<?xml version="1.0" standalone="yes"?>
<Paper uid="C96-1054">
  <Title>Semantic-based Transfer*</Title>
  <Section position="3" start_page="316" end_page="316" type="metho">
    <SectionTitle>
2 Semantic Representations
</SectionTitle>
    <Paragraph position="0"> The different Verbmobil semanti(: construction components use variants of UDRS as their semas&gt; tic formalisms, el. (Bos et al., 1996; Egg and Lebeth, 1995; Copestake et al., 1995). The ability to underspecify quantifier and operator scope together with certain lexical ambiguities is impof rant for a practical machine translation system like Verbmobil because it supports ambiguity preserving translations. The (lisambiguation of different readings couht require an m'bitrau'y amount of reasoning on real-world knowledge and thus should be avoided whenever possible.</Paragraph>
    <Paragraph position="1"> In the following examples we assume an explicit event-based semantics (Dowry, 1989; Parsons, 1991) with a Neo-Davidsonian representation of semantic argument relations. All semantic entities in UDRS are uniquely labeled. A label is a pointer to a semantic predicate nmking it easy t,o refer to. The labeling of all scntantic entities allows a fiat ret)resentation of the hierarchical structure of arguinent an(1 ot)erator and quantifier scope embeddings as a set of labeled conditions.</Paragraph>
    <Paragraph position="2"> The recursive embedding is expressed via additional subordination constraints on labels which occur as arguments of su(:h olmrators.</Paragraph>
    <Paragraph position="3"> Example (la) shows one of the classical Verbmobil examples and its possible English  translation (lb).</Paragraph>
    <Paragraph position="4"> (1) a. Das paflt echt scMecht bei 'mir.</Paragraph>
    <Paragraph position="5"> b. That really doesn't suit me well.</Paragraph>
    <Paragraph position="6">  The corresponding semat)l,ic representations are given in (2a) and (2b), respectively. 2</Paragraph>
    <Paragraph position="8"> Scnlantic entities in (2) are reprc, scnted as a Prolog list of laImlcd conditions. After the unificationt)ased s(~illantic construction, the logical wu'iables for labels and nmrkers, such as events, states and individuals, are skolemize(l with special constant symt)ols, e.g. 11 for a label att(t +-1 for a state. Every condition is prefixed with a lal)el serving as a unique identifer. Labels are also useflll for grouping sets of conditions, e.g. for i)artitions whidt belong to the restriction of a qmmtifier or which are part of a specific sub-DRS. Additionally, all these special constants can be seen as pointers for adding o1&amp;quot; linking informal, ion within ml(t between multiple levels of the VIT.</Paragraph>
    <Paragraph position="9"> Only the set of semantic conditions is shown in (2); the ot;her levels of the multi-dimensional VIT ret)resentation , whidt contain additional senmntic, 2For l)resentaJ;ion purposes we have simplilied the actmd VIT representations.</Paragraph>
    <Paragraph position="10"> pragmatic, morpho-syntactic mid prosodic information, have been left, out here. If necessary, such additional information can be used in transfer and sem;mtic evaluation for resolving ambiguities or in generation for guiding tile realization choices. Furthermore, it .allows traasfer to make fine-grained distinctions between alternatives in cases where the semantic representations of source mid target lmlguage do not match up exactly.</Paragraph>
    <Paragraph position="11"> Semantic operators like negation, modals or intensitier adverbials, such as really, take extra label arguments fi)r referring to other elements in the flat; list which m'e in the relative scope of these operators.a This form of semantic representation has the following adwmtages for transfer: * It; is possible to preserve the underspecification of quantifier amd operator scope if there is no divergence regarding scope ambiguity  between Sollrce and target languages.</Paragraph>
    <Paragraph position="12"> * Coindexation of labels and markers in the  source and target parts of transfer rules ensures that the semantic entities are correctly related and hence ()bey any semantic constraints which may be linked to dram.</Paragraph>
    <Paragraph position="13"> To produce an adequate target utterance additional constraints whirl1 arc important for generation, e.g. sortal, topic/focus constraints etc., may be preserw;d.</Paragraph>
    <Paragraph position="14"> * There need not be a 1 : 1 relation between semantic entities and individual lexical items.</Paragraph>
    <Paragraph position="15"> Instead, lexical units may be decomposed int;o a set of seinantic entities, e.g. in the case of deriwztions arid fi)r a nmre line grained lexical semantics. Lexical decomposition allows us to express generalizations attd to apply transDr rules to parts of the decomposition.</Paragraph>
  </Section>
  <Section position="4" start_page="316" end_page="318" type="metho">
    <SectionTitle>
3 Our Transfer Approach
</SectionTitle>
