<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1021">
  <Title>the The Transfer Phase of Mu Machine Trans l at i on Nakoto NAOAO</Title>
  <Section position="2" start_page="0" end_page="97" type="metho">
    <SectionTitle>
2 The Construction of the Transfer Phase
</SectionTitle>
    <Paragraph position="0"> The transfer phase of our system consists of the following three sub-phases (Fig. i).</Paragraph>
    <Paragraph position="1">  Analysis results 'of input sentences of the source language (SL) are represented in the form of annotated tree structures, from which one can retrieve various levels of information contained in the input sentences. The MTP is performed recursively from the top of the tree to the bottom (Fig. 2). At each step of the recursion, the sub-structure governed by a node (current node, node X ira Fig. 2) is transferr(~d to the corresponding structure of the target language (TL). This phase mainly performs lexJcal tran~'{fer, that is, it chooses an appropriate target lexical item for node X by examining the annotation parts of that node and its dependent nodes (nodes YI,..Y2,..Yn). The selection of a target lexical item may place certain constrains on the transfer of the dependent nodes. These constrains are expressed as property-value pairs and added to the annotation parts of the dependent nodes, which are to be utilized in the subsequent recursion steps. For language pairs such as Japanese and English which belong to quite different language families, however, the lexical transfer is not so straightforward. It often happens that single lexical items of SL correspond to complex (-~xpressions of TL and vice versa. Furthermore, certain structural changes are also required.</Paragraph>
    <Paragraph position="2"> Because the MTP recursively transfers sub-structures governed by single nodes, certain global structural differences cannot be naturally treated. Such global changes are deal t with by the Pre-TP and Pest-TP sub-phases (See section 7) .</Paragraph>
  </Section>
  <Section position="3" start_page="97" end_page="98" type="metho">
    <SectionTitle>
3 The Form-Oriented Dictionatz and Lexical Transfer
</SectionTitle>
    <Paragraph position="0"> of Nouns The following four columns are provided in the form-oriented bilingual dictionary of nouns (Fig. 3). The description of each column is expanded into corresponding types of lexical rule (i.e. GRADE rules which are stored in the dictionaries and invoked at specified processing phases : GRADE is the programming language for writing grammar rules used in our project).</Paragraph>
    <Paragraph position="1"> (1)Column A : Transfer based on the Relationship to the Governor (Fig. 4) Even major parts-of-speech such as nouns need not correspond between two languages. Certain Japanese nouns correspond to words of different parts-of-speech in English, depending on the surrounding context.</Paragraph>
    <Paragraph position="3"> However, the same nouns should be translated as nouns in Engl~sh in contexts such as :</Paragraph>
    <Paragraph position="5"> Surface Case Relation Surface Marker Equivalent \[ Speech~ RATerial null wooden I ADJ __\] This indicates that: if '~' is used as a noun a~ifier and the case relation between '9};' and the governor is 'RATerial' then the translation equivalent is 'wooden', else 'tree'.  (2) Column B : Selections of Prepositions(Fig. 5)  Certain nouns determines the prepositions which relate their dependents of certain deep cases to them, as predicates (verbs) usually do.</Paragraph>
    <Paragraph position="6">  The semantic relationships are not sufficient for selecting such preposJ tions appropriately. Same postpositional case particles should be translated into different prepositions, even though the deep semantic relationships in Japanese are the same (thai: is, Japanese native speakers cannot intuitively distinguish the difference.) In Ex-4, for example, English native speakers can clearly distinguish the semantic relationship of 'results' and 'effects ~ (GOAL or OBJECT case) from that of 'family' and 'duty' (BENEFICIARY case), whereas in Japanese the same postpositional case particles are used and so the distinction of the semantic relationships is not so clear for Japanese natives. We cannot, therefore, even expect the stability of semantic relationships across the two languages.</Paragraph>
