{
"paper_id": "A88-1023",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T02:04:08.468696Z"
},
"title": "V~TJ~CY AND MT: RECENT DEVELOPMENTS IN THE METAL SYSTEM",
"authors": [
{
"first": "Rudi",
"middle": [],
"last": "Gebruers",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Siemens-METAL project Katholieke Universiteit Leuven Maria",
"location": {
"addrLine": "Theresiastraat 21",
"postCode": "B-3000",
"settlement": "LEUVEN",
"country": "BELGIUM"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper describes a valency model, developed within the Belgian METAL project, aimed at enhancing the modularity and multilinguality of the METAL system. The introduction provides background, section 1 discusses the existing valency framework, and section 2 presents the alternative model. The final section deals with some results and problems with this model.",
"pdf_parse": {
"paper_id": "A88-1023",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper describes a valency model, developed within the Belgian METAL project, aimed at enhancing the modularity and multilinguality of the METAL system. The introduction provides background, section 1 discusses the existing valency framework, and section 2 presents the alternative model. The final section deals with some results and problems with this model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The task of MT is to map between equivalent linguistic objects. One of the central design questions in MT is that of the best method to decompose the translation relation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "0."
},
{
"text": "The ideal would be to have a system that produces a (natural) language-independent representation from a source language (SL) text, which could then be synthesized in any target language (TL).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "0."
},
{
"text": "However, this ideal not being feasible for real-world texts, it has become customary to adopt a model where a transfer module, specific to one language pair, defines a mapping between language-dependent structural representations. In principle it should be possible to design a 'transfer' model in such a way that the analysis module for mapping surface strings onto structural representations and the synthesis module for mapping structural representations onto surface strings remain the same, regardless of the TL and SL, respectively. The advantage of this 'multilingual' design is that existing modules will not be seriously affected by the addition of a new language to the system. A still more attractive, but also more ambitious, design would be one in which the same grammar can be used for both parsing and generating, and the same translation rules for translating between two languages in either direction (see Jin and Simmons, 1986 for an example of a 'symmetric' translation system).",
"cite_spans": [
{
"start": 923,
"end": 944,
"text": "Jin and Simmons, 1986",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "0."
},
{
"text": "systems blended the rules of grammar and the analysis procedure for efficiency reasons, it has also become customary, given current system optimization techniques, to make a clear separation between programming logic and data on the one hand, and linguistic logic and data on the other. This separation is convenient for the division of labour between the linguist and the programmer, and it enables the former to revise and complete his rule systems without the latter having to constantly change his programs. The METAL automatic translation system tries to be multilingual in the above sense.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Whereas early MT",
"sec_num": null
},
{
"text": "Moreover, it makes an attempt at separating software and lingware (= linguistic knowledge written in a specialised user language).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Whereas early MT",
"sec_num": null
},
{
"text": "In the following, I will show how the adoption of a new kind of valency framework, developed at the K.U.Leuven in the course of the last two years, enhances the multilinguality and modularity even further. For the sake of clarity, I will first review the relevant aspects of the valency framework in the current METAL system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Whereas early MT",
"sec_num": null
},
{
"text": "Since the main claim to be advanced in this paper bears on the relation between a valency framework and the general design of an MT system, I will first say a few words on the METAL system architecture. I will then review and comment upon the valency framework in this system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "I. Valency in the METAL system",
"sec_num": null
},
{
"text": "In METAL the translation process proceeds in three phases.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "i.I. The METAL architecturo",
"sec_num": null
},
{
"text": "During transfer instructions are paired one-to-one with the grammar rules used to perform the SL analysis. However, provisions have also been made to complement this \"direct transfer' approach with an independent transfer grammar (see Root, 1985) . The latter approach is becoming more and more important in METAL because it greatly enhances the modularity of the system (viz. with an eye on using it for several different language pairs).",
"cite_spans": [
{
"start": 252,
"end": 263,
"text": "Root, 1985)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 7,
"end": 214,
