File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/abstr/90/c90-2023_abstr.xml
Size: 20,451 bytes
Last Modified: 2025-10-06 13:46:52
<?xml version="1.0" standalone="yes"?> <Paper uid="C90-2023"> <Title>amp;quot;TRANSLATION GREAT PROBLEM&quot; - ON THE PROBLEM OF INSERTING ARTICLES WHEN TRANSLATING FROM RUSSIAN INTO SWEDISH</Title> <Section position="1" start_page="0" end_page="0" type="abstr"> <SectionTitle> &quot;TRANSLATION GREAT PROBLEM&quot; - ON THE PROBLEM OF INSERTING ARTICLES WHEN TRANSLATING FROM RUSSIAN INTO SWEDISH </SectionTitle> <Paragraph position="0"> The problem to be discussed here - i.e. how to generate exponents of a morphosyntactic feature which is systematically used in the target hmguage, but not in the source hmguage is ch)sely related to the development of SWETRA - a multilanguage MT system for translating between fragments of Russian, Swedish, English and German (Sigurd & Gawroffska-Werngren, 1988). Anyone working on translation between Russian and Germanic languages must face difficulties as Russian NPs do not have either indefinite or definite articles.</Paragraph> <Paragraph position="1"> The solutions proposed here have been implemented in the SWETRA - program, which is based on a functional GPSG formalism called Referent Grammar (RG; Sigurd 1987).</Paragraph> <Paragraph position="2"> RG-rewriting rules, implemented in Definite Clause Grammar, are used both for analysis and synthesis. The result of parsing is a so-called functional representation (f-representation), containing descriptions of the constituents and information about their syntactic functions. An f-representation of a simple transitive sentence like &quot;a boy met a girl&quot; looks like this: s(subj(np(r(_,m(boy,sg),indef, sg .... ), Attrl,Relcll)) pred(m(meet,past)), obj(np(r(_,m(girl,sg),indef, sg ...... ) Attr2,Relcl2)), sadvl(\[\]),sadvl(\[\]),advl(\[\]), advl(\[\]),advl(\[\])).</Paragraph> <Paragraph position="3"> The entity with the functor r, called &quot;referent nucleus&quot;, is a description of the head noun. Slots Attrl/Attr2 and RelcllRelcl2 are used, respectively, for storing possible attributes and relative clauses.</Paragraph> <Paragraph position="4"> Given an instantiated f-representation, the program can generate the target equivalent of the input string according to target-specific rules. But if a certain value required in the target language (as definiteness in Swedish and English) is unspecified in the source language (as definiteness in Russian), the information stored in the f-representation may be insufficient for generating a grammatically correct output (although the output may be comprehensible). So there is a need of an intermediate (transfer) stage between analysis and synthesis. The most probable definiteness values must be derived from the context before the target rules for marking definiteness start to work. Since the notions of reference and co-reference are crucial when choosing definiteness values, this intermediate stage will be called &quot;referent tracking&quot;.</Paragraph> <Paragraph position="5"> A preliminary discourse model for referent tracking Informally, discourse referents are often defined as &quot;things the sender is talking about&quot;. Referring means primarily pointing out objects and facts in the external world, but we have also to pay attention to those linguistic factors which enable identifying two or more phrases as co-referential. Obviously, two co-referential words or strings of words do not have to point out a physically existing thing: they may allude</Paragraph> <Paragraph position="7"> to an event or an abstract concept. So discourse referents nmst be understood as cognitive entities existing in the mental world.</Paragraph> <Paragraph position="8"> In the program for referent tracking discussed below, a distinction is drawn between nominal referents - alluding to objects: cats, unicorns etc. - roughly, to things which can be pointed out by non-linguistic means, in potential (unicorns may be pointed out on a picture or drawn) and &quot;event referents&quot; referents of whole predications or predicative (verbal) NPs. &quot;Event referents&quot; correspond to situations;, actions or relations between objects. This distinction is not unproblematic (there are obviously borderline cases), but it is useful for translation purposes, since definiteness may be triggered not only by an NP, but also by a predication as a whole. As will be shown below, the rules for discovering co-reference have to be formulated in different ways depending on which kind of referent (nominal referents or events) is involved.</Paragraph> <Paragraph position="9"> Referent tracking and generation of definiteness values A model for generating definiteness cannot be based on the simplistic principle: if an NP with a given meaning has been translated previously (in the cm'rent text), provide it with the value &quot;definite&quot;; otherwise, treat it as indefinite. In order to instantiate the definiteness value, we have to investigate the internal structure of the NP, the interplay between the current NP and the other syntactic constituents of the analyzed sentence as well as the relations between the current NP and the previously translated part of the text.