<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1030">
  <Title>STRATEGIES AND HEURISTICS iN THE ANALYSIS OF A NATURAL LANGUAGE IN MACHINE TRANSLATION (In the memory of Bernard Vauquois)</Title>
  <Section position="7" start_page="136" end_page="138" type="concl">
    <SectionTitle>
5. FORWARD PROPAGATION
</SectionTitle>
    <Paragraph position="0"> Forward propagation is the problem of choosing a rule and an object for application in each cycle of the analyser. This is the execution of steps (b) and (d) in section 2, which is then fell.owed by step (e), completing the cycle. As we are aiming for a solution in one pass of the analysis, the choice is critical, as even a wrong choice of a sequence of applications may lead to a dead end. This can be seen J n the fo\] \].owing example where the grammar contains the rules (omitting the details) :  R 1 : NP VK NP + VCL ; R 3 : RELCL PNP / RELCL R 2 : NP RELCL / NP ; R 4 : VCL PNP + VCL  Taking an example in \[Lytinen 85\], the analysis may find itself in the stats given in figure 2 (the candidate objects are circled and the corresponding rule indicated). The sequence of applications needed in this example is R2 R1 R 4. If we happen to choose R 1 , before R2, we will find that the analysis will not lead to a solution.</Paragraph>
    <Paragraph position="1"> ....</Paragraph>
    <Paragraph position="2"> .... :i: .!:, ,!, ,,o / h6 The situation given here is one of the major problems faced by analyses which predefine the sequence of applications of rules. There is nothing more frustrating than not obtaining a complete analysis and yet knowing that the required rules are present in the grammar.</Paragraph>
    <Paragraph position="3"> Instead of predefining a sequence of rule applications, we prefer using heuristic rules which apply independently in each cycle of the analyser. These heuristic rules act to determine a priority ordering of the candidate rules and objects (each rule will be tied to the object it is applicable on), the highest priority rule or object being chosen for application (taking along the object or rule it is tied to).</Paragraph>
    <Paragraph position="4"> The big question is, what should these heuristic rules contain ? First and foremost, coming from the discussion on solving ambiguities in section 4, we need the treatment of semantics put down as heuristics. An example of such a case is in figure 2 where rule R 3 should be accompanied by a heuristic rule to check for semantics. Here, one does not &amp;quot;find&amp;quot; something &amp;quot;in the garbage&amp;quot; &amp;quot;for ~i0&amp;quot;, and so the heuristics would advise that R 3 should not apply (unless, as discussed before, following this heuristics leads to a dead end, and so we come back to apply R3).</Paragraph>
    <Paragraph position="5"> We shall refer to the type of heuristics just used as the &amp;quot;to-apply-or-not-to-apply&amp;quot; heuristics. The type of heuristics mainly needed is the &amp;quot;afteryou-or-after-me&amp;quot; heuristics. This is the case for the choice between R1 and R2 in figure 2.</Paragraph>
    <Paragraph position="6"> For the said problem, one may argue that VCLs are higher up inthe hierarchy of phrases and clauses than NPs \[Vauquois &amp; Chappuy 85\], and so rules building NPs should be applied before rules building VCLs. This may be true in this example, but care must be taken when we deal with complex clauses and phrases (the hierarchy given in the reference is for simple clauses and phrases). For complex clauses and phrases, we may obtain cyclic hierarchies between NPs and RELCLs, APs and NPs, etc. For such cases, ad hoc heuristics are needed, for instance, rules building RELCLs should apply before rules building NPs if the former is found to the right of the latter, and the inverse otherwise.</Paragraph>
    <Paragraph position="7"> Apart from some hierarchy given, context can also be used to solve the &amp;quot;after:you-or-after-me&amp;quot; problem. (Recall that context is also needed to solve ambiguities). As examples, suppose the grammar for figure 2 also contains the rules (still omitting</Paragraph>
    <Paragraph position="9"> Checking the context, namely the conjunction &amp;quot;when&amp;quot; or &amp;quot;that&amp;quot;, can be used to choose R5 on &amp;quot;the king rides&amp;quot; in figure 3, while RI is chosen in figure 4 (this also gives an example as to why we would not use heuristics like &amp;quot;apply the rule with the longer LHS&amp;quot;).</Paragraph>
    <Paragraph position="10"> .~,d Rl ,C~) &amp;quot;~ vK ) When the king ries the/t h~orse i s-gr!omed That ~~c h6rse is unbelievable F i/gur e 4 In the example in figure 3, it so happens that the two occurrences of the rules R5 are independent, in the sense that the application of one before the other has no great consequence, and so an arbitrary choice will do. However, not competing on intersecting objects does not necessarily guarantee independence. Had we been two steps before figure 2 with rules : RIO : NPR NP VK + RELCL ; 1{I 1 : PNP PNP + PNP included in the grammar, the situation would be as given in figure 5. Here, semantic heuristics can advise that RII should not apply. However, we need to make sure that RIO applies before RI, otherwise we can never arrive at a complete analysis even though these objects seem to be independent.</Paragraph>
