfy all of the constraints Attack must therefore remove John and Fred from PCL-L leaving hits as the sole member of that list. It must also remove guys and him from PCL-R leaving attack as the only word in that list. The configuration of these lists thus indicates that the only possible coordinates in (lb) are hits and attack.</Paragraph>
    <Paragraph position="4"> These same principles determine that predicate coordination cannot obtain in (lc). As Attack begins, PCL-L for the conjunction in this string contains John, hits', and Fred. PCL-R contains guys, attack, and him. Since there is more than one predicate in this string, the predicates will have to be in an ordination relationship, but it will have to be a relationship of subordination rather than coordination. Hits will have to be subordinate to some predicate in this sentence by vil'tue of the fact that it is associated with the subordinator when. (We do not state the means by which this is established here; see Rindflesch op. cit.</Paragraph>
    <Paragraph position="5"> for details.) Since hits is necessarily non-main, any predicate co-ordinated with it would also have to be non-main, by the Equiordination Constraint. Therefore it is not possible to coordinate attack with hits in (lc) since such a construal would cause the Main Predicate Constraint to be violated. The only possible ordination relationship which can obtain between the predicates in (1 c) is one in which hits is subordinate to attack. Therefore, hits must be removed from the PCL-L and attack must be removed from the PCL-R. From this it can at least be determined that (lc) does not involve predicate coordination.</Paragraph>
    <Paragraph position="6"> 3. Empirical Support for the Approach To test the effectiveness of the strategy described above, we subjected to analysis a corpus of nearly 16,000 words (15,985 to be exact). The texts used were specifications and design requirements (5 in all) applying to hardware manufactured by Control Data Corporation, supplied to us in machine-readable form. Each text was run through a concordance program which identified all tokens of and and or; and for each token of each conjunction, tile containing sentence was then analyzed (by hand). A total of 431 tokens of the two conjunctions occurred in the corpus, 362 of them in complete sentences (as opposed to section heads or fragments, which were ignored). As noted earlier, we did not, in undertaking the analysis, take into account the fact that there is widespread category-label ambiguity ('CLA')  in English; this represents a significant idealization of the data, but it is nol a cheat. The problem with regard to coordination with which we m'e concerned is that even in cases where no CLA occurs, problems of the sort exemplified by (1) arise. That the overall problem is even worse than we make it out to be does not invalidate our claims, though it meaus -- and we are fully aware of this -- th,~t tile account is incomplete.</Paragraph>
    <Paragraph position="7"> Of the conjunctions occurring in complete sentences, the type of coo,-dination in which each was involved was correctly ascertainable via application of the five constraints in 91% of the total number of cases, given only tile information made available by Reconnaissance plus the ORS-verb associations made early in Attack. 82 % of the total nmnbcr of cases were correctly identified solely on thc basis of the Limits Constraint and the Position Constraint. Of the remaining cases, at least 51% snbmit to re.solution during tile Attack phase on the basis of the comparativ,',ly low-level structural information concerning ordination relations (Main Predicate, Equiordination, and Multipredicate C,mstraints). (This figure is conservative in that further principles may be identified in the future which would improve performance.)</Paragraph>
  </Section>
class="xml-element"></Paper>