</Paragraph> <Paragraph position="10"> The preliminary procedure inserting definiteness values used in the RG-model conrains the :following stages: A. Investigating the functional representation of the first sentence of the input text in order to create a &quot;preliminary discourse frame&quot;. B. Storing the descriptions of noun phrases (including their referent numbers) and representations of &quot;events&quot; in a data base. C. Comparing the representations of noun phrases in the current sentence with the stored information in order to discover possible coreference; storing new &quot;events&quot; and new &quot;nominal referents&quot;, if any.</Paragraph> <Paragraph position="11"> The right noun phrase form is then generated according to language specific rules - e.g. rules which do not allow NPs like *the my book or Swedish *rain boken (my book+def) and rules inserting possessive pronouns before nouns denoting close relationship, like &quot;brother&quot;, &quot;neighbour&quot; etc. A Russian sentence likeJa vstretil soseda (I met neighbour) is translated into Swedish as Jag trgiffade rain granne I met my neighbour.</Paragraph> <Paragraph position="12"> Stage A includes subprocedures like: - checking if the current sentence is a predicative construction as &quot;X is a great linguist&quot;; if yes - the referent representation of X has to be provided with the attribute meaning &quot;great linguist&quot; before storing in order to enable co-reference identification in the later part of the text, where X may be referred to by an NP like &quot;this great linguist&quot;.</Paragraph> <Paragraph position="13"> - checking whether the sentence contains specific time and/or place adverbials, whether the current NP contains any attributes which may be interpreted as definiteness indices and whether there are any constituents having clearly specific reference. The aim is to classify the current NP and the whole predication as to their reference: if the sentence evokes many specific concepts and/or the NP contains refercn(:e restricting attributes, we may assume, that the event referred to is highly specific, and that the probability for definite articles may become greater (if no counterindices can be found). The results are not always plausible and can probably be improved by more work on topic - comment relations. Currently, when translating a sentence fi'agment like: v&ra ve~erom Michail Gorba~ev yesterday evening Michail Gorbachev vydvinul predlofenie oh...</Paragraph> <Paragraph position="14"> made proposal about the program inserts tile vahle &quot;prodcf&quot; (probably definite) in the representation of the noun meaning &quot;proposal&quot;, as the discourse frame is highly specific: it contains a specific time value, a specific subject referent and a specific:ation of the noun meaning &quot;proposal&quot; by means of a prepositional phrase. Thus, the Swedish translation version below gets greater preference: igdr kv4ill lade Michail Gorbatjov yesterday evening put Michail Gorbachev fram ,/~Wslaget ore...</Paragraph> <Paragraph position="15"> forward proposal+def about although many native speakers of Swedish would prefer the alternative variant: igdr kvOll lade Michail Gorbau'ev yesterday evening put Michail Gorbachev fram ett f/Srslag ore...</Paragraph> <Paragraph position="16"> forward a proposal about The second wuiant is of course not excluded by the subprocedure. Nevertheless, even if the first output is not always the most preferred one, checking the degree of specificity is often useful, tf we deleted this part of the translation procedure, every NP in the first sentence of a text would be understood as indefinite, something which would lead to many &quot;strange&quot; translations (a professor at a depar#Jzelzt of linguistics at a universiO, qf Lund).</Paragraph> <Paragraph position="17"> If the first sentence in the text does not contain any definiteness indices, the definiteness slot remains anonymous and gets tile default value &quot;indef(inite)&quot; during the generation process, if no target-specific rules prevent it, The information supported by the sentence is stored in two lists: a &quot;nominal referent list&quot; for characteristics of those NPs which have been interpreted as establishing nominal referents, and an &quot;event list&quot;, where representations of predications (including those expressed by verbal nouns) are placed. Each new NP to be translated is now compared with the stored information - the aim is to discover possible definiteness triggers. The simplest case of definiteness triggering is that of nominal co-reference (d~e current NP points out a nominal referent which has been introduced before). Nevertheless, a procedure handling this &quot;simple&quot; case must be quite elaborated, as it has to cover at least the following cases: - co-reference between NPs with identical head nouns: here, the program must check if the current NP contains attributes which exclude co-reference with a previously translated NP having the same head-meaning code, In a sequence like A boy played with a little dog.