Recently, work in computational semantics and lexical semantics has made an interesting shift. Motivated by a concern for lexical organization and global coherence in the structure of the language lexicon, some researchers have moved towards more expressive semantic descriptions ([16, 1, 5, 10]), as well as more powerful meth­ods of composition ([22, 3] ).

Some, however, have expressed reservations as to the general applicability of type-changing op­erations such as coercion, as well as the notion of a generative lexicon itself ([7]). In this paper, we address these criticisms directly, and show that, upon closer examination of the data, those cri­tiques cither miss the point or are not substan­tiated by the data. Still, without a proper no­tion of constraints on coercion, there can indeed be overgeneration of forms and interpretations in the semantics, and in fact, the notion of con­ditions on coercion has always been integral to the basic spirit of Generative Lexicon Theory (cf. [19]). The empirical study of the range and lim­its of type change and cocomposition operations in natural language is an essential part of the re­search in formal semantics. The advantages ac­companying generative mechanisms and the char­acterization of languages as polymorphic in welldefined ways far outweight the explanatory in­adequacies inherent in traditional approaches to lexical design and semantic' projection, what [22] have called word sense enumeration approaches.

We will assume some general familiarity with the framework of generative lexicon theory, as out­lined in [16, 18, 1]. We feel it is important, how­ever, to clarify the motivating principles and gen­eral methodology behind such work, since these points seem to bo overlooked or misunderstood by some authors ([7]).

In order to help characterize the generative power of natural languages in terms of seman­tic expressiveness, it is natural to think in terms of semantic systems with increasing functional power. Furthermore, a natural way of capturing this might be in terms of the type system which the grammar refers to for its interpretation. It has been argued elsewhere ([19, 20]), that there are reasons for describing how semantic systems fall on a hierarchy of increasing expressive power, ft seems clear at this point that the current enu-merative techniques for lexical description are too impoverished to adequately describe the richness of semantic data, much less to explain either how word senses relate to one another or the creative use of words in novel contexts ([16]).

liriefly, a generative lexicon can be character­ized as a system involving at least the following lour levels of representations: Argument Struc­ture, Event Structure, Qu alia Structure, and Lex­ical Inheritance Structure. A set of generative do-vices connects these four levels, providing for the compositional interpretation of words in context. The exact nature of these devices will determine the polymorphic expressiveness of the semantics in fairly definite ways. The best studied illustra­tion of this is the phenomenon, of type coercion, but it is by no means the only one.

As argued in Section 2 above, the well-formedness of object complement coercion with aspectual predicates such as begin is conditioned by the event sort of the qualia associated with the NP itself. Tims, only NPs having associated transi­tion, events will allow coercion and control. This is not to say, however, that begin selects only for transition events. There are, of course, perfectly grammatical examples of process complements, as shown in (8) below:

(8) a. The snow began to fall at midnight.

b. John began to feel ill.

c. The war began to reach into Bosnia.

These examples illustrate the use of begin as a raising verb. We will follow Perlmutter [Id], in distinguishing between two senses of the verb be­gin, distinguishable not by the selectional proper-tics given in Godard and J ayez, but rather, con­forming to the distinction that. [11] made; namely, as either a Raising or a Control verb.

The analysis is as follows. There are indeed two grammatical expressions of the verb begin, as Raising and Subject-Control forms: As a control verb, the event sort specified as the complement is a transition. As a. Raising verb, however, the event may be any sort. This follows the typing assignments below:

The examples above and in (9a) and (9b) below illustrate the raising interpretation of begin:

(9) a. The acid began to corrode the marble, b. It began to rain.

We will assume that raising is accomplished by function composition, in the manner of [8]. The manner in which liaising is treated as func­tion composition (KC) is as follows: begin is ea --> c' , to corrode the. marble in the exam­ple above is e —> r'\ Then, FC(begin,V P) — \P\begin! [corr otle(P, the-marble)].

As pointed out in [?], VP ellipsis can be used as a. diagnostic for determining whether a comple­ment is part of a. raising or control construction. Some predicates permit both a control and non-control reading, such as (10) below, where John may be intentionally dieting or lie may be ill.

(10) John began to lose weight.

Notice however, that in English the sentence in (11) has only the intentional inchoative reading, and not the raising version.