    <Paragraph position="0"> Transfer equivalences are stated as relations between sets of source language (SL) and sets of target language (TL) senlantie entities. They are usually based on individual lexical items but might also involve partial phrases for treating idioms and other collocations, e.g. verb-noun collocations (see example (8) below). After skolemization of the sen, antic representation the input to transfer is variable free. This allows the use of logical variables for labels and markers in transfer rules to express coindexation constraints between individual entities such as predicates, operators, quantifiers and aFor tim concrete exaanple at hand, the relative scope ha.s been fully resolved by using the explicit labds of other conditions. If the scope were underspecifled, explicit subordination constraints would be used in a speciM scope slot of the VIT. The exact details of subordination are beyond tim scope of this paper, of, Prank aa~d Reylc (1995) and Bos et al. (1996) for implementations.</Paragraph>
    <Paragraph position="1">  (abstract) thematic roles. Hence the skolemization prevents unwanted unification of labels and markers while matching individual transfer rules against the semantic representation.</Paragraph>
    <Paragraph position="2"> The general form of a transfer rule is given by SLSem, SLConds Tau0p TLSem, TLConds.</Paragraph>
    <Paragraph position="3"> where SLSem and TLSem are sets of semantic entities. Tau0p is an operator indicating the intended application direction (one of &lt;-&gt;,-&gt;, &lt;-). SLeonds and TLConds are optional sets of SL and TL conditions, respectively. All sets are written as Prolog lists and optional conditions caa be omitted.</Paragraph>
    <Paragraph position="4"> On the source language, the main difference between the SLSem and conditions is that the former is matched against the input and replaced by the TLSem, whereas conditions act as filters on the applicability of individual transfer rules without modifying the input representation. Hence conditions may be viewed as general inferences which yield either true or false depending on the context.</Paragraph>
    <Paragraph position="5"> The context might either be the local context as defined by the current VIT or the global context defined via the domain and dialog model. Those inferences might involve arbitrarily complex inferences like anaphora resolution or the determination of the current dialog act. In an interactive system one could even imagine that conditions are posed as yes/no-questions to tile user to act as a negotiator (Kay et al, 1994) for choosing the most plausible translation.</Paragraph>
    <Paragraph position="6"> K the translation rules in (3) are applied to the semantic input in (2a) they yield the semantic output in (2b). We restrict the following discussion to the direction from German to English but the rules can be applied in the other direction as well.</Paragraph>
    <Paragraph position="7">  (3) a. \[L:echt(A)\] &lt;-&gt; \[L:real(h)\].</Paragraph>
    <Paragraph position="8"> b. \[L: passen (E) ,L: arg3 (E,Y) ,LI : bei (E,X)\] &lt;-&gt;</Paragraph>
    <Paragraph position="10"> d. \[L:ich(X)\] &lt;-&gt; \[L:ego(X)\].</Paragraph>
    <Paragraph position="11"> e. \[L:pron(X)\] &lt;-&gt; \[L:pron(X)\].</Paragraph>
    <Paragraph position="12">  The simple lexical transfer rule in (3a) relates the German intensifier echt with the English real 4. The:variables L and A ensure that the label and the argument of the German echt are assigned to the English predicate real, respectively.</Paragraph>
    <Paragraph position="13"> The equivalence in (3b) relates the German predicate passen with the English predicate suit. The rule not only identifies the event marker E, but unifies the instances X and Y of the relevant thematic roles. Despite the fact that the German bei-phrase is analysed as an adjunct, it is treated exactly like the argument arg3 which is syntactically subcategorized. This rule shows how structural divergences can easily be handled within this approach.</Paragraph>
    <Paragraph position="14"> 4The semantic predicate real abstracts away from the adjective/adverbial distinction.</Paragraph>
    <Paragraph position="15"> (4) \[L:passen(E), Ll:bei(E,X)\] &lt;-&gt; \[L: suit (E), L: arg2 (E,X)\].</Paragraph>
    <Paragraph position="16"> The rule in (3b) might be further abbreviated to (4) by leaving out the unmodified arg3, because it is handled by a single metarule, which passes on all semantic entities that are preserved between source and target representation. This also makes the rule for (3e) superfluous, since it uses an inter-lingua predicate for the anaphor in German and English.</Paragraph>
