Clothing Meaning in Syntax: Aspects and Applications of Multilingual Generation

Clothing Meaning in Syntax: Aspects and Applications of Multilingual Generation
Greg Lessard
Queen's University at Kingston
lessard@francais.queensu.ca
Michael Levison
Queen's University at Kingston
levison@qucis.queensu.ca
Keywords: natural language generation, multilingual
Background
Despite the dominance of English, and despite the loss of small languages at an enormous rate (Robins and Uhlenbeck 1991), we live in an increasingly multilingual world. The growth of telecommunications, the rise of the Internet and the increase of global commerce have led to exposure to a richer spectrum of languages than ever before. This poses a challenge both for education, where we must more and more quickly provide people with access to basic tools for communication, and for documentation, where it is frequently necessary to provide multilingual versions of the same document.
Clearly, one solution to this problem lies in machine translation. In many systems of the sort (we exclude statistically-based approaches) a source language text is first transformed into a metalanguage which is then used to produce a target language text, as the following figure illustrates.

source language text --> METALANGUAGE --> target language text
While such an approach can be effective (particularly in restricted domains) it poses problems stemming from the difficulty of parsing input texts into the metalanguage. Natural languages are notoriously ambiguous out of context, and even with contextual clues provided, parsing is often incomplete.
In what follows, we will present an alternative model for producing multilingual texts. In the proposed model, the starting point is a representation in a common metalanguage, which is then interpreted appropriately by separate linguistic representation tools for each language to be output. The following figure illustrates this approach.

METALANGUAGE --> L1 linguistic representation --> L1 text
--> L2 linguistic representation --> L2 text
--> etc.
This approach is not entirely new, as the following examples (based on English and French) illustrate.
Contant 1988 describes the FRANA system (based on the earlier ANA system for English (Kukich 1983)) which uses stock market information as a starting point to drive text generation in French. In both languages, input consists of unordered information, and output of paragraphs of coherent descriptions of stock market activity. As Contant points out, the task is simplified by the use of a closed domain, and further by the choice of phrases as the minimal unit of generation, rather than words. At the same time, the manipulation of the metalanguage is embedded in the code of the program itself, which reduces the possibility of transfer to other domains. Finally, there is no interaction between the two languages which are generated.

More recently, Danlos (1987) describes a syntactic component for the generation of French and English. Her system has the advantage of greater domain-independence, the use of lexical as opposed to phrasal units, while sharing the use of a more easily manipulable non-language specific metalanguage. However, it too fails to provide for simultaneous generation of two languages and provides no interaction between the two. At the same time, Danlos restricts her attention to relatively coarse syntactic and morphological differences between the two languages, such as agreement phenomena.

Finally, Evraert and Van Steenberghe (1994) discuss the problem of representing referential expressions in English and French in the GENESE system. As do others, they use a non-language dependent metalanguage combined with attribute-value representation of information and inheritance of traits. However, they do not discuss entities beyond the noun phrase and like the others, provide for no interaction between the utterances generated in different languages.

Framework
In the proposed paper, we attempt to deepen the focus on multilingual generation, with particular emphasis on the following problems:
constraints on the nature of the metalanguage;
the problem of interlanguage variation;
the problem of intralanguage variation;
interaction between languages.
We will illustrate the approach using examples from English and French. These are generated using the VINCI system, which is a set of linguist- friendly metalanguages for representing the semantics, syntax, lexicon and morphology of a language, together with a set of interpreters for generating output from these metalanguages. Among other features, the system includes partially ordered compoundable attribute classes which capture both inheritance and synthesis of traits, as well as deconstruction of compound traits, attributed phrase structure rules including guarded syntax rules, complex lexical specifications including pointer mechanisms for interword relations, and morphological interpretation. (For details, see Levison and Lessard, 1992; Lessard and Levison 1995.)
Examples will be drawn from two applications of the system, the first as a tool for generating bilingual instructions, the second as a tool for teaching translation theory with specific reference to English and French.

Metalanguage
Ideally, the metalanguage for a generation system must have at least the following pairs of traits, which, unfortunately, are usually mutually exclusive. On the one hand, it must be precise, allowing the user to specify to the appropriate level of detail the utterances to be produced. At the same time, it must be compact so as not to impose a significant strain on the user in terms of retention and ease of use. Similarly, it must be intuitive, reflecting as closely as possible what the user would think of expressing in his or her native language. However, at the same time, it must be as far as possible non language- specific in order to be accessible to the greatest range of users.
Clearly, an appropriate balance must be struck between these competing objectives. Consider some specific cases. Take the problem of explaining how to attach a knob to a cabinet, in which the knob has two parts, the knob itself and the screw used to attach it. In such a case, it would be sufficient for the metalanguage to provide a representation for each of the two entities, screw and knob. Compare this with the problem of explaining how to assemble a set of shelves, in which three separate sizes of screws are required. Should the metalanguage provide a separate representation for each type of screw? Clearly, this will depend on the amount of world knowledge the reader will be capable of supplying, as well as the importance of specifying which screw is used where. Ideally, the metalanguage should allow both options, either precision or generalization, as well as the superclass-subclass relation which links the two. VINCI provides such a mechanism by means of the partial ordering of attribute information. So in the example discussed above, an appropriate metalanguage might take the following form:

quarterinchscrew<screw, {< = subclass-superclass}
halfinchscrew<screw,
inchscrew<screw,
The distinction between intuitive and non language-specific is perhaps even more complex. On the lexical level, there is clearly an advantage in the metalanguage being similar to the language of use of the person producing the text (thus, screw in English, but vis in French). One obvious solution would consist in providing a metalanguage close to the native language of the user, together with mappings from lexical items of the first language to those in each of the others. Thus, if the first language is English, and the English string generated is a digital watch, each of the lexical items might contain a pointer to a French equivalent: a --> un, digital --> digital, watch --> montre. However, such a system fails to account for agreement and word order (in French, we need to produce une montre digitale).
A more abstract solution consists in defining a non-ambiguous metalanguage which is lexically similar to one of the target languages, but which at the same time is prior to and separate from questions of syntax. To the extent that the range of metalinguistic items is controlled to avoid ambiguity, translation from one metalanguage to another is possible, in order to reflect the needs of different users. (So for example, the attributes screw and knob in English could be mapped to vis and bouton in French.)

At the same time, the metalanguage should be capable of capturing questions of salience, as argued by McDonald (1993), who rejects the use of first-order logic for this reason. For example, as Flores d'Arcais (1987) has shown, test subjects shown two objects of which one is either larger or to the left of the other will tend to construct their sentences with the more salient object (the larger or the left-hand one) earlier in the sentence. One means of capturing this involves using compound traits in the metalanguage to indicate degrees of salience.

Furthermore, there are advantages in allowing maximum syntactic freedom in the order of metalanguage terms without allowing this to affect the final output. Consider, for example, the position of the inflected verb in English and German.

Ich zeige dir, wie man das macht.
I will show you how one does that.
Here, we would prefer that the syntax of English produce the order VERB + PRON, while the syntax of German produces the order PRON + VERB. At the same time, we would prefer that the metalanguage not be forced to deal with such questions, or rather that it allow the user to choose the order of terms which is normal for him or her, without this affecting the final output. This increases in importance as the range of languages increases. It is well known that the world's languages may be divided according to their placement of Subject, Verb and Object in simple clauses. In principle, the metalanguage itself should be insensitive to this variation, relying on the different linguistic representation filters to sort this out. So, for example, in the VINCI system, the following metalanguage expressions are equivalent:
S[courier.subj, deliver.past, parcel.objd, yesterday]
S[courier.subj, parcel.objd, deliver.past, yesterday]
S[yesterday, deliver.past, courier.subj, parcel.objd]
As we will see below, it is the configuration of attribute traits which triggers appropriate (and potentially different) syntactic patterns in various languages.
Interlinguistic Variation
The detailed analysis of the differences between linguistic systems forms the object of differential linguistics. Such differences go far beyond simple morphosemantic traits. In what follows, we will draw our examples from Vinay and Darbelnet (1967), a seminal work in the comparison of English and French. These are in no way exhaustive, but serve only to illustrate how interlinguistic variation can be handled.
Consider first the relatively simple problem of adjective order. With a small number of exceptions, English adjectives precede the modified noun, as in (the red door, the fourth candidate, the parliamentary debate, a big truck). In French, on the other hand, most adjectives follow the noun (la porte rouge, le débat parlementaire). There does exist, however, a set of adjectives which precede the noun, including ordinal adjectives (le quatrième candidat) and a set of (usually) monosyllabic and frequent adjectives relating to size, evaluation, etc. (un gros camion). This is further complicated by the fact that certain syntactic structures such as coordination cause such adjectives to follow the noun (un camion gros et vert but not *un gros be vert camion).

If we suppose that the metalanguage itself is insensitive to the order of elements, such that S[truck.subj, red.subj] and S[red.subj, truck.subj] are equivalent, then the linguistic representations of English and French must account for any differences. This is accomplished by means of guarded syntax rules which take the following form:

PARENT NODE =
< GUARD {attributes present on the parent node}
SYNTACTIC STRUCTURE {of child nodes, with attributes}

< GUARD {another possible collection of attributes}
SYNTACTIC STRUCTURE {of child nodes, with attributes}

< GUARD etc.