    <Paragraph position="7">  (3) Column D : Transfer based on Collocations with Predicates (Fig. 6)  Certain combinations of nouns and predicates (verbs, adjectives, etc.) in Japanese are translated into single words in English : the combinations are lexicalized by single English words, combinations of nouns and adjectives, for example, are often \].exicalized by single English adjectives.</Paragraph>
    <Paragraph position="9"> These combinations can appear in the position of an embedded clause and cause structural changes between two languages(Fig.7). Combinations of verbs and their case elements may also correspond to single English verbs.</Paragraph>
    <Paragraph position="11"> It also happens that Japanese verbs with wide ranges of meaning should be translated into certain English verbs when collocated with specific nouns. Rules which choose appropriate English verbs by referring to semantic class:ifications of case elements (nouns) are defJ ned as \].exical rules in the verb dictionary (See Section 4). However, certain collocations of verbs and nouns are so specific that choices of appropriate target verbs cannot be expressed by the lexica\] rules of verbs of this type. column D in the form-oriented dictionary of nouns is used to specify such collocations between nouns and verbs.</Paragraph>
    <Paragraph position="12"> Ex-7. Ih~ll~ %1 ~b W ~ ....... &gt; leave a space (space) (??) The description of this column is not only expanded into \]exica\] rules of the nouns but also into the property (J-DIC-DER:V) of those nouns. The values of this property is a set of Japanese predicates which require special treatments when they are collocated with the nouns. When node X ill Fig. 2 is a predicate, the rules in the main grammar check whether there exists a dependent node whose property J-DIC-D\]'\]RIV contains the predicate, and i f such a node exists, they invokes the lexical rules of the corresponding nouns.</Paragraph>
    <Paragraph position="13"> In the dependency structures produced by the Japanese analysis grammar, simple modification of nouns by adjectives such as Ex-8. ~ i~J ...... &gt; heavy rain (to be strong, (rain) to be big, to be heavy, etc.) are also represented as embedded clauses (Fig. 8), and so column D is used to choose adjectives in English. Note that most of the combinations between nouns and appropriate adjectives are highly conventionalized in each language and so it is hard to imagine the conceptual primitives for adjectives that could be language independent.</Paragraph>
    <Paragraph position="14">  J$C: Japanese surface case marker JDC: Japanese deep cam ESC: English surface case marker EDC: English deep case BOD: Additional specification ('C/dei' shows the case element should be deleted in English, i.e. the meaning of the case element is lexically included by English predicate.) This shows :If '~'(efficiency) is collocated mith ',~#~t~'(high), then the combination should be translated into 'efficient', else the translation equivalent is 'efficiency'.</Paragraph>
    <Paragraph position="16"> (the predicate J ~the translation equivalee of the embedded clause ~f '~,b~' whoa it modifies</Paragraph>
    <Paragraph position="18"> (4) Column C : Transfer based on Collocation with</Paragraph>
    <Section position="1" start_page="98" end_page="98" type="sub_section">
      <SectionTitle>
Modifiers
</SectionTitle>
      <Paragraph position="0"> We have several words of minor parts-of-speech in Japanese which roughly correspond to determiners, quantifiers, etc. Jn English. Almost the same kind of information as column D is specified in this column for these words.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="98" end_page="100" type="metho">
    <SectionTitle>
4 The Form-Oriented Dictionary and Lexical Transfer
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="98" end_page="100" type="sub_section">
      <SectionTitle>
of Verbs
</SectionTitle>
      <Paragraph position="0"> When the node X in Fig. 2 is a verb or an adjective, the transfer grammar for predicates is invoked. If it dose not exist:, the other rule in the grammar invokes the lexical rules derived from the form-oriented dictionary of the predicate (node X). By examining the annotation parts of that node and its dependents, the lexJcal rules of a predicate may determine (\]) the trans\].ation equivalent in English (2) the deep case relationship and surface ease maker of those dependents strictly governed by the predicate.</Paragraph>
