<?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="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
I. INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> We interest ourselves in the analysis phase of a machine translation system which adopts the indirect, transfer and global approach (see \[Slocum 84\]). The aim of this paper is to clarify the problem areas, suggest a few solutions, and point out the loose ends. There is no current implementation for the analyser we describe, and the ideas are basically a reflection of ~hat ~c would like to see in an MT system, based on previous experience in the field. A very important issue is to separate the linguistic knowledge (the grammar) from the algorithmic and technical knowledge (the organ\]sat\]on of the analyser, pattern matching, etc.). &amp;quot;Approximate&amp;quot; linguistic knowledge is also separated and used as a means to guide the analysis rather than considered as absolute (semantics and context constraints as heuristics instead of grammar rules).</Paragraph>
    <Paragraph position="1"> Due to space restrictions, we shall immediately specify the basic type of analyser we shall be working with, without giving any reasons for the choice. The interested reader is referred to \[Zaharin 85\] for an uncondensed version of this paper, and \[Zaharin 86\] for more details.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2. THE ANALYSER
</SectionTitle>
    <Paragraph position="0"> In general, an analyser can be viewed as a black box with two holes, where we insert the input text through one and it gives the output linguistic structure through the other (in our case, an annotated tree representing the &amp;quot;meaning&amp;quot; of the input text). Peeping into the box, we would notice that it works in cycles doing the following five steps until it triggers off some stopping mechanism and hence furnishing the output : a) computing the object set ; b) choosing an object ; c) computing the rule set ; d) choosing a rule ; e) applying the chosen rule on the chosen object.</Paragraph>
    <Paragraph position="1"> Naturally, depending on the various models, these steps need not be executed in the given order, nor are they necessarily as clear out. Indeed, some may even execute the cycles in parallel.</Paragraph>
    <Paragraph position="2"> Our analyser will do the five steps in the following manner. Steps (a) and (c) will be done together, computing all objects on which some rule is applicable, and to each of these objects, the set of all applicable rules is computed. The result is a set of \].inked pairs O-R where R is a rule applicable on the object O. A linked pair is then picked, i.e. steps (b) and (d) together, and the chosen rule applied on the chosen object. The cycle repeats. The motivation for the above choice is that we are aiming for a one-go analysis, for which we shall be needing the maximum of information before we apply a rule, hence the computation of all candidate objects and rules. Strategies and heuristics are then needed for the critical choice of object-rule pair in each cycle.</Paragraph>
    <Paragraph position="3"> ~L~ natural language treated will bedescribed by a grammar containing a set of rewrite rules with a context free base of the form Xl...Xn ~JX where Xl,...,Xn,X are annotated trees ; in other words, an augmented context free grammar. What we actually have in mind is a grammar containing rules of the form given in figure i, as discussed in \[Zaharin 86\]. Nevertheless, the discussion remains valid for any system using a similar representation of data.</Paragraph>
    <Paragraph position="4"> AS in most machine translation systems, the analysis looks for only a single solution, i.e. a single representation of meaning for each input text. If the text is ambiguous, the &amp;quot;best&amp;quot; solution is taken. In the search for a solution, a depth first approacb is taken, and the analyser allows for backtracking in case the solution is not found in one go. Backtracking is also required in cases where an input sentence is not in the language of the gralnmar, but most important, to ensure that the analyser finds a solution if there is one.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="136" type="metho">
    <SectionTitle>
3. THE PROBLEM AREAS
</SectionTitle>
    <Paragraph position="0"> With the type of analyser we have chosen, the problems that arise are basically the following :  - pattern matching ; ambiguities ; - forward propagation ; checking for correctness ; ~ backtracking.</Paragraph>
    <Paragraph position="1">  Pattern matching seems to be the bottleneck of the real\]sat\]on of any system. Fortunately, the literature already contains some efficient pattern matching procedures that can be modified to suit our model.</Paragraph>
    <Paragraph position="2">  The choice of an augmented context free gralmnar means that the rules are basically in the form of strings of symbols, where each symbol is augmented with an annotated tree structure. Figure i gives an example of a rule we use (see \[Zaharin 86\]). In this form, the pattern matching can be carried out in two stages : one for strings, followed by one for trees, ~\]ere the latter (the more costly one) is triggered only in cases of success of the former.</Paragraph>
