ANOTHER LOOK AT NOMINAL COMPOUNDS
Pierre Isabelle D~partement de linguistique Universit~ de Montreal C.P. 6128, Succ. A, Montreal, Qua., Canada H)C )37
ABSTRACT semantics. And since noun meaning appears to be closely connected with knowledge of the world, one is led to explore the modes of interaction We present a progress report on our research between linguistic and conceptual knowledge. on nominal compounds (NC's). Recent approaches to this probiem in linguistics and natural ianguage From an NLP perspective, NC's have turned processing (NLP) are reviewed and criticized. We out to be an important stumbling block for sys- argue that the notion of "roie nominal", which is tems that attempt to deaI with real-iife text, at the interface of linguistic and extraiinguis- especiaIly in technical domains, for purposes tic knowledge, is crucial for characterizing NC'e such as machine translation (IsabeIle, to as weII as other Iinguistic phenomena. We examine appear), information retrieval, etc. a number of constraints on the semantic interpre- tation ruies for NC's. Proposals are made that Our ultimate goal is to develop an NLP sys- shouid improve the capability of NLP systems to tem capable of analyzing large classes of NC's in deaI with NC's. the sublanguage (Kittredge and Lehrberger, i982) of aircraft maintenance manuals. We do not aim at solving all cases, since we believe the problem to be exceedingly difficult. At the present stage I INTRODUCTION of our inquiry, we concentrate on the design of s suitable theoretical framework. A. Problem Statement In section II, we present a brief review of As a first approximation, we define a previous work in linguistics and NLP. In sections "nominal compound (NC) as s string of two or more III and IV, we examine two aspects of the seman- nouns having the same distribution as a singie tics of nouns that are crucially relevant to the noun, as in example (I): analysis of NC's: predicative nouns and role nominals. Finally, in section V, we explore pos- (1) aircraft bomb bay door actuating cylinders sible constraints on the semantic interpretation of NC's. We will see below that provisions have to be made in some cases for intervening adjectives.
NC's can exhibit various degrees of lexica- II BACKGROUND lization, but we will focus our attention on productive rules for forming novel compounds. As to their surface syntax, NC's can assume any A. Approaches in Linguistics structure generated by the rule N --> N N; accor- dingly, their structural ambiguity grows exponen- The early study of Lees (1963) classified tially with their length (following the "Catalan NC's on purely grammatical criteria, and it sequence"). How, then, do we determine that the failed to provide constraints that could explain normal interpretation for (i) imposes the bracke- how NC's are semantically interpreted. ting shown in (2), rather than any of the other 41 syntactically possible bracketings? In s number of more recent studies, such as Levi (1978)9 there has been an attempt to view (2) ((aircraft ((bomb bay) door)) NC's as governed by tight semantic constraints. (actuating cylinder)) Thus, according to Levi, any novel NC realizes a pattern where either: a) the head noun is a deverbal naminalizstion and its modifier is interpreted as an argument of the related verb; B. Goal__ssof th___ee study or b) the two nouns are related by one of exactly nine deletable predicates ("is the cause of", "is We believe that the analysis of NC's repre- for", etc.). sents an important and largely unsolved problem. From a theoretical point of view this problem This reductionist attempt has been criti- raises the question of how to deal with noun cized, most notably by Downing (1977), on the
509 grounds that the interpretation of NC's crucially (3) a. measure nouns involves pragmatic knowledge, and that numerous map: objects onto quantities cases of NC's (such as thalidomide parents) will examples: speed (of), temperature (of), resist any analysis in terms of a closed set of volume (of), size (of) relations. These criticisms have led theorists like Selkirk (1982) to adopt the position that b. "area" nouns only "verbal compounds" (those constructed on a map: objects onto subparts pattern "argument + nominalization") are amenable examples: top (of), side (of), bottom to linguistic characterization; all other NC's (of), center (of), core (of) would have to be explained in separate extralin- guistic theories. c. collective nouns map: individuals onto sets examples: group (of), set (of) C. Approaches in NLP d. representational nouns Several systems have been developed in an map: objects onto representations of ob- NLP framework to deal with the problem of inter- jects preting NC's; two recent exempIes are reported in examples: picture (of), diagram (of), Finin (1980) and McDonaId (i982). sense (of)
