Title: From above to under: Enabling the Generation of the Correct Preposition from an Interlingual Representation

(1) Three of the most advanced institutions in the ACCION network started their programmes as non-profit organizations and have, in the last five years, converted into regulated financial institutions that are part of the financial system in their countries but specialize in serving the microenterprise sector. One of the most successful of these institutions is BancoSol in Bolivia.

 

In addition to this polysemy, further reflection will lead to the realization that there are certain cases in which it is difficult to attribute any specific meaning component at all to the preposition itself. That is, prepositions seem not only to vary in meaning according to their context, but in certain cases they seem to be semantically vacuous. As a case in point, it would be difficult to ascribe any isolable meaning to the use of in in the phrase "specialize in serving the microenterprise sector" from the text above.

 

The concept of the semantic vacuity of prepositions is not a new one: one finds accounts of semantic vacuity of prepositions both in the theoretical and computational literature. On the most obvious level, we might note that much of the information contained in prepositions in English is carried by (grammatical) case-marking inflections in more highly inflected languages (Comrie 1987; Fromkin and Rodman 1988). Stockwell, Schachter, and Partee extend this idea to semantic case roles, asserting that whenever a preposition marks case, it is semantically void of any other content, stating that prepositions in their role as case markers "appear to lack independent semantic significance" (Stockwell, Schachter, and Partee 1973). They provide following examples

 

(2) Someone opened the door with the key.

The key opened the door.

(3) The clown was amusing to the children.

The clown amused the children.

 

noting that with and to are used to mark the Instrument used to complete the action, and the Dative case, respectively but do not carry independent semantic content. Van Langendonck (1974) opens his discussion with an example very similar to (2), and states, "But we cannot possibly imagine the meaning of the actant with in the phrase with the key of sentence [2] ... except in terms of the language-internal notion `instrument' 1 . Therefore the preposition with is here a case marker without any lexical content of its own." When we talk of "semantically void" prepositions, then, what is intended is that the only semantic function served by the preposition is to indicate a semantic relationship between a verb and its argument. Thus, prepositions in English (and other analytic languages) and inflections in synthetic languages sometimes serve only to indicate grammatical case, and sometimes also serve to indicate the corresponding nominal's semantic relation to the predicate without containing any additional semantic content. Diachronic support for this commonly accepted concept can be found throughout (Comrie, 1987), and more specifically for Germanic languages and for English in particular in the articles dedicated to these topics (Finegan 1987; Hawkins 1987). Finally, we may note that this phenomenon is not restricted to the Germanic languages, as is demonstrated by Spang-Hansen's (1963) lengthy discussion of the distinction between prépositions vides (`empty prepositions') and prépositions pleines (`full prepositions') in French.

 

3. Semantically Vacuous Prepositions in (Interlingual) Machine Translation

 

Given that some prepositions are semantically void (`content-free' prepositions) and that others maintain their full semantic value (`content-full' prepositions), one may wonder, "What impact does this have on the MT process?" The answer is that the necessity to account for both content-full and content-free prepositions has a great impact, whether the system in question is an IL based MT system, or a more traditional transfer-based MT system. To account for content-less prepositions, a transfer-based system must have the ability to account for null elements in its transfer rules, thus allowing translations of NP ´PP.

In an Interlingual MT system, if a given preposition in an input text has no semantic content, it is not necessary to include it in the IL meaning representation of that text, but it will be necessary to generate any necessary preposition (assuming that one is necessary) in the target text. Given an IL approach, then, in cases where the preposition is semantically void, we can reduce the problem to one of TL generation. That is, we do not need to have an explicit representation of the prepositional component in the IL representation. Rather, we can treat the prepositional translation problem as one of generation of TL text from an underspecified IL representation. This is similar to the view taken in (Barnett, Mani, & Rich, 1993), who view the problems of lexical mismatch and divergence as problems for (monolingual) generation. This is beneficial, because, "By forcing us to focus on the lexical and syntactic structures of a single language at a time, it simplifies our attempt to isolate the sources of information that are necessary to solve the problem" (Barnett, Mani, & Rich, 1993).

 

3.1 Locating the Content-full Prepositions

 

If we accept that certain prepositions do not need to be represented in an IL representation, then a method for determining which ones to include and which to exclude must be devised. Do we include only the ones that are not recoverable during generation? Only semantically non-null (content-full) ones? If the latter, then how do we determine which are content-free and which are content-full?

 

Turning to syntactic cues, Hutchins & Somers (1992) provide some insight into a potential solution to this dilemma. They write:

 

If a prepositional phrase is a complement, we must look to the verb to see how to translate it. On the other hand, adjunct prepositional phrases can usually be translated independently of the verb.

