<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-2102">
  <Title>Interaction between Structural Changes in Machine Translation</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
* Effective Exploitation of Domain/'l~xt Type-
</SectionTitle>
    <Paragraph position="0"> Specificity (or Sublauguageness) in MT and Discovery Processes for such Specifieities Though new frameworks of MT such as Statistics-based MT (SBMT), F, xample-bascd MT (F, BMT), Analogy-based MT (ABMT), Knowledge-based MT (KBMT) etc. look radically difihrent from conventional liuguisties-based MT (LIIMT) anch as Transfer-based MT, they address role or two of the above focal issues and ignore the rest\[7\]\[10\]. In particular, the new paradigms of MT tend to ignore t.he first issue ie. declarative representation of bi-lingual knowledge and the complexities involved in the process of combining units of bi-liagual knowledge.</Paragraph>
    <Paragraph position="1"> It is our contention that any MT system, whichever paradigm it belongs to, has to store a set of translation-equivalent units for a pair of languages and combine these to produce larger units of translation. In EBMT, for example, a set of translation examples has to be stored and several of them have to be combined properly to produce translation. Because of the declarative nature of translation exampies, EBMT inevitably encounters the same complexitics of combining translation units in declarative representation o.~ I,BMT does.</Paragraph>
    <Paragraph position="2"> tLescarch in LBMT\[1\]\[2\]\[3\]\[ll\]\[12\]\[13\] has revealed that dittieulties in tim declarative representation of bi-lingual knowledge stem mainly from the treatment of idiosyncratic structural changes caused by lexieal items, and interactions of such idiosyncratic structural changes when they co-exist in single sentences. These type of structural changes also cause problems when they are combined with general or ordinary linguistic phenomena soeh as coordination. A formal framework to cope with these matters is essential in other aPl)roaches, such as EBMT, as in LBMT, if the translation is produced in a compositional way.</Paragraph>
    <Paragraph position="3"> In this paper, we discuss problems caused by such interaction, and give solutions in our logic-based transfer framework\[8\]. Our transfer model views the transfer process as deduction; it produces the target linguistic description without changing the source description. This giw:s a clear contrast to the conventional tree-traasducer model, which gradually tranSforms the source description, and is therefore hard to control.</Paragraph>
    <Paragraph position="4"> Because of the logic oriented nature of our framework, it can also be extended easily to cope with the other local issue of MT such as the integration of knowledge-based processing with translation\[4\].</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Formalism
</SectionTitle>
    <Paragraph position="0"> We use a simple formalism, representing a sentence as a network of semantic dependencies. The basic units of the formMism are indices, properlies, and re-Istwna. A logical form consists of an unordered net of Ierms; each term is either a property predicated of an it~dex, or ~t relation between two indices. The written notation depicts properties and relations as unary and binary functors, respectively, and indices as theiL&amp;quot; arguments, i.e. within brackets. A logical form representing 'John saw Mary' might be</Paragraph>
    <Paragraph position="2"> This representation in simplified particularly in that  the relation tense here stands for what should be a complex representation of tense, aspect, and aktionsart, related anaphorieally to the preceding discourse. It can be seen that the representation includes the names of the concepts, both objects and events, which are mentioned in the sentence, aud the semantic relationships between them. Each of the three indices, j, m and e, is a unique label given to a node in the dependency network. The indices also serve as discourse referents when the logical form is incorporated into a discourse representation. The root node of the network is specified at the beginning of the representation, in tiffs ease e, the node representing the seeing event. In terms of discourse representation theory, e is tim discourse referent of which the logical form is a description. As long as the root node of the network is specified, it is possible to convert a logical form mechanistically into a structured representation such as a dependency graph or a feature structure such as tense : past subject : \[predicate : john\] object : \[predicate : mary\] Some other examples of logical forms are shown in figure 1. The particular characteristics of the formalism -- the types of relations and predicates used and the manner of their use -- are justified elsewhere\[9\]. We only state here that tbe forrualism can treat the phenomena which are treated by traditional formalisms, along with additional phenomena relating to discourse structure and lexical semantics.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Transfer
</SectionTitle>