"text": "transfer instructions are paired one-to-one with the grammar rules used to perform the SL analysis. However, provisions have also been made to complement this \"direct transfer' approach with",
"ref_id": null
}
],
"eq_spans": [],
"section": "i.I. The METAL architecturo",
"sec_num": null
},
{
"text": "It is well-known that the dependency relations between a verb and its arguments can influence greatly the lexical and structural transfer of both, as well as the structural transfer of the clause as a whole.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "1.2o The valency framework",
"sec_num": null
},
{
"text": "Though the dependency relations themselves may be languageindependent, their encoding varies from one language to another, and, within one language, from one verb to another. The structure of valency frames is languageindependent, and can be defined as follows: Finally, the third frame requires a nominal filler and a measure constituent.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "1.2o The valency framework",
"sec_num": null
},
{
"text": "The respective frames are illustrated in the following sentences:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "1.2o The valency framework",
"sec_num": null
},
{
"text": "(1) Je ($0) fais ce Jouet ($I) pour mon ami ($2) \"I make this toy for my friend\" (2) Ella ($0) fair vieille ($5) \"She looks old\" (3) Carte volture ($0} fair 100 km/h ($3) \u2022 This oar does i00 km/h, An example of a (French) frame test is given in fig. i This implies that two kinds of preference mechanisms had to be introduced in the valency procedure: one to choose the best candidate from a set of potential fillers for a given slot (instead of always choosing the first constituent matching the specifications of that slot), and one to choose the most ambitious frame from a set of successful frames (instead of always choosing the first frame that matches a given analysis tree).",
"cite_spans": [
{
"start": 44,
"end": 48,
"text": "($2)",
"ref_id": null
},
{
"start": 108,
"end": 112,
"text": "($5)",
"ref_id": null
},
{
"start": 166,
"end": 170,
"text": "($3)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 245,
"end": 251,
"text": "fig. i",
"ref_id": null
}
],
"eq_spans": [],
"section": "171",
"sec_num": null
},
{
"text": "The first preference mechanism is implemented in the following way. When checking whether a frame matches the tree, the valency procedure collects for each slot ($i keyl...keyN) all sentence constituents which pass all the frame tests associated with the keys in that slot.",
"cite_spans": [
{
"start": 161,
"end": 177,
"text": "($i keyl...keyN)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "171",
"sec_num": null
},
{
"text": "Furthermore, in the action part of frame tests, each potential filler gets a number indicating the probability that this constituent will be taken as the ultimate filler for a slot. As the linguist can easily alter this number, he has significant control over the assignment of constituents to slots.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "171",
"sec_num": null
},
{
"text": "The actual assignment procedure is fairly economical and runs as follows. Whenever there is only one potential filler for a slot (which may be either an obligatory or an optional one), this constituent loses its marking as a candidate filler for other slots, if it was one.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "171",
"sec_num": null
},
{
"text": "Furthermore, it is marked as the only remaining filler for the slot that is being matched. A side-effect of this marking may be that one of the other slots will now have only one candidate filler.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "171",
"sec_num": null
},
{
"text": "In that case, the latter constituent will be marked as the actual filler for that slot and lose its marking as a potential filler for still other slots, if it was one.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "171",
"sec_num": null
},
{
"text": "This may again cause the number of potential fillers for yet another slot to be reduced to one, in which case the above marking (and unmarking) procedure starts over again.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "171",
"sec_num": null
},
{
"text": "If, eventually, there is more than one potential filler for a slot, the procedure takes the leftmost candidate which received the highest preference value for the slot in the frame tests.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "171",
"sec_num": null
},
{
"text": "The second preference mechanism is fairly economical as well. The general assumption is that, in order to find the most 'ambitious' frame, one should look for the frame which has the largest number of slots realized. In order to avoid superfluous processing, we let the more complex frames (i.e. the ones with the larger number of slots, whether obligatory or optional) precede the less complex ones in the ARGS value of a verb. Consequently, the system comes across the former before it sees the latter. Whenever two or more alternative frames happen to have the same number of slots, the lexicographer has to determine (e.g., on the basis of frequency of occurrence) which frame to try first. Given that the system has found a matching frame, it will only explore an alternative frame if the number of (optional and non-optional) slots contained in the alternative frame outnumbers the number of slots found to be realized during the matching of the first frame.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "171",
