<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-3141">
  <Title>HIGH-PROBABILITY SYNTACTIC LINKS</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Bottom-up Parsing
</SectionTitle>
    <Paragraph position="0"> The processing of a sentence begins with morphological analysis. As a result, each word is given a set of its possible lexico-morphological interpretations, henceforth called &amp;quot;homonyms&amp;quot;. A homonym is a list that includes a lexeme identifier, a part-of-speech marker, and morphological features of the wordform.</Paragraph>
    <Paragraph position="1"> For instance, the morphological module of the ETAP-2 system (Apresjan et al. 1989) will give for the word sawtbe following three homonyms: SEE, V, pt (-past tense); SAWl, V, mf (-main form); SAW2, N, sg.</Paragraph>
    <Paragraph position="2"> All morphological data are concentrated in a special morphological dictionary. The key role in parsing proper is played by a combinatorial (syntactic) dictionary that contains versatile information on syntactic properties of lexemcs, i.e. on their ability to participate in various syntactic constructions (for details see Mel'~uk 19&amp;quot;/4, 1988; Apresjan et al. 1989, 1992).</Paragraph>
    <Paragraph position="3"> The general scheme of parsing is as follows. After the morphological analysis, for each word there appears one or more homonyms. By &amp;quot;fragment&amp;quot; we shall understand a set of homonyms occupying one or more successive positions in the sentence (one homonym in each position) plus a tree of syntactic links defined on these homonyms as nodes. For instance, an isolated homonym is a trivial fragment; the whole dependency tree of a sentence is also a fragment. It should be noted that in trees (2) - (4) each word is represented by a certain homonym (for example, saw is represented by SEE, V, pt).</Paragraph>
    <Paragraph position="4"> Lejkina and Tsejtin (1975) described a bottom-up process for constructing dependency trees. It is based on the operation of adjunction. This operation is applied to two adjacent fragments and consists in establishing a link, marked by a certain syntactic relation, from a certain node of one fragment to the root of the other. The result of adjunction is a new fragment on the union of segments occupied by the initial fragments. null This action is similar to generation of a new constituent from two adjacent constituents. However, unlike constituents, fragments at the moment of adjunction may be &amp;quot;incomplete&amp;quot;, i.e. they need not contain all the nodes that will be direct or indirect dependents of their roots in the structure of the sentence. These nodes may be added to them later (also by the operation of adjunction).</Paragraph>
    <Paragraph position="5"> Mitjushin (1985) described the class of trees that can be constructed from isolated homonyms by repeated adjunction, i.e. that can be built by the bottom-up process. Consider a tree with an ordered set of nodes. Let a &amp;quot;tangle&amp;quot; be a quadruple of nodes (a, b, c, d) with the following properties:  1) a&lt;b&lt;c&lt;d; 2) a and c are linked by an arc (in any direction); 3) b and d are linked by an arc (in any direction); 4) the path between a and d contains neither b nor c (here, orientation of arcs is ignored, so the  path always exists).</Paragraph>
    <Paragraph position="6"> The following criterion is true: a tree can be constructed from its nodes by repeated adjunction if and only if it contains no tangles.</Paragraph>
    <Paragraph position="7"> The simplest tangle looks as follows: a b c d (direction of the arcs does not matter; there can be other nodes between a, b, c, and d). According to the criterion, a tree that contains such a subtree cannot be built by the bottom-up process.</Paragraph>
    <Paragraph position="8"> The class of trees obtainable by adjunction is much wider than the class of so-called projective trees (on projectivity see, for example, Gladkij 1985; Mel't~uk 1988). For the model of Russian syntax presented by Mel'~uk (1974) and Apresjan et al. (1992), this class includes all syntactic structures permissible in scientific, technical, and business texts (however, it is not so for spoken language and poetry). We suppose all the structures considered below to belong to this class.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Rules
</SectionTitle>
    <Paragraph position="0"> In our system, in contrast to those based on forulal grammars, the rules are not a tool for the exhaustive description of the set of correct syntactic structures.</Paragraph>