    <Paragraph position="0"> We follow the standard division of a machine translation system into stages of analysis, transfer, and generation. The parts of the system include algorithms for analysis, transfer, and generation, descriptions of the strneture of individual languages (used by the analysis and generation algorithm.s), and descriptions of equivalence relationships between pairs of languages (used by the transfer algorithm). It is a requirement for the transfer description that it should state all and only the equivalence relationships between expressions of a pair of languages. It should contain no general linguistic information, eltiler universal or specific to any single language: these types of information belong ill the formalism and content, respectively, of the descriptions of individual languages. In fact most of our transfer descriptions consist of direct equivalences between predicates or groups of predicates of the language pair.</Paragraph>
    <Paragraph position="1"> A transfer rule shows a particular translation equivalence between two languages. It consists of three parts: a pair of logical forms, one for each language, representing tile equivalence, and a third logical form giving the conditions under which the equivalence is applicable. We call these the equiva~ lenee pair and the condition respectively. Two logical forms form an equivalence pair if the natural language expressions they represent have overlapping denotations (using 'denotation' in a very broad sense to encompass not only the referential possibilities of nouns but the possibility in general of applying a predicate to a particular index). The rule can therefore be used for translating in either direction: the two logical forcers of the eqnivalenee pair are always translationally equivalent if the condition is satisfied. The logical forms of the equivalence pair will be indistinguishable from logical forms of the respective languages, using the same predicates and relations. The logical forms of the condition use meta-predieates which allow reference to to the logical form of the current local translation unit (sentence or clause) and to the linguistic context. In practice, most transfer rules have no condition; they consist just of an eqnivalenee pair. Some examples of rules follow:</Paragraph>
    <Paragraph position="3"> hypothetical(El,like(El -tai(F),omou(E) would like iff inLeft(obj(E,F)) ~tal omou The following is the basic definition of transfer between two texts (sentences, clauses, or whatever): A source logical form and a target logical form are correctly transferred if tile terms of each can be divided into non-overlapping sub-sets such that the source subsets can be placed into one-to-one correspondence with the target subsets by each corresponding pair being 'unified' with the two halves of the equivalence pair AcqEs DE COLING-92, NANTES, 23-28 AOt~T 1992 6 8 0 PROC. OF COLING-92, NAI~rEs, AUG. 23-28, 1992 of a transfer rule, and if the accumulated conditions of the transfer rules used are satisfied.</Paragraph>
    <Paragraph position="4"> 'Unification' in this definition is rattmr similar to graph unilication.</Paragraph>
    <Paragraph position="5"> q~ansfer rules specifying larger sets of terms can be used to state trauslational equivalences for idioms and fixed expressions. Simple idioms present no particular problem. One translation of the English idiom to kick the bucket into Japanese is sinu 'die', and this can be represented by the transfer rule kick(K) $z. obj(K,n) &amp; the-bucket(B) ~ sinu(K) Fixed expressions, appropriate only to a specific context, may require conditions which refer to the discourse representation.</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 'Complex' transfer
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.1 Types of complex transfer
</SectionTitle>
      <Paragraph position="0"> Lindop &amp; Tsujii (1991) list and discuss a variety of examples that are alwaYs problematic for machine translation systems. We suggest that what makes these examples difficult is that different dependency structures are associated with otherwise synonymous lexical items. We break the problem down into four subparts.</Paragraph>
      <Paragraph position="1"> i) Argument-switching as in the translation of the German Der Wagen gef'~llt mir -- I like the car The car pleasen me</Paragraph>
      <Paragraph position="3"> In argument-switching, the relations between the lexical item and its dependents are not translated standardly. Although the German nominative normally corresponds to an English subjcct, it must in this example be translated as a dative.</Paragraph>
      <Paragraph position="4"> it) IIead-switching as hi the translatlolt of Gernlan null tlaas sehwimmt gem -- John likes swinmfing John swinm ~ladly</Paragraph>
      <Paragraph position="6"> The Germau sentence is a description of a situation to do with swimming; the English is a description of a situation to do with liking. The shift predicate is explained below.</Paragraph>
      <Paragraph position="7"> iii) Decomposition as in the translation of the</Paragraph>
      <Paragraph position="9"> llere the single Japanese word jisatusnrn is tra~l~lated as the English phrase to commit suicide. Some types of decomposition can give rise to special problems when there arc modifiers or dependents involved.</Paragraph>