In both systems, the individual nouns of an e. other examples NC are first mapped onto conceptual representa- location (of), goal (of), tions where concepts are characterized by a set brother (of), king (of) of "roles" or "slots", and are arranged in an abstraction hierarchy. Interpreting a compound Most if not all of these argument-taking nouns then amounts to forming a derived concept on the can have their argument satisfied by a modifier basis of the constituent concepts; in most cases, noun in an NC: this is done by interpreting one of the concepts as a slot filler for the other. For example, the (a) a. oil temperature interpretation of steel adapter would involve b. box top inserting the concept associated with steel into c. tank group a RAW-MATERIAL slot within the representation of d. circuit diagram adapter. But the authors do not examine in any e. component location detail the question of eventual constraints on this interpretation process, such as the effect A treatment in terms of predicate/argument pat- of word order. terns for this type of NC seems far superior to Levi's (1978) use of the semantically empty Another crucial issue which has not been "deletable predicate" HAVE ("the component HAS a explored in sufficient detail, is the nature of, location"). Although these NC's are excluded from and the justification for the particular set of Selkirk's (1982) class of verbal compounds, they slots which is assigned to a given concept. Finin are amenable to the same type of semantic des- is somewhat more explicit on this question. Some cription. nouns have slots which represent standard case roles. Role nominals, on the other hand, are nouns which refer to a particular case role of B. Action and State Nominalizations another concept; for example, food refers to the object role of eat, and this fact provides the Most studies on NC's have recognized the key to the interpretation of cat food. This fact that deverbal nominalizations exhibit a notion will be examined in detail below. But semantic behavior closely related to that of the Finin also resorts to other types of slots (e.g. verb, and subcategorize elements that can occur "raw-material") which are not discussed. as modifier nouns in NC's. As mentioned above, these are in fact the only cases that Selkirk (1982) deems characterizable at the level of linguistic competence.
III PREDICATIVE NOUNS 8ut there seems to be no reason why deadjec- tival nominalizations should be handled in a different way. In examples (5) and (6), an action A. Root Nouns with Arguments and a state are nominalized, with the argument occurring either in a prepositional phrase or as The fact that several classes of non-derived a modifier in an NC: nouns strictly subcategorize phrases which are semantically interpreted as arguments has (5) a. Someone removes the pump. received very little attention in the literature. b. removal of the pump This phenomenon would deserve an extensive study, c. pump removal and we can only give here some relevant examples, together with an indication of the semantic cate- (6) a. Uranium is scarce. gories between which the relation expressed by b. scarcity of uranium the noun effects a mapping: c. uranium scarcity
510 We are not claiming that there is exact A related claim underlies Zholkhovskij & synonymy betweenthe (b) and (c) examples, but Hel'cuk's (1971) use of certain "lexical func- only that they exhibit the same predicate/ tions": argument pattern. Action nominals can take various types of arguments in NC's, and sometimes (15) a. Sl(bUy) = buyer several of them simultaneously: b. S3(buy) = seller c. Sloc(battle) = battlefield (7) a. pump failure (subject) d. Sinst(See) = eyes b. Montreal flight (source, goal) (8) a. poppet chatter tendency where Si, $1o c and Sinst are defined as functions t~t yield the typical name of, respectively, the i- arrant, the location a-nd the instrument.