 

Support for the plausibility of this assertion can be found in the work of numerous linguists. For example, with respect to learnability of prepositions heading complement phrases, Celce-Murcia and Larsen Freeman (1983: 253) provide a perspective from within the field of applied linguistics, writing:

 

Many English verbs and adjectives occur with one and only one preposition. Such a preposition must be entered in the lexical entry of the verb or adjective concerned, and it must be learned as an integral part of the verb or adjective even though these are still combinations of V + PP or adj + PP from the structural point of view.

 

Beneath the surface of Hutchins and Somers' and Celce-Murcia and Larsen Freeman's assertions is the assumption of the lack of semantic value of heads of complement PPs in their role as Case assigners, which, as has been shown, is a common interpretation of the role of prepositions heading complement phrases (see section 2).

 

Although Hutchins and Somers' heuristic provides a very good first approximation as to which prepositions are likely to be content-free and which are likely to be content-full, for our purposes, we must go one step further. Specifically, in cases in which the preposition heads a L(ocative) complement prepositional phrase, the heuristic will fail. That is, according to the heuristic, it is necessary to look to the verb to translate (locative) complements, whereas in reality the preposition itself should be the focus. Unfortunately, the workshop text does not afford examples of this phenomenon, but the following simple example will serve to illustrate the point. Although it would be generally agreed that the (Locative) PP following a verb like put is subcategorized for by the verb, many different prepositions are permitted in the head position, and these prepositions seem to translate fairly directly, or at least independently of the verb:

 

(4) He put the book in/ on/ near / under / next to / behind the box.

Il a mis le livre dans/ sur/ près de/ sous / à coté de/ derrière la boîte.

 

In fact, many authors have commented on the distinct nature of Locative and Temporal prepositional phrases 2 . Somers himself confirms this observation, stating "For those verbs taking a Locative complement (especially verbs of motion like go, walk, etc.), a huge range of prepositions can introduce the complement" (1987: 15). The motivation for a special treatment of locative and temporal PPs comes from the fact that, as stated in Fillmore (1968), "the L and the T (for Time) prepositions are either semantically nonempty (in which case they are introduced as optional choices from the lexicon), or they are selected by the particular associated noun...." Nilsen (1972: 24) further expounds on this idea by adding the notion of compositionality:

 

.... And sometimes the preposition carries semantic as well as grammatical significance. This is especially true with Locative and Temporal prepositions. In sentences using these types, the semantic meaning of PP is not the same as that of NP, but rather is the same as P + NP.

 

A similar sentiment is echoed by Somers (1987), who is working from a computational viewpoint. Somers, however, also adds the observation that "Locatives ... are most commonly realized by prepositional phrases."

 

3.2 An Additional Complication

 

An additional complication worth noting (with respect to the use of the complement/adjunct heuristic to distinguish content-full from content-free prepositions) is that the distinction between complements and adjuncts is not always clear-cut. Thus, in his Valency and Case in Computational Linguistics, rather than a binary complement / adjunct distinction, Somers (1987) proposes a five-level 'range' of complementarity, from 'integral complements' (essentially frozen expressions and collocations) to 'extra-peripherals' (which can be equated to sentence modifiers). This finer-grained distinction is important, because the varying levels of the complementarity scale seem to correspond to the degree to which prepositions retain (or fail to retain) their semantic content. For example, consider sentences (5) - (8), which progress from greater to lesser levels of complementarity, accompanied by increasing content-retention on the part of the preposition involved.

 

BancoSol took on large amounts of risk in its early years.

The ACCION network specializes in serving the microenterprise sector.

a. PRODEM talked to BancoSol before investing.

b. PRODEM talked about BancoSol before investing.

BancoSol opened its doors in 1992.

 

If the decision to include the meaning of a preposition in an IL representation were to be based on the complement / adjunct status of the prepositional phrase, it would be necessary to work out a well-defined model of complementarity and to specify the criteria by which distinctions would be made. More likely than not, in an implemented system, the choice would be reduced to a lookup in a lexicon containing subcategorization information.

 

4. Requirements for the IL Representation

 

After having decided what not to include in the IL representation, the next logical question is "what do we include in the IL representation?" Requisite components that must either be present in the IL representation or derivable during generation to enable the choice of the correct TL preposition include:

 

Distinction between locative and non-locative prepositions

 

For non-locative prepositions:

Semantic roles of arguments

 

For locative prepositions:

Stativity of prepositions

Fine-grained locative representation

Relationships between items `linked by' the preposition

 

The following sections will discuss each of these in turn.