      <Paragraph position="10"> These are discussed in detail by Tsujii et al. (1992), Shortage of space aml tile variety and complexity of examples prevents their discussion here.</Paragraph>
      <Paragraph position="11"> iv) ttaistng as in the translation of the Welsh Fe ddigwydd rod Si6n yma happells be ~l o|ln here John happens to he here</Paragraph>
      <Paragraph position="13"> In the English, the subject of happen is John, but in the Welsh, the subject of digwydd ('happen') is the situation of John's being here.</Paragraph>
      <Paragraph position="14"> The predicate shift is introduced to define a translation equivalence where certain dependencies are differently arranged in the target language, as compared to the source language. It can be understood in terms of semantic type-shifting - - the types are differently distributed in the source- and target-language logical forms. Shifl(H,E) means that in any instance of subj(H,X) or l'(X, tt) (where V stands tbr auy rela~ tion) in the logical form representing the English text, the H corresponds to an E in the logical form representing the other language. The following example shows how the example under (it) above would work in practice.</Paragraph>
      <Paragraph position="15"> lch welt\] Hans schwimmt gern - i know John likes</Paragraph>
      <Paragraph position="17"> The columns of the table show tile German source logical form, the English logical form as it would be without shift, and tile actual English target logical form. The horizontal divisions of the table represent the fields of operation of separate transfer rules.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.2 Interaction between types of com-
</SectionTitle>
      <Paragraph position="0"> plex transfer When several of these syntactic changes are combined in one example, the logical form resulting from transfer is incoherent. The following examples are types on which most, if not all, existing machine translation Acq~s DE COLING-92, NAN-I~S, 23-28 Aour 1992 6 8 I PROC. OF COLING-92, NANTES, AUG. 23-28. 1992 systems will fail, as will the mechanism proposed here so far.</Paragraph>
      <Paragraph position="1"> English - Welsh; raising with argument-switching: John happens to like swimming fe ddigwydd fod yn dda gan John nofio happens be nice by John swim  John happens to like swimming (also 'John likes happening to swim~ There are two separate causes of difficulty in these examples. Firstly, the long-distance movement caused by raising canses problems in the transfer of structures which have been the target of raising. The simplest way to get rmmd the problem is to require syntactically-deleted arguments to be marked explicitly on all the verbs of wbich they are logical arguments, treating the phenomenon syntactically as an example of lexieally-triggered extraction, on a par with the extractions of syntactic objects in a man I knowand an easy man to please. Transfer rules which introduce raising predicates will then have to bring in the new subject explicitly. For instance, the rule for happen and digwydd, (iv) in SS4.1, will be re-written as</Paragraph>
      <Paragraph position="3"> The second point is that the shift predicate must be defined in such a way that it can cope with recursive changes in the dependency structure brought about by the occurrence of several interdependent head-switching translations. It seems that shift can be made to do this simply by having e.g. shifl(H,E) affect all instances of snbj(tl, X) or P(X,H) (including shifl(X,H)) not transferred by the transfer rule which introduced the shift(ll, E).</Paragraph>
      <Paragraph position="4"> Together, these two stipulations enable the transfer of examples involving head-switching, argumentswitching, raising, and most types of decomposition.</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5 Transfer of Coordinate
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Structures
5.1 Problems in transfer of coordi-
</SectionTitle>
      <Paragraph position="0"> nate structures Coordination often reqnires exceptional treatment in parsing and generation of natural language. Transfer is no exception. Transferring a coordinated structure sometimes produces a miserable result due to the accompanying combination of structural changes or the fact that coordinated linguistic objects require different target words for a shared source word. However, few attempts have been reported to formalize this problem.</Paragraph>
      <Paragraph position="1"> We here divide the problem into two categories: t  The first type of problem occurs when semantic features of the coordinated source words require a source word to be translated into two different target words. A typical example can be seen in translation of the following English sentence into Japanese.</Paragraph>
      <Paragraph position="2"> (la) Site wears a bat and shoes.</Paragraph>
      <Paragraph position="3"> (lb) kanojo-ga boushi-wo kaburi, kutsu-wo she-suhj lmt-ohj wear shoe-obj haku.</Paragraph>