IV ROLE NOMINALS Fillmore (1971) also makes a comparable proposal, when he suggests that the lexical entry of knife should include (16) as a component: A. Nominalizations (16) use: I of
ill) 18H employee Possibility (a) seems to describe correctly = one who is employed by IBH numerous cases of deverbal nouns. For example, it seems clear that anyone who is employed is an (12) pump assembly employee. However, with agents and instruments, = the result of assembling a pump there is a tendency to reserve the role nominal for habitual fillers: one hesitates to apply the With agent and instrument nominals (9, i0), there term writer to a person who only wrote a letter. is a strong tendency to assign an "habitual" Horeover, with this definition, knife would not aspect to the underlying verb and generic refe- be a role nominal for cut, since it is perfectly rence to its object; this kind of interpretation possible to cut bread with a sword. The notion is awkward when the argument appears in a PP: would then loose much of its power, since we need it to explain why bread knife is interpreted as (I)) ?a driver of trucks "a knife used as an instrument for cutting bread".
B. Root Nouns On the other hand, definition (c) seems too weak: even if a sword can be used to cut bread, The term "role nominal" is due to Finin bread sword is odd in a normal context (contex- (1980) who uses it to cover not only nominaliza- tual factors will be discussed below). Thus, tions of the type described above, but also any definition (b) seems the most appropriate. noun which can be semantically interpreted as referring to a role of a given verb, whether or not this verb happens to be morphologically D. 3uatifyin9 Role Assignments related. This claims amounts to saying that, semantically, we have relations such as the For those role nominals where there is no following: morphological evidence of relatedness with the underlying verb, one is forced to rely mostly on (14) a. pilot:fly :: driver:drive intuitions. However, the risk of arbitrariness b. gun:shoot :: lubricant:lubricate can be reduced by looking for further evidence c. food:eat :: employee:employ from other linguistic phenomena.
511 i. Evaluative Adjectives example, if that same Phil uses different types of trombones for men and women, we might speak of When Fillmore (1971) introduced his notion his women trombones, to mean trombones of the of role nominal, he was attempting to characte- type that Phil sticks into the nose of female rize the behavior of evaluative adjectives, not bypaasers. the behavior of NC's. He noted that a ~ood ~, where X is a role nominal, means: But C & C claim that, more frequently, deno- minal verbs are based on ~eneric knowledge about concrete objects, knowIedge which is accessibIe a. if X is an agent: one who performs the to aii speakers in a Iinguistic community. This associated activity skilfully claim is to be linked with Downing's (i977) (s good driver, a good pianist); remark that aithough NC's are sometimes "deic- tic", they are most often based on generic or b. if X is an instrument: a thing which permanent reIationships. permits the associated activity to be performed easily Obviously, the relevant knowledge (whether (a good knife, a good broom); particular or generic) is at least as much about the world as about language. In fact, it is clear c° In other cases, it seems that the resul- that both types condition each other. For exam- tant meaning is less predictible ple, objects that are used as vehicles will tend (good food has certain properties concer- to be verbalized on the syntactic pattern of ning nutritiousness and taste; a good movement verbs; and in the absence of other evi- house is comfortable, built to last, dence, one is likely to infer from its syntax etc.). that The fairy pumpkined the kids to Narnia allu- des not to an "satin 9 pumpkin" but to a "trans- As far as we can tell, the evaluation domain portation pumpkin". for agents and instruments is precisely the acti- vity which is relevant for the understanding of In order to predict the range of meaning of NC'B. Thus while good driver evaluates the dri- large classes of denominal verbs, C & C propose ving, car driver specifies its object; moreover, to encode some generic knowledge in the lexicon, in ~ood car driver, car falls within the evalua- by means of "predominant features" such as: tion domain: car drivers and truck drivers are evaluated on different scales. Evaluative adjec- (17) a. x is the agent of Act (to pilot y) tives can thus be used as a further source of b. x is the instrument of Act (to pump y) evidence in the description of role nominals. c. x is the result of Act (to group y) d. x is the location of y (to can y) e. x is the locatum of y (to cover y) 2. Denominal Verbs f. x is the time of Act (to weekend in y)