      <Paragraph position="1"> If a Japanese predicate is really ambiguous, a separate case frame is provided for each of its meanings. The Japanese analysis grammar disambiguates the meaning and a separate lexical entry is prepared in the bilingual dictionary, i.e. different meanings of a predicate are treated as different lexica\], items  in the transfer phase. However, the real problems in translation is not caused by ~mbiguity such as this but by vagueness of meaning. When we treat a language pair such as Japanese and English, it is not exceptional but rather usual that a verb of SL has to be translated into different target lexical items, even though the native speakers of SL cannot clearly realize the meaning difference.</Paragraph>
      <Paragraph position="3"> The above three usages of the verb '~ ' are continuous in the sense that all of them have the core meaning in common. If we considered them as different meanings, the Japanese verb '~' would have plenty of different meanings and be a highly ambiguous word. Note that the English verb 'to wear' should be translated into more than five differnt Japanese verbs depending on what is to be worn. 'To wear shoes', 'to wear a watch', 'to wear spectacles', etc. should all be translated differently. These facts show that it is impossible to establish the set of lexical items of the interlingua. In our system, then, these 'translationally ambiguous' lexical items are treated in the transfer phase.</Paragraph>
      <Paragraph position="4"> The form-oriented dictionary of verbs is shown in  an Example of Descriptions Because certain Japanese verbs are translated into complex expressions in English, one can specify not only the English main verbs but also arbitrary phrases governed by the verbs as constants.</Paragraph>
      <Paragraph position="6"> ..... &gt; to produce trans\[X\] on a trial base Note that X in the above example indicates a variable an arbitary Japanese noun phrase can fill in. T RANS(X) shows the result of transferring the noun phrase into English.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="100" end_page="100" type="metho">
    <SectionTitle>
5 Recursive Processing and Multiple Layer of Grammars
</SectionTitle>
    <Paragraph position="0"> Fig.10 shows that the schematic construction of the MTP. The substructure governed by the node X is transferred recursively by an appropriate subgrammar, depending on the properties of that node. For simplicity, only the subgrammars for main clauses and noun phrases are shown in detail. The blocks surrounded by bold lines in this figure are those which are performed by lexical rules described in section 3 and 4. The lexical rules are invoked at the beginning of the subgrammars, because they are the rules for treating idiosyncratic phenomena specific to lexical items and are applied before the other, moregeneral rules are applied. Constraints on some of the dependent nodes such as syntactic forms (a prepositional phrase with a specific preposition,etc.) are determined by these lexical rules, and added to their annotation parts. These constraints are utilized at the next recursion step where the substructure under each dependent node is  treated by the \].exieal rules are dealt with by general rules. That is, general rules are applied to the dependent nodes of a predicate, only when the  node Wi in Fig. i0 is not treated by (i) lexical ru\].es from the column D of the node Wj (2) lexical rules of the predicate, the node X (3) lexica\] rules from the column A of the node Wi (4) lexical rules from the column C of the node Wi  The general rules are further divided into the following two categories : (i) Rules in the case when the deep case relation between the node Wi and the governor (the node X) is uniquely determined by the analysis grammar : They determine the surface preposition of English by referring to the deep case interpretation, the surface case marker of Japanese, and the lexica\] item and the semantic markers attached to the node W. 1 (55 semantic markers are used in our system to classify Japanese nouns) .</Paragraph>
    <Paragraph position="1"> (2) Rules in the case when the deep case relation is not uniquely determined in the analysis phase : They determine the preposition just by referring to the surface case marker of Japanese. The preposition roughly corresponding t:O the Japanese surface case marker is selected.</Paragraph>
    <Paragraph position="2"> The above two categories of rules are just rough c\]assificat\[ons and, in reality, more detailed rules are prepared for each Japanese case marker. In particular, there are several type 2 rules for each surface case marker which take into account the set of possible case interpretations given by the analysis phase.</Paragraph>