(11) John began to lose weight, and Mary began too.

What this indicates is that there are indeed two constructions at play here, as teased apart by cer­tain diagnostics, further evidence comes from imperative structures (12) and /oree-complement constructions, which require the control sense of the verb.

(12) a. *llegin looking for a job, you lousy bum! b. Start looking for a job, you lousy bum!

These data, indicate that begin, in the control in­terpretation, strongly prefers a telic (transition) event complement.

We have argued that there are two senses of the verb begin, corresponding to raising and control predicates. These senses, however, are not arbi­trary types but are logically related to one an­other in the same way that the different senses of unac.cusa.tive/ca.usative verbs, such as break, and .sink are related. In [23] it is shown that verbs such as stink and aj'fondarc are logically poly se-mous in predictable ways, and don't need to be assigned multiple lexical entries. The same gen­eralization holds for verb such as begin: begin is the lexical version of an unaccusative marker, but for propositions rather than for entities.

In [22] and [23], a general mechanism is de­fined which makes the appropriate type available for a coercion operation. As discussed in [16],

iircommRîim' a m'e (e iure: e—*eJ

(7 h Marie commencer à lire le livre G F Marie commencer le livre : t the qualia, can be seen as partial functions, re­turning the value of a particular quale for an NP. The combined set of qualia provide a set of type aliases for the expression containing them (cf. [1.8]). One particular mechanism, type pumping, has been explored as a. means to generating the alias set ([20]).

Let (J be the typing judgements with respect to a grammar. Then, by convention, G h a : r represents a type assignment of r to the expres­sion a.' Thus for example, the type available to an expression a with quale Q; of type r, can be seen as the following type inference:

This says that, given an expression a of type o, there is a coercion possible between a and r, which changes the type of a in this composition, from o to t. We will illustrate the further appli­cation of this coercion operation below, as used in the begin examples. In (13a), we see how the as­pectual verb commencer selects the complement VP, and how in (13b), an NP is coerced into an event interpretation.

(13) a. Marie a commencé à lire le livre. (VP) b. Marie a commencé le livre, (np)

Following [20], we can view the basic composi­tion of the sentence in (13a) a.s type inference in Figure 1.

'See [6] for explanation of formal mechanisms of type inference within the A-calculus, and [10, 2], and [22, 19] for its application to lexical representation.

For the derivation of (13b), coercion applies to the complement NP, resulting in the appropriate type selected by the verb, as illustrated below in Figure 2.'*

In the case of begin with NP complements such as the. symphony or the motorway, the coercion is not possible, given the type mismatch in the in­tended qualia relation (i.e. listening and driving arc process events). Notice, however, since the agbnt1vb for each has an event of sort transi­tion, these are possible coercive interpretations; i.e. perform the symphony, or build the highway.

Notice that one might expect there to be rais­ing constructions involving coerced NP comple­ments, lint these do not exist, as the ungram-maticality of *.John began his nap (non-control reading) illustrates. This is due to the fact that coercion is governed by the type of the controller, in this case of type e" -, e'. 'Phis coercion will be successful if such a type exists in the alias set of the complement. Since function composition is an operation at the level of t he VP, there is no point in the derivation such that the appropriate type is available for the rule to apply.

As a final observation, it should be obvious now why verbs such as enjoy allow a much broader range of complement coercions (cf. [16] for de­tails). They are typed for taking an event of any sort, thereby allowing the process events of the tei,ic roles in enjoy the symphony / the movie.

•'We ignore for now the type distinction between in­dividuals, e, and generalized quantifiers, «e,t>,t>. In the full version of the paper, we show the type shift taking this distinction into account.

Wc have attempted to respond to specific crit­icisms regarding coercion operations in the se­mantic interpretation of natural languages. The problems pointed out by Godard and J aye/, do illustrate that conditions on coercion are a. nec­essary part of the semantics, but as we demon­strated, these are already an integral component of Generative Lexicon Theory. In the process of this discussion, we have reiterated the advan­tages of a generative lexicon in the context of the larger theoretical and methodological issues. More specifically, we showed how begin and com­mencer exhibit both raising and control behavior, and that this is an instance of the larger alterna­tion class between causative and inchoative verbs, itself an example of logical polysemy.