Another phenomenon which is relevant to the It is quite apparent that these features are question of role nominals and NC's is the crea- meant to capture the same type of facts as role tion of denominal verbs. Clark & Clark (1979) nominsIs. It is easy to find NC's which paraIiel examine this very productive process, in which a each ciass of verb singled out by them: verb formed by zero-affixation is understood "in such a way that the parent noun denotes one role (17') a. aircraft piiot in the situation, and the remaining surface argu- b. oii pump ments of the denominal verb denote other roles in c. tank group the situation" (p. 787). For example, Max sub- d. oil can wayed downtown means that Max went downtown on e. pump cover a subway. Intuitively, it appears obvious that f. Montreal weekend the knowledge involved in this interpretation process is very closely related to whatever per- We believe that the semantic mechanisms that mits interpreting the NC downtown subway as "the are at work in the interpretation of NC's, deno- subway that goes downtown". minal verbs and evaluative adjectives are basi- cally similar. By using independent evidence from An important aspect of C & C's work is to these phenomena, and an adequate generalization show that the formation of denominal verbs is of the notion of role nominal, one can go s long heavily dependent on contextuai knowIedge. Thus way toward uncovering the relevant semantic if you and me both know that Phii has iong had mechanisms. the crazy habit of sticking trombones into the nose of bypassers, I can inform you that PhiI has The notion of role nominal, as we understand just tromboned a poiice officer. Notice that in it, is at the interface between linguistic and the same context, 9opd trombone wouId presumably extralinguistic knowledge. Nouns may have severaI mean a trombone that is easy to stick into some- different roies, depending on the contingencies one's nose. of the entities that they denote; in those cases context usuaiiy makes one roie more saiient. In In such cases, the interpretation is based fact, context may even impart a noun with an on particular, situational knowledge. NC's can unusuai roie, as we have seen. However, we also be based on this type of knowledge. For beiieve that in ordinary texts, such as technieai
512 manuals, NC's that are analyzable in terms of such that P(x)". We do not think that diacritic role nominals are most often based on the usual, markers such as "typical function" are required; generic roles of the nouns. But this can only be rather, notions such as typicality, should be a shown through a large scale description of the consequence of the semantic rules which interpret relevant NC's. lexical entries.
We give below a few examples of the type of lexical specification for role nominals with E. A Tentative Scheme which we want to experiment. At this stage of our work this should be taken as nothing more than a Since the knowledge we have to encode is first approximation. The material enclosed in tied to world knowledge, there is a risk that it brackets represents selectional restrictions. could become overwhelmingly compiex. However, we will minimize this risk, by limiting the scope of (18) pilot our inquiry to a suitably restricted subianguage. x such that FLY(x,y,p) Technical manuals are interesting in this respect
A role nominal will contain within its lexi- cal entry one or more statements of the form "x
513 A. Relative Order of the two Nouns min9 unacceptable? In the latter case, it cannot be, as suggestedby Selkirk (1982) that subject I. Predicative Nouns arguments are prohibited in general: pump failure is fine. It may be that subjects are ruled out Selkirk (1982) claims that the modifier noun for -ing nominals, unless they express a result can satisfy an argument of the headnoun but not (cf. consumer spending). vice-versa. Finin (1980) claims that argument satisfaction (or slot filling) is possible in But on the whole, the pattern where the both directions. Finally, McDonald (1982) takes predicate comes first is much more constrained, the intermediate position that satisfaction of since it permits oniy action nominaIs, and pro- the head by a modifier is much more frequent. duces a semantic resuit very simiiar to =oie Now, consider the following pairs of exam- nominais. Thus, a cooIing device and cooier ples: denote very simiIar entities; the same is true of jacking system and jack. Notice that in exampies (25) a. oil temperature such as: b. ??temperature oil ()3) air temperature monitoring system (26) a. uranium scarcity b. ??scarcity uranium monitorin 9 system forms a constituent, even if air temperature is understood as the object of (27) a. pump removal of the monitoring; this is confirmed by the fact b. ??removal pump that (34) receives a similar interpretation:
(28) a. student invention (34) monitoring system for air temperature b. ??invention student It seems that this modification pattern has the It is clear that the (b) examples, if interpreta- effect of creating a role nominal out of its two ble st all, cannot easiIy realize the pattern constituents. found in the (a) examples. However, there is a very productive process which permits an action nominal to occur to the left of its argument; this pattern is most productive with inanimate 2. Role Nominals headnouns: Here again, the order of the nouns strongly (29) a. repair man conditions the resulting interpretation: b. ?man repair (35) a. truck driver (30) a. cooling device b. ??driver truck b. device cooling (36) a. oil pump (3l) a. bleed valve b. pump oil b. ??valve bleed c. valve bleeding (37) a. pump case b. ?case pump (32) a. jacking point b. ??point jacking (38) a. equipment bay b. bay equipment In the first three (a) examples, the headnoun is interpreted as a subject argument (agent or ins- In the (a) examples, the modifier noun is inter- trument). In (28a), point is interpreted as a preted as the object of the underlying verb iocation (where one jacks something). In all of (drive, transfer, detect, hold); here too, a the (a) examples, the action nominai is inter- permanent connection is established. A truck preted as denoting a permanent roie or function driver is a person whose (social) function is to of the headnoun. drive trucks, and that person is still a truck driver when he/she drives a scooter. If you use When order is reversed, meaning and, even- an oil pump to pump your tomato juice, it still tually, acceptability are affected. In the (b) qualifies to be called an oil pump. examples, the action nominal is not interpreted as expressing a permanent function of the other But when the nouns are inverted, there is a noun. Thus, if one can manage subject interpreta- change in meaning and, possibly, in acceptabili- tions in (29b) and (31c) -- "a repair done by men ty. It is hard to find a sensible interpretation (not robots)" and "a valve which bleeds some- for driver truck. The most natural interpretation thing" --, it will not follow that the men are for pump oil is "oil used for lubricating pumps": repair men or that the valve is a bleed valve. in this case, oil has become the relevant role nominal. There is also a possibility for an The pattern "argument + predicative noun" interpretation such as "oil in the pump" or "oil has some pecularities of its own; for example, coming from the pump". We are not sure how that why is (32b) unacceptable no matter the role type of interpretation should be produced. But it assignment that one makes? Or why is girl swim- seems clear that oil can only become an actor in
514 a permanent function of the pump when pump is the interpreted as no more than a location for the headnoun. shelf. However, in (44b), when the scope of righ t includes shelf, cockpit becomes nonreferential; It is easy to see that the other examples cockpit shelf then denotes a type of shelf. These exhibit s similar behavior. We therefore conclude facts seem to indicate that this scoping of that role nominals tend to establish permanent, right, in virtue of syntactic constraints, forces functional connections with their modifiers only. one to take cockpit shelf as something of cate- When they modify a headnoun, either this noun gory N; and when we do so, shelf becomes a role will absorb them itself (if it is a predicative nominal whose argument receives the type of "per- or role nominal) or else a looser connection such manent function" interpretation discussed above. as a locative one will be created. Core NC'B, those which establish argument connections with a predicative noun or role nominal, appear to form a tightly bound unit: B. Multiple Modification (45) a. ??oil high temperature i. Argument Ordering b. ??truck good driver c. ??equipment left bay When a predicative noun or role nominal in head position is modified by more than one argu- Some adjectives, the "nonpredicating" adjectives ment, it seems that ordering (39) will apply, as of Levi (1978) can appear within core NC's. But illustrated in the examples below= they are in fact nouns in disguise, and they receive an argument interpretation; in (46), it ()9) time < paths < subject < object is understood that one repairs a structure. loc (46) aircraft structural