    <Paragraph position="17"> The rule in (3c) illustrates how an additional condition (ILl :passen(E)\]) might be used to trigger a specific translation of schleeht into not good in the context ofpassen. The standard trailslation of schlecht to bad is blocked for verbs like suit, that presuppose a positive attitude adverbial. 5 One main advantage of having such conditions is the preservation of the modularity of transfer equivalences because we do not have to specify the translation of the particular verb which only triggers the specific translation of the adverbial. Consequently, the transfer units remain small and independent of other elements, thus the interdependencies between different rules are vastly reduced. The handling of such rule interactions is known to be one of the major problems in scaling up MT systems.</Paragraph>
    <Paragraph position="18"> A variation on example (1) is given in (5).</Paragraph>
    <Paragraph position="19">  (5) a. Das paflt mir echt schleeht.</Paragraph>
    <Paragraph position="20"> b. That really doesn't suit me well.</Paragraph>
    <Paragraph position="21"> Tile translation is exactly the same, but tile German verb passen takes an indirect object mir instead of the adjunct be/-phrase in (1). The appropriate transfer rule looks like (6a) which can be reduced to (6b) because no argument switching takes place and we can use the metarule again.</Paragraph>
    <Paragraph position="22"> (6) a.\[L:passen(g) ,L:arg2(E,X) ,L:arg3(E,g)\]&lt;-&gt; \[L: suit (E), L:arg2(E,X) ,L:arg3(E,Y)\] .</Paragraph>
    <Paragraph position="23"> b. \[L:passen(E)\] &lt;-&gt; \[L:suit(E)\].</Paragraph>
    <Paragraph position="24">  In a purely monotonic system without overriding it would be possible to apply the transfer rule in (6b) to sentence (1) in addition to the rule in (4) leading to a wrong translation. Whereas in the underlying rule application scheme assumed here, the more general rule in (6b) will be blocked by the more specific rule in (4).</Paragraph>
    <Paragraph position="25"> The specificity ordering of transfer rules is primarily defined in terms of the cardinality of matching subsets and by the subsumption order on terms. In addition, it also depends on the cardinality and complexity of conditions. For the passen example at hand, the number of matching predicates in the two competing transfer rules defines the degree of specificity.</Paragraph>
    <Paragraph position="26"> ~Instead of using a specific lexical item like passen the rule should be abstracted for a whole class of verbs with similar properties by using a type definition, e.g. type (de, pos_att itude_verbs, \[gehen, passen .... \] ). For a description of type definitions see (11) below.  The following example illustrates how conditions are used to enforce selectional restrictions from the domain model. For example ~rmin in German might either be translal;ed as appointment or as date, depending on the context.</Paragraph>
    <Paragraph position="28"> The second rule (7b) is more specific, because it uses an additional condition. This rule will be tried first by calling the external domain model R)r testing whe.ther the sort assigned to X is not suhsumed by the sort, letup_point. Here, the first rule (7a) serves as a kind of default with respect to the translation of Terrain, in cases where no specific sort information on the marker X is awfilable or the condition in rule (7b) Nils.</Paragraph>
    <Paragraph position="29"> In (8), a light verb construction like einen Terminvorsehlag aachen is translated into su.qgest a date by decomposing the compound and light verb to a simplex verb and its modifying noun.</Paragraph>
    <Paragraph position="30">  (8) \[L : machen (E) , L : arg3 (g, X) , LI : terminvorschlag (X) \] &lt;-&gt; \[L : sugge st (E) , L : arg3 (E, X), LI : date (X) \] . We close this section with a support verb example (9) showing the treatment of head switching in our approa.ch. The German comparat;ive construct,ion lieber sei'n (lit.:bc more liked) in (9a) is t;ranslated by the verb prefer in (9t)).</Paragraph>
    <Paragraph position="31"> (9) a. Diensta9 ist air lieber.</Paragraph>
    <Paragraph position="32"> h. \[ would l, refer Tuesday.</Paragraph>
    <Paragraph position="33"> (IO) \[L : suppore (S, LI ), L2 : experioncer (S, X)</Paragraph>
    <Paragraph position="35"> The tra.nsfer rule in (10) matches the decoinposilion of the comI)at'at;ivt; form lieber into its positive forin lieb atnt an additional comt)arative predicate toget;her with l;he. support verb sei'n such t;IKtl; tile comparative construction lieber sein (g ist X liebeT) is translated as a whoh; to the English verb prefer (x prefers Y).</Paragraph>
  </Section>
  <Section position="5" start_page="318" end_page="319" type="metho">
    <SectionTitle>
.4 Discussion
</SectionTitle>
    <Paragraph position="0"> The main motivation for using a senmntic-based at)proach for transfer is the abilil;y to abstract aww froln morplioh)gical and syntactic idiosyncrasies of individual languages. Many of the, traditional cases of divergences discussed, e.g. by Dorr (1994), at'e already handled in the Verbmobil syntax-seniantics interface, hence they (lo not show up in our transfer at)proach. Examples include cases of categorial and thematic divergences.</Paragraph>
    <Paragraph position="1"> These are treatt;d in tile linking between syntactic arguments and their corresponding thematic roles.</Paragraph>