    <Paragraph position="3"> Thus, tl~e dependent nodes of a predicate are transferred by several kinds of rules with differing preferences in order to utilize as much of the in \[ormatJon extracted duri ng the ansi ys is phase as possible. The same multiple layer construction of subgrammars is a\].so adopted for the other grammatical units such as subordinate clauses, embedded clauses, etc.</Paragraph>
    <Paragraph position="4"> 6 Lexica\] Rules and the Form-Oriented Dictionaries ............................. The form-oriented dictionaries are provided just for preparing dictionary descriptions for a large number of ordinary lexica\] J terns. The forms given in the preceding sections are used by lexicographers who do not have any detail ed knowledge about the transfer grammar. At execution time, these descriptions are expanded into GRADE lexica\] rules (Fig. ll).</Paragraph>
    <Paragraph position="5"> If one wants to specify complicated transfer rules which cannot be expressed in these forms, one can provide \]exica\] rules directly i n GRADE, and attach them to specific lexica\], items. Such lexica\] rules are also invoked at the appropriate time during tha transfer phase. The rules in the transfer grammar checks whether the \]exical items have \]exJca\] rules of certain types, and if they have, the \]exical rules are i nvoked whether they are derived from the form-oriented dictionaries or prepared directly in the GRADE form. On&lt;! can specify arbitrary treetransformations in GRADE so that one can prepare very flexible transfer rules specific to \]exica\] items. Lexical rules for certain items such as Japanese verbs with wide ranges of usages are given directly in GRADE.</Paragraph>
    <Paragraph position="6">  Fig.ll Relationships betveen Format Dictionaries and Dictionaries for Execution Though mosl Japanese verbs and nouns are respectively translated into English verbs and nouns, certain parts-of-speech in Japanese do not have such direct equivalents in English, and the word of such parts-of-speech need to be translated into various kinds of expressions. The transfer rules are highly specific to each lexica\] item, and therefore, we do not prepare any dictionary formats for these words.</Paragraph>
    <Paragraph position="7"> The tranfer rules are directly prepared in the GRADE form. The adverbial postpositional particles in Japarlese, for examp\].e, are translated into adverbs, prepositions, sty\].es of sentences, etc., depending on each \]exica\] item and its surrounding context.</Paragraph>
    <Paragraph position="8">  .... &gt; 1) such as A, B and C ---&gt; A, B, C and so on NoLe a\]so that, because the roles of these particles are to give additional, meaning to the core meaning (i.e. the propositional contents) of the sentences and some of them are related to the attitude or the intention of the speaker, the inter\] ingua\] representation for these patti cles can be very comp\]ieated. \]it is a\] so the case for the Japanese auxi.\]iary verbs which follow the main predicates and express vari ous kinds of modal i ty, aspect, politeness, and so on. The correspondence of these expressions is highly dependent on language pairs and so should be treated in the transfer phase.</Paragraph>
  </Section>
  <Section position="6" start_page="100" end_page="101" type="metho">
    <SectionTitle>
7 The Structural Transfer
</SectionTitle>
    <Paragraph position="0"> The MTP transfers Japanese structures into corresponding English structures recursive\]y by traversing the Sapanese analysis trees (dependerlcy trees) from the top te tbe bottom. The order of processing is })re-determined by the dependency trees given by the analysis phase. If the substructures are changed unexpectedly durLng the recursion, the upper \]eve\] of recursi ve processing cannot respond to the change ; if it could, tee rules which cbeck the results of the transfer of the substructures and take appropriate actions would need ko be provided, making the construction of subgrammars unnecessarily complicated.</Paragraph>
    <Paragraph position="1">  There are pairs of stereotypibal sentence sty\] es J n the two languages which cannot be translated by phrase-to-phrase trans\]ation.</Paragraph>
    <Paragraph position="2"> translated irlto an adverb or a preposJ tiona\] phrase in EnglJ sh. Rules are prepared in Pre-TP, which transform clausal constructions into phrasal constructions.</Paragraph>