    <Paragraph position="1"> We suppose that the correspondence between sentences and their syntactic structures is defined by some other means. The task of the parsing algorithm and the rules it employs is to build, for a given sentence, some set of its syntactic structures or their fragments, without losing the semantically correct ones.</Paragraph>
    <Paragraph position="2"> The concrete function of the rules is to check whether the given case of adjunction is feasible and, if so, to perform the operation of adjunctioa. Some additional operations can also be performed. The rules have access to any information about the structure of fragments to be adjoined and the homonyms they contain (their lexeme names, morphological features, and syntactic properties stated in the combinatorial dictionary). The rules may also use data on punctuation and limited data on homonyms not belonging to the given two fragments; they have no access to information about fragments built by the algorithm earlier.</Paragraph>
    <Paragraph position="3"> While formally the rules could be strongly context-sensitive within the limits of two given fragments, in most cases they only use information on nodes X and Y (those to be linked) and their nearest syntactic Context. In fact, the rules currently emloyed do not ACTES DE COLING-92, NANTES. 23-28 Am~'r 1992 9 3 1 PROC. OF COLING-92, NANTES, AUG. 23-28. 1992 consider nodes for which distance from X or Y exceeds 3 (where distance is the number of links in the path connecting two nodes in the dependency tree of a fragment).</Paragraph>
    <Paragraph position="4"> A rule is a program written in the form of a transilion graph, with an elementary predicate or operation associated with each arc. The rule interpreter performs ordered search to find a path along &amp;quot;true&amp;quot; arcs that starts at a fixed entry node and ends at one of fixed exit nodes. No backtracking is used: if forward motion from some node proves to be impossible, interpretation is terminated. The fact that backtracking is not necessary has been discovered empirically; it is connected with the nature of syntactic events considered by the rules. On the other hand, when desirable, an explicit return may be made to a point passed earlier, with simple measures taken against infinite cycling.</Paragraph>
    <Paragraph position="5"> Each rule contains at least one operation LINK(X, Y,R) that establishes a link marked by a certain syntactic relation R between the given node X of one fragment and the root Y of the other (that is, performs the adjunction). The corpus of rules covers only those situations for which the probability that the established links are correct is estimated as close to 1. For instance, the rules do not establish links like alter and ad~v in structures (3) and (4) because altachment of prepositional postmodifiers is known as a &amp;quot;classical&amp;quot; case of structural ambiguity.</Paragraph>
    <Paragraph position="6"> It should be noted that the probability close to 1 characterizes here not individual links (it would be too strong a demand) but all complex of links established for the given words. This can be illustrated by the segment 1 saw, for which two fragments will be built with different homonyms for the word saw: pred pred I~-- SEE V.pt, I~ SAWI V,mf.</Paragraph>
    <Paragraph position="7"> Both these alternatives are syntactically correct. At the same time, they are mutually exclusive, and it is only their disjunction that has probability close to 1.</Paragraph>
    <Paragraph position="8"> This ambiguity is also inherited by larger fragments.</Paragraph>
    <Paragraph position="9"> (As a result, the sentence 1 saw a man with a telescope has four different parses, two of which are semantically abnormal.) Thus, high probability is a &amp;quot;rollerfive&amp;quot; and not an &amp;quot;individual&amp;quot; property of links. Rigorous definitions can be found in the paper by Miljushin (1988).</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 The Parsing Algorithm
</SectionTitle>
    <Paragraph position="0"> The simplest method of bottom-up parsing is to consider all opportunities for adjunction, starting from adjacent one-element fragments. We employ a faster algorithm, in which certain heuristics are used to reduce search (Mitjushin 1988).</Paragraph>
    <Paragraph position="1"> The algorithm builds a growing sequence A of fragments. At any moment of time A contains some homonyms of the sentence and certain fragments constructed of these homonyms. The algorithm moves from the beginning of the sequence A to its end and tries to perform adjunction between the current fragment F E A and the fragments that appear in A earlier than F. New fragments are added to the end of the sequence.</Paragraph>