 

4.1 Distinction between Locative and Non-Locative Prepositions

 

To say that the distinction between locative and non-locative prepositions must be included in the IL representation is a bit misleading 3 . In actuality, locative and non-locative prepositional phrases should be distinguished in the IL representation implicitly due to differences in the representation of each. That is, locative prepositions could potentially be treated as predicates, as suggested by Davidson (1980), Becker & Arms (from Nilsen (1973)), and as is consistent with the treatment put forth in (Hobbs 1985) 4 . Non-locative prepositions, on the other hand, would have an entirely different representation, which will be discussed in the next section.

 

4.2 Semantic Roles of Arguments: The Potential Role of Case Grammar

 

In sentences (5) and (6) above, the (non-locative) preposition heading the complement prepositional phrase is basically content-free. As previously stated, there is no need to represent such prepositions in the IL representation of these sentences. In (7a) and (7b), however, although the prepositional phrases can still be considered complements, it is evident that these prepositions are somewhat more content-full than those in (5) and (6). Thus, the meaning of these prepositions must be somehow accounted for in the IL representation. We may notice that in (7a), the grammatical object of the preposition is in some sense the recipient of the communication described in the sentence, whereas in (7b), the object of the preposition is in some sense the object of the communication described in the sentence. This is the type of information that it is crucial to retain in the IL representation.

 

In this vein, it is interesting to note that Fillmore's early work on Case (preceding his 1968 'The Case for Case') was actually centered around prepositions, appearing in The Georgetown University Round Table on Languages and Linguistics 1966 under the title 'A Proposal Concerning English Prepositions' (Fillmore 1966). At one point in his career, Fillmore went so far as to propose that there is a direct correspondence between (semantic) case roles and lexical prepositions that are used to realize these roles at surface structure. He posits the following for English (Fillmore 1968: 32):

 

The rules for English prepositions may look something like this: the A[gent] preposition is by; the I[nstrument] preposition is by if there is no A, otherwise it is with; the O[bject] and F[actitive] prepositions are typically zero [Ø]; the B[enefactor] preposition is for; the D[ative] preposition is typically to ....

 

Many others picked up on the idea that the prepositions appearing with a given verb in the surface structure would be predictable, based on the specification of the Case Frame involved. Conversely, others took this to mean that the prepositions appearing in surface structure could be used to determine the underlying Case Frame. For example, Van Langendonck (1974: 2) states, "As a rule, each case is connected with its typical prepositions." Stockwell, Schachter, and Partee (1973), working toward a computational approach, adopt Fillmore's Case Grammar, and state that "certain generalizations can be made to the effect that a given case is typically marked by a given preposition, so that only the use of some atypical preposition to mark that case need be indicated in the lexicon" (Stockwell, Schachter, and Partee 1973: 35). They are careful, however, to note the aforementioned special status of Locative prepositional phrases: "With Locatives there is no single unmarked preposition: all locative prepositions have semantic content that includes more than the feature [+locative]" (Stockwell, Schachter, and Partee 1973: 39). They further comment that a Case Grammar facilitates the making of many generalizations about prepositions that are not possible within other grammatical frameworks.

 

Not all Case Grammarians, however, are enthusiastic about the notion that prepositions are direct surface representations of deep Case Roles. Very early in his discussion of Case Grammar, Nilsen (1972) states, "... I will first show that there is not a perfect correlation between deep cases and overt surface manifestations of these cases," and later, "The most obvious problem is that there is by no means a one-to-one relationship between prepositions and case" (Nilsen 1972: 3, 20). Nilsen's strongest argument, though, is in his counterexamples to the notion of equivalence of Case and prepositions: he cites examples using the preposition with marking no fewer than six Cases, and provides examples of five different prepositions marking the Instrumental Case.

The arguments against a one-to-one relationship between preposition and Case, then, are numerous. It is interesting to note that even in Fillmore's "rule for English prepositions" (above) there is not a simple one-to-one mapping between semantic cases and lexical prepositions: he links the preposition by to both A and I cases, lists both by and with as possible mappings for the I case, and later in the same article, he states that with has a comitative function as well (Fillmore 1968). Surely due in part to its beginnings as a theory focused on prepositions, however, as Stockwell et al. point out, Case Grammar does, provide a uniform way of dealing with and of making generalizations about prepositions. However, even Fillmore finds it necessary to adopt a weaker version of his hypothesis of the strict correspondence between prepositions and case in his later work.