    <Paragraph position="4"> For instance, at the string level, this falls into the category of many pattern/many object pattern matching for strings,for which the procedure of \[Aho &amp; Corasick 75\] which finds all patterns and all objects in one pass seems suitable. Only in cases of success do we pass on to a tree pattern matching process, for instance that of \[Hoffmann &amp; O'Donnel 791\]. Repetitive work can be avoided if we factorise the results of the pattern matching from one cycle to the other.</Paragraph>
    <Paragraph position="5"> Ambiguities and forward propagation are the two major problems for the model. We defer the discussion on t~ese to the next two sections.</Paragraph>
    <Paragraph position="6"> In genecal, it is very difficult to describe a natural language exactly by means of a formal grammar, no matter how sophisticated the formalism.</Paragraph>
    <Paragraph position="7"> In spike of this, the criteria for the correctness of the result of the analysis is usually with respeck to the natural language treated and not that of the grammar, i.e. finding an axiom in the gra1~nar may not be sufficient. So, rather than writing very strict rules at the risk of excluding correct structures, it is better to have more genera\], rules which may accept anomalous structures, and then provide a filter to reject such results. \[Boitet &amp; Gerber 84\] suggests an expert system to do this post-analysis checking.</Paragraph>
    <Paragraph position="8"> For backtracking analysis, a simple model is to store failed configurations (dead states) in some file. At the beginning of each cycle, the new configuration can be checked against this file, backtracking further if it compares. This may seem a huge effort, but natural language analysis is such that many identical nodes may be found in different parts of the search space. As for forward propagation in the backtracking analysis, the priority orderings to be discussed in 5 can be preserved and made use of here.</Paragraph>
  </Section>
  <Section position="6" start_page="136" end_page="136" type="metho">
    <SectionTitle>
4. AMBIGUITIES
</SectionTitle>
    <Paragraph position="0"> Ambiguities haunt every treatment of a natural language. \[Lepage 85\] summarises the types of ambiguities that we face, both lexical and structural, while \[Van Klinken 84\] writes on the methods used to solve some of the cases. However, until a formal treatment of ambiguities can be proposed, the solutions will remain ad hoc, treating case by case as we meet them.</Paragraph>
    <Paragraph position="1"> In general, lexica\] ambiguities are solved either grammatically, with context, or with semantics. Grammatically is as in the sense of using agreement in number to obtain &amp;quot;that&amp;quot; as a conjunction instead of a determiner in the sentence : We know that ambiguities are difficult to solve.</Paragraph>
    <Paragraph position="2"> We use context to distinguish the past participle &amp;quot;collected&amp;quot; from the verb in the two sentences : The corals co\] lected at the bottom of the sea are beautiful.</Paragraph>
    <Paragraph position="3"> The corals co\] lected at the bottom of the sea.</Paragraph>
    <Paragraph position="4"> Finally, \[Lytinen 85\] polnted out the need of semantics to determine the attachraen't of the prepositional noun phrase &amp;quot;:\[or '~iO&amp;quot; in the two sentences (based on the verb &amp;quot;found&amp;quot;) : The cleaners dry-cleaned the coat that Mary found the rummage sale for ~. at The cleaners drv-cleaned_ -- the coat that Mary f oun&lt;!. in the garbage foE__,}! q.</Paragraph>
    <Paragraph position="5"> Whereas we can be quite certain of the so\]utio~ obtained grammatically, the use of context or semantics does not inspire the same confidence. Context can pose problems when locating the elements the context refers to, which can be arbitrarily far away from the ambiguous word. Furthermore, the problem can he aggravated by the elements looked for being ambiguous themselves. Sometimes, negative constraints are used ~n context e\].ements, and these carl pose interpretation problems (see \[Zaharin 86\]). As for semantics, the arguments can be endless.</Paragraph>
    <Paragraph position="6"> Bearing the above in mind, we prefer to treat * k\]-te solution of \]6~5{I(Ia1 amhlg|\]i~je m m~ hel lrist{C/ ~, rather than steadfast rules. By this we mean that context and semantics should not be incorporated into the grammar rules used to describe the language treated, but instead should be placed in related heuristic rules which advise on the apmlicability of their counterparts. This also means that if their advice has not led to a success, it is possible to backtrack to the same rule and recommence, this time ignoring the advice. The case would not have been possible if the grammar rule and the context and Semantics had been put together in one rule.</Paragraph>
    <Paragraph position="7"> In the case of structural ambiguities, the sentence can be inherently ambiguous, in which case context and semantics heuristic rules can only aid to pick the preferred reading. It is also possible that structural ambJguitLes occur only at the level of substrings of the sentence, but some of the possibilities will not lead to a solution. In such a case, heuristic rules for preferred readings will also help, but the problem is more of choosing a rule or object to avoid leading to a dead end. This falls into the categorv of problems to be discussed in the next section.</Paragraph>
  </Section>
class="xml-element"></Paper>