      <Paragraph position="4"> wear As is understood from its translation, &amp;quot;wear&amp;quot; is translated &amp;quot;kaburu&amp;quot; or &amp;quot;haku&amp;quot; in Japanese, depending on whether its object is something worn on the head or on the foot(or leg). This means that, in this example, coordination of objects in English should be altered to that of verb phrases in Japanese.</Paragraph>
      <Paragraph position="5"> This type of knowledge for lexical choice is very common in a transfer or bi-lingual dictionary, and plays an essential role in lexical transfer of most current transfer-based MT systems. The problem is that ueither a transfer program or a transfer-rule writer expects such an awkward problem caused by coordination. To translate &amp;quot;wear&amp;quot; iuto &amp;quot;kaburu&amp;quot; in the above example, n rule writer may usually write the following rule in our notationS: wear(X) ~ kaburu(X) iff obj(X,Y)&amp;HAT(Y) But the condition part of this rule is implicitly expected to be interpreted as follows.</Paragraph>
      <Paragraph position="6"> wear(X) ~ kaburn(X) iff V Y obj(X,Y)&amp;;HAT(Y) The precise definition may change depending on how the coordinate structure is represented. But the 1 There is eamther type of problem wlfich is baaed on the syntactic differences between coordination constructlom in source arid target languages. For example, &amp;quot;2 or 3 potuld~&amp;quot; in English should be trmmlated &amp;quot;2 pondo ka 3 pondo'(2 poundJ or 3 pounds} in JaF.mlese and &amp;quot;dwybunt ncn clair&amp;quot; (2 pounds or 3) in Welsh. (The Welsh expression is used only for a sum of money. Another exists for weight.) This type of problem tlaould alto be solved in transfer, but we do not mention it here. 2In this section, we coltsider translation whose source and target logical forms are on the left and right sides of a trantfer rule. For the sake of simplicity, trmasfer rules hereafter are described as uni-directional ones.</Paragraph>
      <Paragraph position="7"> AC'TES DE COLING-92. NANqT~. 23-28 Ao~r 1992 6 8 2 PROC. OF COLING-92, NANTES. AUG. 23-28. 1992 point is that &amp;quot;wear&amp;quot; may be trmmlated &amp;quot;kaburn&amp;quot; only if all the objects have a feature of being a &amp;quot;HAT&amp;quot;. A simple transfer algoritlnn, for example, may choose the target word when it finds the first applicable transfer rule for a source word: this algorithm may produce &amp;quot;boushi-to kutsu-wo kaburu&amp;quot; for the sentence (la), which memm that tile semantic relation between &amp;quot;wear&amp;quot; and &amp;quot;shoes&amp;quot; is ignored. There may be another type of transfer algorithm which simply fails beeanse it cannot provide two dilfcrent target words to one identical source word.</Paragraph>
      <Paragraph position="8"> The second type of the problem occurs when one of the coordinated objects triggers a complex transfer which is described in SS3. This type of problem can bc seen in the following translation.</Paragraph>
      <Paragraph position="9"> lie connnitted nmrder then suicide.</Paragraph>
      <Paragraph position="10"> kare-ga satsujin-wo okaslfi, jisatsu-shita.</Paragraph>
      <Paragraph position="11"> he~subj murder-obj conmlit commit-ed suicide This problem is more complicated than the previous one because complex ti'ansfer, in this example &amp;quot;manyto-one transfer&amp;quot;, causes a structural change. Our simple transfer algorithm mentione&lt;l in tile previous section may produce a disastrous result for this type of translation.</Paragraph>
      <Paragraph position="12"> There are several possible solntions to this problem.</Paragraph>
      <Paragraph position="13"> The simple.st one is that a transfer:rule writer writes all the transfer rules which explicitly describe every possible sub-structure with coordination. This is of course unrealistic.</Paragraph>
      <Paragraph position="14"> Another solution is to make a transfer program which modifies the transfer result dyuanfically when, tbr example, a source word is given two differeut tar: get words. But such a dynamic modilication of tile result during transfer is against our policy of logic: based transfer, because this means gradual transformation of the source strueturc and therefore transfer cannot be formalized as logical inference.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
5.2 'lYansfer with Coordinate Expan-
</SectionTitle>
      <Paragraph position="0"> sioll  tiereafter we concentrate on a case where coordinated objects of a verb cause the problem, though there is apparently an example where other eases such a.~ &amp;quot;snbj&amp;quot; cause the same problem. The basic reqnirement in logic-based transfer is that coordination of objects should he reorganized to that of verb phrases or sentences, which is not supposed to cause problems in transfer. We call this reorganization &amp;quot;coordinate expansion&amp;quot;. The following is a part of logical form for (ta), which involves a coordinate structure.</Paragraph>
      <Paragraph position="1">  In this form, o is a linguistic object, and a predicate coord represents a relation between the linguistic object and its coustituents. The following is a r~ult of expansion.</Paragraph>