Thus, the role for Case Grammar (or some variant thereof) proposed for use is not to be considered a shorthand or coded version for including lexical prepositions in the IL representation. Rather, it is proposed that in certain cases there is information about the semantic role of the argument that must be included in the IL for successful translation. In fact, this is almost always true, except in constructions containing a single complement prepositional phrase that is filling the semantic role variously labeled as Object, Theme, or Patient. In all other situations, case roles used to indicate the semantic roles of arguments, information which is instrumental during the generation of the target language.

 

4.3 Stative Prepositions

 

Sentences (9) and (10) demonstrate another important meaning component of prepositions: they may be marked [stative]. In both sentences, the prepositional phrase specifies a location, in one case lexicalized as to ([-stative]), and in the other lexicalized as at ([+stative]) 5 . Thus, some Case Grammar models that would view the prepositional phrase to school as filling a directional locative Case role in sentence (9) and the prepositional phrase at school as filling a stative locative Case role in sentence (10)

 

([ stative])

John went to school.

... loans granted every year are rising to $80 million.

 

(10) ([+stative])

John stayed at school.

... loans granted each year are remaining at $80 million.

 

In these sentences, it is possible to posit a specific function for to and at marking [ stative] and [+stative] locative complements, respectively. Sentences (11a) and (12a), the French counterparts of (11) and (12), however, display no such prepositional differentiation:

 

(9a) Jean est allé à l'école.

Lit. `Jean is gone to the school.'

 

Les prêts consentis chaque année s'élèvent à 80 millions de dollars.

Lit. `The loans granted each year themselves-raise to 80 million of dollars.'

 

(10a) Jean est resté à l'école.

Lit. `Jean is stayed at the school.'

 

Les prêts consentis chaque année restent à 80 millions de dollars.

Lit. `The loans granted each year remain at 80 million of dollars.'

 

So, again, it is misleading to say that we should include an explicit indication of the stativity of the preposition in the IL representation. Rather, we could follow the example set forth in Anderson (1971), who uses verbal features to distinguish `static' and `dynamic' (a.k.a. stative and directional) locatives. Adoption of Anderson's placement of the [stative] feature on the verb eliminates the necessity of labeling the nouns filling the argument slots for these verbs as distinctly filling stative and directional locative Case roles. Thus, we can say that while the French verbs aller (`to go') and rester (`to stay') are marked for [ stative] and [+stative], respectively, the French preposition à is undetermined for the feature [stative]. Choosing the correct preposition during generation, then, can be reduced to unification of verbal and prepositional features.

 

4.4 Fine-grained Locative Representation

 

As has been noted in (Dorr & Voss 1993; Trujillo 1992), even seemingly straightforward locative prepositions such as the one in (11) are potentially ambiguous.

 

(11) The mouse ran under the table.

 

Possible interpretations include a path, destination, and location (area) reading of the preposition under. As this ambiguity cannot be preserved in many languages, it would be desirable to choose between these various meanings, and to accommodate a distinct representation for each in the IL. Similarly to the stative distinction, however, Trujillo has demonstrated that, at least for English and Spanish, it might be possible to glean much of this information by means of constraints imposed on possible meanings by other elements in the sentence (namely the verb and the noun that is the object of the preposition).

 

4.5 Relationships between objects `linked by' the preposition

 

So far, we have been concerned with cases in which the choice of preposition depends on the locative status of the preposition, and perhaps on features of the verb to which the prepositional phrase attaches. In many cases, however, the object of the preposition imposes selectional restrictions of sorts on the actual preposition chosen. Take, for example, sentences (12) and (13):

 

(12) ACCION International est un organisme privé ...basé aux États-Unis.

Lit. `ACCION International is an organization private ...based at-the States-United'

Eng. ACCION International is a private ... organization based in the United States.

 

(13) Je l'ai vu sur la photo.

`I saw it on the photo.'

I saw it in the photo.

 

In (12), French selects the preposition à (`to/at') before États-Unis, whereas English chooses `in'. Japkowitz and Wiebe (1991) suggest that in these cases, it is the conceptualization of the NP object that differs between languages, such that while a photo is conceptualized as a container in English (thus the choice of `in'), a photo is conceptualized as a surface in French (leading to the choice of the preposition sur (`on') in French). A similar explanation could be given for the sentences in (12). While this does not necessarily impose any constraints on the IL representation itself, it does suggest the necessity of a language-specific KB during the generation phase. The content of this KB would minimally consist of a conceptual hierarchy to assist in resolving such generation problems. This may not be a simple matter, since a change in the verb can cause differing salient conceptual characteristics of NPs to come into focus.