    <Paragraph position="3"> Ex-11AtZW$~a~12t~&lt;, B_~....8~6 ---&gt; not only h is ~ but also B is &amp;quot;o&amp;quot; Because the clues of such stereotypical sentence patterns are embedded deeply in the analysis trees (see Fig.14) and resultant tree structures of the transfer phase become quite different from those of the analysis phase, it is difficult for the recursive processing in MTP to treat them.</Paragraph>
    <Paragraph position="4"> The two phases, Pre-TP and Post-TP, are provided for such global structural changes. In these two phases, each rule traverses the whole tree independently in order to find where the rule is applicable.</Paragraph>
  </Section>
  <Section position="7" start_page="101" end_page="101" type="metho">
    <SectionTitle>
8 Pre-Transfer Phase
</SectionTitle>
    <Paragraph position="0"> In this phase, certain structures in Japanese which do not have direct Engli sh equivalents are transformed into more English-oriented structures.</Paragraph>
    <Paragraph position="1"> For example, the following structural transformation (Fig. 12) is taken place, because Japanese quantifiers such as ~b ~ (many, much), ~f~b~ (few, little) etc. can be used as predicates but English quantifiers cannot.</Paragraph>
    <Paragraph position="2">  Furthermore, the following structures in Japanese are treated.</Paragraph>
    <Paragraph position="3"> (i) Subordinate clauses which should be expressed by adverbs or prepositional phrases in English : Ex-13 (to pay attention to) (machine) (to control) ............ &gt; to control (the) machine carefully A subordinate clause with a predicate of 'TE'-form (~,\[~7\[ - to pay attention to) is usually expressed in Engllsh as a subordinate clause with a conjunction such as 'while', 'when' etc. or a participle clause. However, when the predicate of the subordinate clause is a verb of a certain type, the clause expresses the 'manner' case of the main clause and it should be Such structural transformations are often required in the translation of Japanese into English. Some complex expressions which contain subordinate clauses or embedded clauses are used in Japanese only for expressing deep case relationships.</Paragraph>
    <Paragraph position="4"> (A with a subject marker)(cause) (to become) ..... &gt; (d irect translat ion) Because 'A' becomes the cause .......</Paragraph>
    <Paragraph position="5"> (deep interpretation) 'h' fills the deep case 'cause' of the main clause --&gt; 'by' Because direct translations of these expressions result in clumsy English, they are reduced into much simpler structures in the transfer phase (Fig. 13). main clause main clause</Paragraph>
    <Paragraph position="7"/>
  </Section>
  <Section position="8" start_page="101" end_page="102" type="metho">
    <SectionTitle>
SUBJ GOAL
</SectionTitle>
    <Paragraph position="0"> Fig. 13. Reduction of a Structure in Pre-1? (2) Stereotypical expressions of Japanese which have corresponding English stereotypical expressions : The correspondence of stereotypical expressions in the two languages such as given in Ex-ll cannot be recovered appropriately by the phrase-to-phrase transfer in MTP. Rules are prepared in Pre-TP, which discover the existence of such stereotypical expressions and insert special nodes in the dependency trees to facilitate the recursive processing of MTP (Fig. 14).</Paragraph>
    <Paragraph position="1">  (3) Complf~-x Expressions following predacates :  Though auxi liary verbs in Japanese fo\] \] ow main predicates to express various kinds of information such as modality, aspect, po\]iteness, etc., some complex expressions which contain embedded clauses are also used to express the same kind of information and should be expressed in English by auxiliary verbs, aspect verbs (ex : to begin, to continue, keep --ing, etc.), etc. (Fig. 15).</Paragraph>
    <Paragraph position="2">  Note that what kinds of information can be expressed hy simp\].e &amp;uxiliary verbs and so by simple linguistic constructions is dependent on each language, and the adjustment of the di. fference requires certain structural changes.</Paragraph>
    <Paragraph position="3"> (data (advance) (case (automated (to increase processing) particle ships) -'past tense) -'by') .... &gt; (direct translation) Due to advances of data processing, the number of autonlated ships increased.</Paragraph>
    <Paragraph position="4"> (natura I transl at ion) Advances of data processing increase(\] the number of automated ships.</Paragraph>
  </Section>
class="xml-element"></Paper>