    <Paragraph position="2"> The fragment cmtsidered at the given moment is called active. All fragments of A (including isolated homonyms) become active successively, without leaps or returns.</Paragraph>
    <Paragraph position="3"> While the algorithm moves along the sequence A, tile sequence grows longer because of the addition of newly built fragments. Nevertheless, a moment will necessarily come when the active fragment is the last in A and further motion is impossible. In this case, the next homonym of the sentence is added to the sequence; it becomes active and the work is continued. When a new deadlock arises, another homonym is added, and so on. If in such a situation it turns out that all homonyms of the sentence are exhausted, then the work is finished.</Paragraph>
    <Paragraph position="4"> Homonyms are added to the sequence in the order they are arranged in the sentence from left to right (which is essential), and those occupying the same position are added in an arbitrary order (in this case, the order has no influence m~ the results). At the initial moment A contains a single element, namely one of the homonyms occupying the leftmost position of the sentence, and it is declared active.</Paragraph>
    <Paragraph position="5"> For each active fragment Flhe algorithm selects in A its left neighbors, i.e. fragments thai are adjacent to F on its left. A preference relation is defined between tile neighbors of F: fragments of greatter length are preferred, and lhose of equal length are considered equivalent.</Paragraph>
    <Paragraph position="6"> For the given F, the algorithm considers its left neighbors E in order of their preference, and for each E tries to adjoin it to F. If for some E adjunclion is successful, subsequent search is limited to the neighbors of F equivalent to E; less preferred fragments are not considered.</Paragraph>
    <Paragraph position="7"> An attempt to adjoin E to F is made as follows.</Paragraph>
    <Paragraph position="8"> IJnks are considered that connect a certain node X of fragment E with the rightmost node Y of fragment F.</Paragraph>
    <Paragraph position="9"> A preference relation is defined between the links: those of greater length are less prefen'ed, and those of equal length are equivalent. In other words, more preferred are links X -- -Y and X ~ Y with nodes X that are nearer to the Jight end of E; links with the same X are equivalent.</Paragraph>
    <Paragraph position="10"> For the given E and F, nodes X ~E are considered from right to left (i.e. in order of the preference of links between X and Y), and for each X the rules applicable to these X and Y are activated. The list of such rules is determined by parts of speech of X and Y, and by possible direction of fhe link. If during interpretation of a rule an operation LINK(X, Y, .) or LINK(Y,X, .) is performed then a new fragment is built which is the ~'esult of joining X and Y with the given link. It is placed at the end of tile sequence A. After flint, for these E and F the search is limited to the links equivalent to the established one; less preferred links are not considered.</Paragraph>
    <Paragraph position="11"> When the sequence A is built, its subset C of maximal fragments is formed. A fragment is called maximal if its segment is not a proper part of the segment of any other fragment belonging to A. The set C is the final result of partial parsing. Below, when speaking ACTES DE COLING-92, NANTES, 23-28 AO(ZI' 1992 9 3 2 PROC. OI' COLING 92, NANI KS. AUG. 23-28. 1992 about fragments built by the algoritlnn, we shall al.ways mean exactly tim set C.</Paragraph>
    <Paragraph position="12"> The first experiments with this algorithm have shown titbit, tit sonte c;.tses, the preferences attd restrictions adopted arc too strong and pvtme away semantically correct parses. To intprove tire situation, special operations were defined that made it possible to cancel (from inside lhc rule) priority of longer neighbors or shorter links, and also to make the algm'ithm cmrsider not only the rightmost node of tire right fragment. ()wing to them, the search can be made exhaustive in all cases when the rule &amp;quot;considers it desirable&amp;quot;. In tile real process of pursing, these operations are fired not too often, so tile main part of search remains limited.</Paragraph>
  </Section>
class="xml-element"></Paper>