repairs (40) s. computer fuel testing b. ??fuel computer testing Thus, if we are willing to make a distinc- tion between syntactic compounding (similar to (41) a. Montreal jet flights genitive phrases) and core NC's, it can be b. ??jet Montreal flights claimed that true compounds form anaphoric islands and cannot be separated by adjectives. (42) a. evening Montreal trains b. ??Montreal evening trains Finally, there are certain types of nominal modifiers (for which we have as yet no analysis to offer) whose degree of cohesion with the 2. A Broader Perspective headnoun is somehow intermediate between nouns interpreted as arguments and nouns in separate So far, we have concentrated on predicative (pseudo-genitive) NP's. For example, nouns which nouns and role nominals, thereby excluding stand in a "material of" relation to the head various types of compounds. When we look at the cannot precede adjectives, but must precede argu- broader picture, different problems are raised. ments= First, (in technical manuals) we find examples with intervening adjectives as in (43): (43) a. large steel tank b. ??steel large tank (43) (Remove) front cockpit right shelf aft con- trol panel. (44) a. steel fuel tank b. ??fuel steel tank While the formation of NC's such as those we have examined in previous sections is usually The constraints that we have seen, and no considered to be a lexical process, there is some doubt others yet to be discovered, are certainly evidence that, in examples like (43), syntactic significant from a theoretical point of view; but processes are at work. Notice that each one of they also have practical value for NLP systems the three main groups in (43) is referential= which have to deal with NC's. "the (...) panel of the (...) shelf of the (...) cockpit", much as if there was an implicit geni- tive marker. But genuine compounds (that is, those of syntactic category N) are usually consi- dered to be anaphoric islands (Levi, 1978). One VI CONCLUSIONS possible explanation is that (43) should in fact be analyzed as containing three NP's. We have discussed two classes of nouns that A comparison of the following examples pro- have particular importance in NC formation= pre- vides some support for that view= dicative nouns and role nominals. We have shown that the latter class is also relevant to the (44) a. Remove front cockpit right shelf. description of evaluative adjectives and denomi- b. Remove right cockpit shelf. nal verbs. A tentative framework for the descrip- tion of role nominals has then been proposed. In (44a), cockpit has definite reference and is Finally, we have seen that rules which interpret
515 NC's obey e number of constraints. A distinction Selkirk E., The Syntax of Words, MIT Press, 1982. has been made between "pseudo-genitive" construc- tions and core NC's, the latter forming a tightly Zholkovskij A., Mel'cuk I., SUr la synthbse sd- bound unit with internal ordering constraints. mantique, in T.A. Informations, 11:2, 1971.
Much work remains to be done on the issues that we have discussed here; it is likely that experimentation with actual descriptions on a larger scale will lead to several refinements and revisions. Moreover, as has been pointed out in the last section, our description still ignores a number of types of compounding, and further dif- ficulties are to be expected. Nonetheless, we are confident that our work will, in the near future, result in a small NLP system capable of analyzing a broad range of NC's.
VII REFERENCES
Aronoff M., Contextuals, in Language, 56:4, 744- 758, 1980.
Clark E., Clark H. When Nouns Surface as Verbs, in Language, 55:4,, 767-811, 1979.
Downing P., On the Creation and Use of English Compound Nouns, in Lanquaqe, 5):4, 810-842, 1977.
Fillmore C., Types of Lexicsl Information, in D. Steinberg and L. 3akobovits (eds.), Semantics, Cambridge Univ. Press, 197I.
Finin T., The Semantic Interpretation of Compound Nominals, Coordinated Science Lab., Univ. of IIIinois, 1980.
Hobbs 3., Sublanguage and Knowledge , paper pre- sented at the NYU Conference on Subianguages, 3anuary 1984.
Isabelle P., Machine Translation at the TAUM Group, paper presented at the Lugano Tutorial on Machine TraneIation, April i984, to appear.
3ackendoff R., Semantics and Cognition , MIT Press, 198).
Kittredge R., Lehrberger 3., Sublanguage: Studies of Language in Restricted Semantic Domains, De Gruyter, 1982.
Lees R., The Grammar of English Nominalizations, Mouton, i96).
Levi 3., Th__ee Syntax and Semantics of CompIex Nominals, Academic, 1978.
McDonald D.B., Understandin 9 Noun Compounds, Carnegie-Meiion University, i982.
Moore R., Problems in Logical Form, Tech. note 241, SRI, 1981.
516