</Paragraph> <Paragraph position="18"> 77~en, a big dog came the two dogs must not be interpreted as co-referential. This is achieved by a subprocedure &quot;attribute_conflict&quot;, which compares the attributes of the NPs involved.</Paragraph> <Paragraph position="19"> - co-reference between synonyms or between a hyponym and a hyperonym: the program must be able to trigger the value &quot;prodef&quot; if the current NP evokes a concept which is not more 3 135 restricted than and not incompatible with a previously stored referent. Thus, the strings my old teacher and man should be identified as co-referential in a sequence like: I met my old teacher. The man was drunk; but not in I met a man. My old teacher was drunk.. Furthermore, if the current NP refers to a set of objects, we have to check if there are at least two previously established referents which - n'eated as elements of a set - constitute a potentially co-referential set (cases like: A boy met a girl. The children ran home). For this purpose, recursive PROLOG-predicates searching for possible hyponyms in the referent list are used. One of the simpler versions of the predicate for co-reference discover, (the one handling cases like boy+girl=children+def) is formulated as follows: possible_coref(m(A,pl),Rlist):hyponyms(m(A,sg),\[HlT\],Rlist). null where m(A,pl) is the meaning code of the current noun, Rlist is a list containing codes of previously translated noun phrases and the possible hyponyms of the singular form meaning A are stored in the list \[HIT\]. The whole rule is to be read as: a plural noun with the meaning code m(A,pl) may co-refer with a set containing referents of previously mentioned NPs, if at least two previously mentioned nouns can be interpreted as hyponyms of the singular form of the cun'ent noun. The predicate &quot;hyponyms&quot; utilizes the semantic features stored in lexical entries in order to establish a hierarchy between meaning codes.</Paragraph> <Paragraph position="20"> - co-reference between evaluating and nonevaluating expressions - as in the following fragment of a Pravda-notice: Israeli airplanes staged three bomb-attacks on Lebanese territory today.</Paragraph> <Paragraph position="21"> Fifteen persons were killed as a result of the barbaric action of the air pirates.</Paragraph> <Paragraph position="22"> The evaluation of israeli airplanes as &quot;airpirates&quot; depends obviously on the sender's attitude, and such aspects as the sender's political and emotional preferences are not accessible to the program. But evaluating components seem not to restrict the potential reference of an NP in a purely linguistic way (any human being may be referred to by an NP like this fool). Therefore, we may assume, that if the general condition for possible co-reference (not incompatible and not more restricted) is fulfilled after extraction of evaluating elements from the semantic characteristics of the current NP, definiteness may be triggered. In the example above, after deleting evaluations from the lexical description of the entity &quot;air-pirate&quot;, the features corresponding to the concepts &quot;airplane&quot; and &quot;pilot&quot; remain. Consequently, co-reference with &quot;israeli airplanes&quot; is not excluded. - whole - part relations: in cases like car engine etc. definiteness should be triggered.</Paragraph> <Paragraph position="23"> Formulating a PROLOG-rule handling this kind of relation is not a difficult task - the problem is to create an appropriate data base (it would be necessary to include much encyclopaedic knowledge in the lexicon).</Paragraph> <Paragraph position="24"> Another type of definiteness triggering rules applies in the case of co-reference between sequences alluding to 'evenls&quot;, as in the following example: An unidentified submarine followed a Swedish trawler.</Paragraph> <Paragraph position="25"> The hunt went on for about two hours.</Paragraph> <Paragraph position="26"> The first step is to check whether the current noun (here: hunt ) may be interpreted as having an &quot;event-referent&quot; - the information is 136 4 provided in the lexicon. Then, a specific rule for possible event-co-reference applies. It would not be sufficient to compare the semm~tic representation of &quot;hunt&quot; with that of the finite verb (&quot;follow&quot;) according to the previously outlined principle: &quot;not incompatible and not more restricted&quot;. &quot;Hunting&quot; is obviously a more specific concept than following (hunting is a special type of following). As the NP meaning &quot;hunt&quot; refers to an event, we have to treat it as a predication and compare it with the previously mentioned predication as a whole.