    <Paragraph position="11"> NP NPR NP RIO VK PNP RII PNP ' 1 /\ I I I. /h/2.o The cleaners dry-cleaned the coat that Mary found in th g g ~u I ! No doubt the above problem can be solved using the same heuristics giving the hierarchy of clauses and phrases, but this situation brings up two important questions which necessitate answering : firstly, how do we expect such situations ? And secondly, do we have to know all such situations before we can write a good set of heuristic rules ? Before attempting to answer these questions, we wish to highlight the heuristics used by \[Nagao &amp; Nakamura 82\] which looks very promising but may suffer from the same drawbacks. The reference uses sentential patterns (SP) that express the global structure of a sentence. If these SPs are satisfied by a certain configuration, the said configuration is chosen for expansion (they use a best-first search where a configuration is then a node in the search space).</Paragraph>
    <Paragraph position="12"> For our purposes, SPS can also be adapted either to place checks testing whether the analysis is on the right track, or to create subgoals if some of the constituents of the SP are satisfied.</Paragraph>
    <Paragraph position="13">  This can be useful for configurations containing specific words which can determine its neJ ghbours. For exalnple, a conjunction necessitates a VCL or PARTCL to its right.</Paragraph>
    <Paragraph position="14"> Going back to the two questions posed earlier on, the problem of expecting the situations where heuristic rules can be written is not a simple one. For a given derivation tree in a context free grammar, any cut in this tree is a possih\]e configuration of the correct analysis. Passing this cut through the. pattern marcher will give the complete configurationdeg Looking at ali possible subskrlngs of this cut, and multiply 'this by the nu~nher of alI cuts in the tier\]-ration tree will give us the situations we need to predict. The result is by no means negligible, to say the ieast.</Paragraph>
    <Paragraph position="15"> Fortunately, rules that can apply on intersecting objects can be precomputed. In particular, if we use the pattern matchi.ng procedure of \[Aho &amp; Corasick 75\] as mentioned earlier, the procedure produces a network equipped with a failure :\[unction indicating to which part of another rule (say Rb) the pattern matcher is to go to after successfully finding a pattern (say rule Ra). This gives a possible clash between rules Ra and Rb, where Ra is on the \]eft of l{b. For example, the clash between R1 and R2 in figure 2 can be predicted by the pattern marcher, to which a heuristic rule can be written, say the one given in figure 7. Figure 6 gives the network for the pattern matcher of the reference for the rules R\] to R4 in our example. The failure function is given by f(i) where i is a state of the network while the output of applicable rules is given by out{}ut(i) (again we omit the details of augmenting ea, ch arc by the TREE value). ~e refe~ the reader to the reference for further details.</Paragraph>
    <Paragraph position="16">  As for having to predict on possible situations we can cut down on some work by making the analyser &amp;quot;reason&amp;quot; a littie. For examp\].e, the analyser should be able to deduce from the situation in figure 5 that it can get to the situation in figure 2 and hence apply the heuristic rule already written for figure 2 (in this case the one given in figure 7).</Paragraph>
    <Paragraph position="17"> This reasoning can be done in the following manner with R1 applicable on &amp;quot;NP VK NP&amp;quot; (see figure 5), the failure function for R1 points to state \] (see figure 6), and with the applicable rule to the immediate right of R1 being RiO which produces a RELCL, this gets us to state 4 with output R2.</Paragraph>
    <Paragraph position="18"> We then obtain a sIightly different situation from figure 2 but the heuristic rule can still apply, giving priority to rule R2 hence RIO.</Paragraph>
    <Paragraph position="19"> Despite the title, we have hesitated on discussing strategies, because experience tel\]s us that it is very difficu\].t to write admissible strategies (i.e. set sequences of heuristic rules). Furthermore, strategies may be as risky as procedural methods unless they are flexible enough. This means that they can be halted, created, interrupted and resumed during the analysis. Furthermore, they ought to be global rather than particular. For example, the hierarchy of clauses and phrases can serve to choose between rules having the same priority after other heuristic rules have applied, and halted when compIex structures are treated. An interesting discussion on global and flexible strategies is \[ound in \[Hayes-Roth 85\] for the expert system OPM.</Paragraph>
  </Section>
class="xml-element"></Paper>