"sec_num": null
},
{
"text": "An alternative frame will be preferred only when it has more slots realized than the previous matching frame.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "171",
"sec_num": null
},
{
"text": "During transfer, since we take argument structures to be language-specific entities, the valency procedure has to accomplish two tasks. First, it has to determine which TL frame corresponds with the SL frame that has been found to be applicable to the analysis tree.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2.3.2.",
"sec_num": null
},
{
"text": "Secondly, it has to specify which slots in the TL frame correspond with which slots in the SL frame.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2.3.2.",
"sec_num": null
},
{
"text": "It performs those tasks in the following way. respectively; (b) it allows for a neat separation of kernel software and application-specific lingware and provides user-friendly facilities to access and update the latter; (c) its methodological underpinning to a certain extent allows languages to be treated independently of one another.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2.3.2.",
"sec_num": null
},
{
"text": "Because of these advantages, the Leuven valency framework has recently been adopted by all sites of the METAL project.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2.3.2.",
"sec_num": null
},
{
"text": "However, serious problems remain to be tackled, with respect to both lexicon coding and grammatical parsing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2.3.2.",
"sec_num": null
},
{
"text": "The first kind of problems can be traced to the rigidity of the frame mapping schema itself.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2.3.2.",
"sec_num": null
},
{
"text": "As has been pointed out in section 2.3.2., the main requirement for frame mapping to be possible is that verb lexicons be coded consistently across languages. This may indeed be profitable in an experimental environment. However, it is doubtful whether we can expect the average end user to have both source, target, and transfer codings in mind at the same time and to make sure that no aspect of the mapping between frames is overlooked. Therefore, we have developed provisional relaxations on the rigid schema to the extent that at least the mandatory slots of the TL frame must have a counterpart in the SL frame. Eventually, however, it may turn out to be more effective not to have these relaxations at run time, but to sort out inconsistencies between verbal lexicons at coding time.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2.3.2.",
"sec_num": null
}
],
"back_matter": [
{
"text": "Having retrieved all the transfers for the SL verb, the system reduces the potential transfer ambiguity of a SL verb in two steps. It first discards any transfers for which not all conditions are fulfilled. Afterwards, it checks which of the remaining transfer entries (there should be at least one) provides a frame equivalence whose 'left-hand side\" can be linked to the frame realized in the SL tree. Once this equivalence has been found, the TL frame matched by the 'right-hand side' of the frame equivalence will be substituted for the SL frame referenced by a feature on the root node of the tree. At this stage, nodes can be pruned from or added to the tree structure, if lexical instructions tell the system to do so.Finally, after translation of the verbal predicate, the system exploits the equivalences between SL slots and TL slots in order to determine how the translations of the sentence-level constituents fit into the slots of the TL frame. Proceedings of the 2nd ACL 1985 , 70-72. Slocum, J. 1983 'A status report on the LRC machine translation system.'",
"cite_spans": [
{
"start": 981,
"end": 989,
"text": "ACL 1985",
"ref_id": null
},
{
"start": 990,
"end": 1014,
"text": ", 70-72. Slocum, J. 1983",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "annex",
"sec_num": null
},
{
"text": "Conference on Applied Natural Language Processing, 166-173. Somers, H.L. 1986 .",
"cite_spans": [
{
"start": 51,
"end": 77,
"text": "166-173. Somers, H.L. 1986",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Proceedings of the ACL-NRL",
"sec_num": null
},
{
"text": "Linguistics.Edinburgh University Press.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Valency Computational",
"sec_num": null
},
{
"text": "Edinburgh:Van der Korst, B. 1987. 'Twelve sentences: a translation procedure in terms of Functional Grammar.\"Working papers in Functional Grammar 19, University of Amsterdam.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "and Case in",
"sec_num": null
}
],
"bib_entries": {},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"uris": null,
"text": " : :-\" (\" + [ \"OPT\" +] \") = ::--\"(\" + \")\" ::-one of a set of user-definable atoms, startinq with a \"S'-sign ::- I ::-one of a set of user-definable atoms ::_ .(m + \")\" ::-one of a set of user-definable atoms : :-one of a set of user-definable atoms"
},
"FIGREF1": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "RET 0 PREP) if there is a PREP feature on the father node (this feature has been retrieved fr~ a verbal entry) (OR (XFM (&:l (--:2 (NP:3 NIL (INT 3 ROL $2)) ---:4)) and if thera is a NP son marked $2, then create a TL-PREP node in front of it and make both dependent on a new PP node marked $2 or if there is a PP node marked $2, then translate its PREP node (&:l (--:2 (PP:3 ((PREP:7 NIL (TRF i PREP)) &:5)) --:6)))))"
}
}
}
}