</Paragraph> <Paragraph position="27"> The event-list contains at this point a representation formulated as: e(hunt,m'gs(r( 1 ,submarine ,unidentified), r(2,trawler, swedish))) The event referred to by hunt has no syntactically represented m'guments - before co-reference checking it gets a representation like: e(hunt,args(_,_)). Co-reference seems to be allowed by the following principle: a verbal noun may co-refer with a prcvious predication, if it is semantically not incompatible with the predicate and if the argulnents of the verbal noun are either not specified or co-referential with the arguments of the previously stored predicate. A PROLOG-implementation of this rule may have the following shape (simplified):</Paragraph> <Paragraph position="29"> The ease of &quot;pseudo-objects&quot; In the example above, both syntactic arguments of the transitive verb were clearly referential they pointed out specific objects. But there m'e cases in which the syntactic complement of a verb does not allude to a referent - though the form of the complement is nominal. The distinction is manifested clearly in Swedish, where the stress pattern of tile string verb + complement varies depending on whether tile complement is referential or non-referential. In tile second case, the stress pattern is identical with the one of particle verbs. Furthermore, tile complement cannot take relative clauses: i. han hall tel._2/ he made speech ii.* han hall ta__I sore var fint he made speech that was fine If hall takes an object proper, as in iii., the stress pattern changes: iii.han hall ett (ldngt) tal sore vat.tint he nmde a (long) speech that was fine The unability versus ability of taking relative clauses is highly significant and can be taken as a criterion tbr referent establishment. According to RG (Sigurd 1989), the head noun, the relative pronoun and the relativized (lacking) constituent in the subordinate (defective) clause are considered as alluding to the same referent. The ungrammaticality of relative clauses other than sentence relatMzing ones can be explained by the fact that the &quot;pseudo-object&quot; lal lacks a referent of its own. The only accessible referent which can be common for the relative pronoun and the lacldng constituent in the relative clause is the referent of the whole predication - as in iv.&quot; iv. hart h61l tal vilket var ,tint he made speech which was fine Vilket is the only Swedish pronoun used for sentence relativization. The sentence above may be paraphrased as: det var ji'tzt art hart h?~l! tal ('it was fine that he made a speech') or as art 5 137 han h611 tal var tint ('that he made a speech was fine') but not as *han h6ll tal sore var .tint ('he made a speech that was fine').</Paragraph> <Paragraph position="30"> Subsequently, components which cannot contain relative clauses are treated as incapable of establishing referents of their own. In the referent tracking procedure, they are interpreted as components of the verbal part of an event.</Paragraph> <Paragraph position="31"> The translation problem arising here is caused by the fact that the distinction between referential objects and &quot;pseudo-objects&quot; is not manifested in Russian. Both v. and vi. are possible: v. on proiznes re~' he &quot;made&quot; speech vi. on proiznes (dlinnuju) reg', kotoraja he made (long) speech that nikomt~ ne ponravilas' nobody+dat not liked v. may thus be translated into Swedish either as hart h61l tal or hart h6ll ett tal.. This translation procedure preserves the anabiquity. If there are neither relative clauses nor other attributes before/after a foma which may be interpreted as a &quot;pseudo-object&quot;, and if there are no counterindices (e.g.clearly anaphoric expressions in the next following part of the text) the non-referential interpretation is preferred, but the second alternative (han h61l ett tal ) is not excluded.</Paragraph> <Paragraph position="32"> Summary The model and procedures discussed above are attempts to utilize text semantic restrictions in machine translation. The current version of the program covers quite a large repertoire of different types of definiteness-triggers and handles generation of correct forms of &quot;pseudo-objects&quot; in phrases like &quot;play the piano&quot;, &quot;play footboll&quot; etc. quite successfully. Nevertheless, there is a need for further study - among other problems, on the &quot;life-span&quot; of discourse referents and on cases where NPs traditionally (i.e. according to Karttunen 1976) treated as non-referential (e.g. predicatives) allow certain instances of definite anaphora (Frarud 1986).</Paragraph> <Paragraph position="33"> The semantic representations of lexical entries require elaboration, and storing non-linguistic knowledge necessary for appropriate definiteness triggering is a problem. Currently, the program works quite efficiently when translating short text fragments, where the number of discourse referents is not too great.</Paragraph> </Section> class="xml-element"></Paper>