    <Paragraph position="2"> Another advantage of a semantic-based t;ransfer approach over a pure interlingua apt)roach, e.g. Dorr (1993), or a direct sl;ructural c()rrespondence approach, e.g. Slocum el; al. (1987), is the gain in modularity by allowing language independent grammar development. Translation equivalences relating semantic entities of the source and target grammars can be fi)mmlated in a grmnmar independent bilingual semantic lexicon. In cases where the semantic representations of source, and target language are not isomorphic, a nontrivial transfer relation between the two representations is needed. But it is cleat'ly much easier to niap between fiat semantic representations than between either syntact;ic trees or deeply nested semantic re, presentations An inl;erlingua approadt presumes thai; a sire gle representation for arbitrat'y languages exists or can be developed. We believe fi-om a grammar engineering point of view it is unrealistic to come tip with such aai interlingua representation without a strict coordination between the monolingual grammars. In general, a pure interlingua approach results in very application and domain specific knowledge sources which at'e difficult to maintain atM extend to new languages anti domains. This holds especially in the Verbmobil context with its distributed gratnmat- development.</Paragraph>
    <Paragraph position="3"> Whereas our approach does not preclude the use of interlingua predicates. We use interlingua representations for time and date expressions in the Verbmobil domain. Sinfilarly for prepositions, cf.</Paragraph>
    <Paragraph position="4"> Buschbeck-Wolf and Niibel (1995), it makes sense 1,o use inore abstract relations which express timdmnental relationships like temporal location or spatial location. Then it is left to the language specific grammars to make the right lexical choices.</Paragraph>
    <Paragraph position="5">  (11) a. type(de ,leap lea, \[an, in,um,zu\] ).</Paragraph>
    <Paragraph position="6"> b. am Dienstag, im Mai, um drei, zu Ostern c. type(en,temp loc,\[on,in,at\]).</Paragraph>
    <Paragraph position="7">  (t. on Tuesday, in May, at three, at E~Lster 'File class deiinitions in (lla) arid (llc) cluster together those prepositions which can be used t,o express a temporal location. The names de and en are the SL and TL modules in which the (:lass is deiined, temp loc is the (:lass natne and the list denotes the extension of the class. (11b) and (11d) show possible German and English lexicalizations.</Paragraph>
    <Paragraph position="8"> (12) \[temp_loc(E,X)\] , \[sort(X)=&lt;time\] &lt;-&gt; \[temp loc(E,X)\].</Paragraph>
    <Paragraph position="9"> The interlingua rule in (12) identifies the abstract teinl)oral location predicates under the condition that the internal argument is more specitlc than the sort time. This condition is necessary because of the polysemy of those prepositions. During comt)ilation the SL class definition will be automatically expanded to the individual predicates, whereas the TL class dclinition will be kept unexpanded such that the tat'get gratnmar might be able to choose one of the idiosyncratic prepositions. null Mixed approaches like Kaptan et al. (1989) can be characterized by mapping syntax ,as well as a predicate-m'gument structure (f-structure). As  already pointed out, e.g. in (Sadler and Thompson, 1991), this kind of transfer has problems with its own multiple level mappings, e.g. handling of verb-adverb head switching, and does not cleanly separate monolingual from contrastive knowledge, either. In Kaplan and Wedekind (1993) an improved treatment of head switching is presented but it still remains a less general solution.</Paragraph>
    <Paragraph position="10"> A semantic approach is much more independent of different syntactic analyses which axe the source of a lot of classical translation problems such as structural and categorial divergences and mismatches. In our approach grammars can be developed for each language independently of the transfer task and can therefore be reused in other applications.</Paragraph>
    <Paragraph position="11"> At first glance, our approach is very similar to the semantic transfer approach presented in Alshawi et al. (1991). It, uses a level of underspecified senmntic representations as input and output of transfer. Tile main differences between out' approach and theirs are the use of flat semantic representations and tile non-recursive transfer rules. Tile set-oriented representation allows much simpler operations in transfer for accessing individual entities (set membership) and for combining the result of individual rules (set union). Furthermore, because the recursive rule application is not part of tile rules themselves, our approach solves problems with discontinuous translation equivalences which tile former approach cannot handle well. A transfer rule for such a case is given in (4).</Paragraph>