      <Paragraph position="2"> coord(w,w 1)&amp;we;tr(w 1 )&amp;obj(w l,n l)$zhat(ol)&amp; coord(w,w2)&amp;wear(w2)&amp;obj(w2,o2)&amp;shoe(o2) The most naive and simplest strategy usiug this expansion is to expand cvery coordination within a sentence and rel)reseut it in sentcnce coordination before transfer. This transfer result will be reorganized again into an appropriate representation of coordination in tim target language. But this solution seems inefficient from the computatioual point of view beo ettase caseswhere expansion is uecessary are rare . Unnecessary expansion and reorganization of coordinate structures should be avoided.</Paragraph>
      <Paragraph position="3"> The strategy wc propose executes coordination expansion only if it is necessary\[5\]. Figure 2 shows a general view of our modified approach to transfer. Transfer is divided into three phases; in the first phase, logical forms are expanded if expand rnles (explained below) find tim necessity of coordinate expansion. This process continues as long as the necessity remains. In tile second phase, transfer described in previous sections is executed. Fiually, in the third phase, coordilmtion is reorganized if the target language has a more appropriate structure for coordination than tile second phase result. (Consider the translation of (lb). Without reorganizing eoordina~ ti(m, the transfer result will contain two &amp;quot;wear&amp;quot;s .) Tile following is all expand rule which detects the necessity of expansion concerned with translation of &amp;quot;wearing a bat&amp;quot;.</Paragraph>
      <Paragraph position="5"> ht tile rule, e~:pand(Y,X) means that coordination of the level Y should be expanded to that of X. This rule suggests that coordinate expansion is necessary AClES DE COLING-92, NAI, CrEs, 23-28 AoI~r 1992 6 8 3 PROC. OF COLING-92, NANTES, AtnL 23-28, 1992 if the object of &amp;quot;wear&amp;quot; is a coordinate structure, and one constituent is a HAT while another isn't.</Paragraph>
      <Paragraph position="6"> We assume that expand rules are produced automatically from transfer rules before the actual transfer operation. The simplest way of extracting such a rule refers only one transfer rule at one time. This means that the necessity of expansion can be detected not by creating expand rules but by referring transfer rules in actual transfer phase. But the former approach seems essential if we wish to optimize expansion detection.</Paragraph>
      <Paragraph position="7">  Our coordinate expansion detection works even if a coordinate structure has more than two constituents.</Paragraph>
      <Paragraph position="8"> What wc have to consider is an appropriate expansion algorithm. For example, in translating (4a), an appropriate expansion should keep coordination of &amp;quot;shoes&amp;quot; and &amp;quot;stockings&amp;quot;, as shown in (4b), hecause both satisfy a semantic constraint on which the system chooses &amp;quot;haku&amp;quot; as the translation of &amp;quot;wear&amp;quot;. Otherwise reorganizing a coordination in the generation phase is inevitable.</Paragraph>
      <Paragraph position="9">  Expand rules from a transfer rule which involves a structural change are little different in forms to the previous ease. The following are a transfer rule for translating &amp;quot;commit suicide&amp;quot; into Japanese and its expand rule.</Paragraph>
      <Paragraph position="11"> Another example is the translation of the English &amp;quot;HUMAN have ADJ NOUN&amp;quot; construction into Japanese. The sentence (2E) is translated (2J) using the rule (3).</Paragraph>
      <Paragraph position="12">  Reorganization of a coordination in the target language does not only occur as outlined in the above case. Since the coordinate expansion is completely separate from the actual transfer process, transfer rules which do not cause problems might be used. There is still a problem to be solved with regard to tile transfer of coordination with &amp;quot;expansion&amp;quot;; expansion is not always straightforward. There is often a ease where coordinate expansion is impossible without deep understanding of the sentences, or it is impossible in that it may change their semantic structures. For example, the sentence (5b) cannot be the expansion of (5a) though it seems so at first glance. (5a) I saw a group of men and women.</Paragraph>
      <Paragraph position="13"> (5b) I saw a group of men and I saw a group of women. The apparent disadvantage of our approach with &amp;quot;expand rules&amp;quot; is that a large number of expansion rules might be created. Though it provides an efficient way of detecting the necessity of expansion, it consumes a lot of memory, which will raise an implementation problem.</Paragraph>
      <Paragraph position="14"> This case is more complicated than the previous ones because the transfer rule refers to two semantic features. Therefore we will get two expand rules, one of which is the following rule.</Paragraph>
      <Paragraph position="16"> In addition, we need another expand rule which checks another type of coordinate construction. This rule will apply when the system translates &amp;quot;she has long hair and beautiful eyes&amp;quot;.</Paragraph>
      <Paragraph position="18"/>
    </Section>
  </Section>
class="xml-element"></Paper>