 

5. Conclusion

 

It has been suggested that in many cases, prepositions are semantically null (`content-free') elements that need not be represented in an IL representation. This places much of the burden of correct prepositional choice on the generation process. Content-full prepositions were then divided into locative and non-locative groups, and necessary information for enabling the generation of the correct preposition from an IL representation was discussed for each. Since not all content-full prepositions have the same `semantic weight', differing amounts of information are necessary for different types of prepositions. It has been further suggested that even in the case of content-full prepositions, much of the information necessary for the generation component can be gleaned from features of the IL representation of other sentential elements.

 

The standard IL problems remain: if the problem of prepositional choice is viewed as a problem of generation from an underspecified IL representation, then all prepositional distinctions necessary for all languages must be provided for in the IL representation. However, as alluded to in section (4.4), it is sometimes not possible to provide information in the representation that is not present in the source text. Finally, the scope of what can be generated or inferred from the IL representation will be constrained by the domain of the IL system. This is particularly true with respect to any conceptual hierarchy that might be used during generation. Providing an IL for unrestricted text, even for the representation of prepositional meaning, is not yet a trivially solvable problem.

 

References

Anderson, John M. 1971. The Grammar of Case Towards a Localist Theory. Cambridge Studies in Linguistics, ed. Sidney W. Allen, Eugenie J. A. Henderson, Fred W. Householder, John Lyons, R. B. Lepage, F. R. Palmer, and J. L. M. Trim, 4. New York: Cambridge University Press.

Celce-Murcia, Marianne, and Diane Larsen-Freeman. 1983. Prepositions. Chap. 19 In The Grammar Book An ESL/EFL Teacher's Course. USA: Newburry House Publishers.

Comrie, Bernard, ed. 1987. The World's Major Languages. New York: Oxford University Press.

---. 1983. Phrasal Verbs. Chap. 20 In The Grammar Book An ESL/EFL Teacher's Course. USA: Newburry House Publishers.

Davidson, D. 1980. The Logical Form of Action Sentences. In Essays on Actions and Events. New York: Oxford University Press.

Dorr, B. and Voss, C.R. 1993. Constraints on the Space of MT Divergences. Working Notes for AAAI Spring Symposium Series, Building Lexicons for Machine Translation. Technical Report SS-93-02. Stanford, CA.

Fillmore, Charles J. 1966. A Proposal Concerning English Prepositions. Report of the Seventeenth Annual Round Table Meeting on Linguistics and Language Studies:19-34.

---. 1968. The Case for Case. Chap. 1 In Universals in Linguistic Theory., ed. Emmon Bach and Robert T. Harms. New York: Holt, Rinehart and Winston, Inc.

Fromkin, Victoria, and Robert Rodman. 1988. An Introduction to Language. Fourth ed. Philadelphia: Harcourt Brace Jovanovich College Publishers.

Hobbs, Jerry. 1985. "Ontological Promiscuity," in Proceedings of the 23rd Annual Meeting of the Association for Computational Linguistics. Chicago.

Hutchins, J. W., and H. L. Somers. 1992. An Introduction to Machine Translation. San Diego: Academic Press, Inc. Harcourt Brace Jovanovich.

Japkowitz, Nathalie, and Janyce M. Wiebe. 1991. A System for Translating Locative Prepositions from English into French. 29th Annual Meeting of the Association for Computational Linguistics 29 (18-21 June):153-60.

Nilsen, Don Lee Fred. 1972. Toward a Semantic Specification of Deep Case. Janua Linguarum Series Minor, ed. C. H. Van Schooneveld, 152. The Hague: Mouton & Co. N.V.

---. 1973. The Instrumental Case in English. Janua Linguarum Series Minor, ed. C. H. Van Schooneveld, 156. The Hague: Mouton & Co. N.V.

Otero, María, "Latin America: ACCION Speaks Louder Than Words", UNESCO Courier,

January, 1997.

Somers, H. L. 1987. Valency and Case in Computational Linguistics. Edinburgh Information Technology Series, ed. S. Michaelson and Y. Wilks. Edinburgh: Edinburgh University Press.

Spang-Hanssen, E. 1963. Les prépositions incolores en français moderne. Copenhagen: G. E. C. Gads Forlag.

Stockwell, Robert P., Paul Schachter, and Barbara Hall Partee. 1973. The Major Syntactic Structures of English. New York: Holt, Rinehart and Winston.

Trujillo, A. 1992. Spatial Lexicalization in the Translation of Prepositional Phrases. in 30th ACL, Student Session. Newark, Delaware. 28 June - 2 July, 1992.

Van Langendonck, W. 1974. Internally Referring Prepositions and the Subcategorization of Space and Time Denominations in a Netherlandic Case Grammar. Leuvense Bijdragen 63.