    <Paragraph position="12"> Out&amp;quot; current apt)roach is strongly related to tile Shake-and-Bake approach of Beaven (1992) and Whitclock (1992). But instead of using sets of lexical signs, i.e. morpho-syntactic lexemes as in Shake-and-Bake, we specify translation cquivalences on sets of arbitrary semantic entities. Therefore, before entering tile transfer component of our system, individual lexemcs can already be decoinposcd into sets of such entities, e.g. for stating generalizations on the lexical semantics level or providing suitable representations for inferences. For example, the wh-question word when is decomposed into temp loc(E,X), whq(X,R), time(R,X) (lit.: at which time), hence no additional transfer rules are required. Similarly, German composita like Terminvorschlag axe decomposed into its compounds, e.g. termin(i2), n n(il,+-2), vorschlag(il) where n_n denotes a generic noun-noun relation.</Paragraph>
    <Paragraph position="13"> As a result a compositional translation as proposal for a date is possible without stating any additional translation equivalences to the ones for the simplex nouns.</Paragraph>
    <Paragraph position="14"> Another major difference is the addition of coilditions which trigger and block the applicability of individual transfer rules. For instance in the specific t,'anslation of schlecht to not good as defined in (3c), without conditions, one would have to add tile verb passen into the bag to test for such a specific context. As a consequence the translation of the verb needs to be reduplicated, whereas in our approach, the translation of the verb can be kept; totally independent of this specific translation of tile adverbial, because the condition functions merely as a test.</Paragraph>
    <Paragraph position="15"> These examples also illustrates the usefulness of labeled conditions, because the negation operator can take such a label as an argument and we can use unification again to achieve the correct coindexation. If we would use a hierarchical semantics instead, as in the original Shake-and-Bake aproach, where the negation operator embeds the verb semantics we would have to translate schlecht (e), passen(e) into not(suit(e), well(e)) in one rule because there is no coindexation possible to express the correct embedding without the unique labeling of predicates.</Paragraph>
    <Paragraph position="16"> Finally, we have filled the lack of an adequate control strategy for Shake-and-Bake by developing a nonmonotonic control strategy which orders more specific rules before less specific ones. This strategy allows the specification of powerfifl defanlt translations. Whereas without; such an ordering special care is needed to prevent a compositional translation in cases where a more specific noncompositional translation also exists.</Paragraph>
    <Paragraph position="17"> The same argument about control holds in comparison to the unification-based transfer approach on Mimimal Recursion Semantics (MRS) (Copestake et al., 1995; Copestake, 1995). In addition, we use matching on first order terms instead of feature structure unification. Full unification might be problematic because it is possible to add arbitrary information during rule application, e.g.</Paragraph>
    <Paragraph position="18"> by further unifying different arguments. The other main difference is our nonmonotonic control component whereas tile MRS approach assumes a monotonic computation of all possible transfer equiv'&amp;lences which are then filtered by the generation grammar. It is difficult to judge the feasibility of their approach given the fact that only a limited coverage has been addressed so far.</Paragraph>
  </Section>
  <Section position="6" start_page="319" end_page="320" type="metho">
    <SectionTitle>
5 Implementation
</SectionTitle>
    <Paragraph position="0"> A more detailed presentation of the implementation aspects of our transfer approach can be found in Dorna and Emele (1996). The current transfer implementation consists of a transfer rule compiler which takes a set of rules likc the one presented in section 3 and compiles them into two executable Prolog programs one for each translation direction. The compiled program includes the selection of rules, the control of rule applications and calls to external processes if necessary.</Paragraph>
    <Paragraph position="1"> Because both the transfer input and the matching part of the rules consist of sets we can exploit ordered set operations during compilation as  well as at runtime to speed up the matching process and for computing common prefixes which axe shared between different rules.</Paragraph>
    <Paragraph position="2"> The compiled trazlsfcr program is embcdded in the incremental and parallel axchitecture of the Verbmobil Prototype. Interaction with external modules, e.g. the domain model and dialog module or other inference components, is done via a set of predefined abstract interface functions whidl may be called in the condition part of transfer rules. The result is a hilly transpaxent and modulax interface for filtering the applicability of transfer rules.</Paragraph>
  </Section>
class="xml-element"></Paper>