> OPTIONAL DEFAULT SYNTACTIC STRUCTURE {with attributes}
Thus, the English syntax may be relatively simple (note that this is only a partial representation).
E_NP = {English noun phrase}

inherit Ns : Nounsemantics.Function, {ex. truck.subj}
As : Adjsemantics.Function; {ex. red.subj}

ADJ[As] N[Ns]
%
The French syntax will be more complex (again, what follows is only a partial representation, which leaves aside questions such as number and gender agreement). We assume that the trait red is a subset of the class preposed
F_NP = {French noun phrase}

< and.Function, preposed.Function :
{if the adjective is of the preposed class and there is
a situation of coordination (and.Function)}

inherit A1 : Adjsemantics.Function,
A2 : Adjsemantics.Function,
Ns : Nounsemantics.Function,
Cs : Coordination.Function;

N[Ns] ADJ[A1] CONJ[Cs] ADJ[A2]

< preposed.Function : {if the adjective is of the preposed class}

inherit As : Adjsemantics.Function,
Ns : Nounsemantics.Function;

ADJ[As] N[Ns]

> {default}

inherit Ns : Nounsemantics.Function,
As : Adjsemantics.Function;

N[Ns] ADJ[As]
%
The same sort of mechanism is used to deal with a variety of differences between English and French, including:
the fact that English has access to a large number of distinct lexical items used to describe sensory phenomena such as sounds (clink, clatter, tinkle, etc.) whereas French tends to use more general terms bruit;
the fact that English uses negative manner adverbs (joylessly, hopelessly where French often relies on prepositional phrases (sans joie, sans espoir);
the use of verb and postposition constructions in English (cut up, chop down) including the possibility of separating the two components (cut the plank up) as opposed to simple verbs in French (découper, couper).
Intralinguistic Variation
As every translator knows, there often exist several possible target language equivalents for a given source language structure. These may be syntactic, as the choice between active and passive voice. In some instances, there are advantages in having access to more than one variant. The mechanism of transformations provides one device to accomplish this. If we assume that output of generation may take the form of an in principle unlimited number of parallel trees, then each tree can represent the product of some range of transformations, or none at all. Thus, given the metalinguistic specification:
S[courrier.subj, deliver.past, parcel.objd, yesterday] %
the application of transformations will provide parellel output including:
NO TRANS: The courrier delivered the parcel yesterday.
PASSIVE: The parcel was delivered by the courrier yesterday.
CIRC_INITIAL: Yesterday the courrier delivered the parcel.
PASSIVE, CIRC_INITIAL: Yesterday the parcel was delivered by the courrier.
In other cases, variant lexical formulations are possible. This is accomplished by means of lexical pointers embedded within lexical items and triggered as needed. In a manner reminiscent of WordNet (Miller et alii, 1993), links are provided to synonyms, antonyms, hyponyms, etc.
Interaction between languages
Ideally, a multilingual generation system should be capable of outputting two or more languages simultaneously, while mapping between them. In the presentation, we will show how the use of transformations allows this to occur.
Applications
In the full presentation, we will illustrate the use of the multilingual generation system to two areas: multilingual text generation and teaching of translation theory. In the first, we will provide a more detailed illustration of generation in a variety of subdomains, showing the interplay between the various components of the system. At the same time, we will discuss problems of scale, as well as the interplay of the various levels of expertise required, ranging from computer scientist (software questions), to linguist (linguistic models) to end-user (formulation of textual content in the metalanguage). In the second, we will show how the system allows the presentation of one language as a source, the ability to sollicit proposed equivalents from learners, and the comparison of these with various possible translations.
References
Contant, C. (1988) Génération automatique de rapports boursiers français et anglais. Revue québécoise de linguistique 17/1:197-222.
Danlos, L. (1987) A French and English Syntactic Component for Generation. In G. Kempen (ed.) Natural Language Generation: New Results in Artificial Intelligence, Psychology and Linguistics. Dordrecht: M. Nijhoff, pp. 191-217.

Evraert, G., T. Van Steenberghe. (1994) The Realization of Individual Instances in a Multilingual Generation System. Meta 39/1:194-205.

Flores d'Arcais, G.B. (1987) Perceptual Factors and Word Order in Event Descriptions. In G. Kempen (ed.) Natural Language Generation: New Results in Artificial Intelligence, Psychology and Linguistics. Dordrecht: M. Nijhoff, pp. 441-451.

Kukich, K. (1983) Knowledge-Based Report Generation: A Knowledge-Engineering Approach to Natural Language Report Generation, PhD, University of Pittsburgh.

Lessard, G., M. Levison (1995) Le logiciel VINCI: lexigrammaire et génération automatique. In Lexiques-grammaires comparés et traitements automatiques (J. Labelle, ed.) Montréal: Université du Québec à Montréal, pp. 175-185.

Levison, M., G. Lessard. (1992) A System for Natural Language Sentence Generation. Computers and the Humanities, 26:43-58.

McDonald, D. (1993) Issues in the Choice of a Source for Natural Language Generation. Computational Linguistics 19/1:191-197.

Miller, G., R. Beckwith, C. Fellbaum, D. Gross, K. Miller. (1993) Introduction to WordNet: An Online Lexical Database. (From web page at www.cogsci.princeton.edu/~wn/w3wn.html)

Robins, R.H., E.M. Uhlenbeck, eds. (1991) Endangered languages. New York : Berg.

Vinay, J.-P., J. Darbelnet. (1967). Stylistique comparée du français et de l'anglais; méthode de traduction. Paris, Didier.