Remembrance of Things Past... 1 Motivation for syntactic transformations In syntax, the American linguistic tradition prior to Chomsky had focussed on immediate constituent analysis of sentences (Chomsky 1957). This approach provides a parse tree of sentence constituents, grouping words into phrases and phrase types. Some examples from a popular work on linguistics are given in (1) (Pinker 1994: 102): (1) Surface structure ambiguity: a. Yoko Ono will talk about her husband John Lennon who was killed in an interview with Barbara Walters. b. We will sell gasoline to anyone in a glass container. In (1a), for example, the prepositional phrase in an interview with Barbara Walters could describe the place where Yoko Ono will talk about her husband or the place where he was killed, a difference which is clearly modelled in the two surface structures (2) and (3): (2) Surface tree for (1a): P PP PP talk P in an interview about her husband John Lennon who was killed (3) Alternative surface tree for (1a): 1
P PP talk P about her husband John Lennon who was killed PP in an interview The tree structure is an excellent model of many aspects of syntactic organization, but a leading linguistic idea within the tradition of generative grammar is that grammatical structure cannot be adequately characterized solely in terms of a single tree structure representing the immediate constituent analysis of overt constituents. A more abstract representation is needed to represent the implicit linguistic knowledge of speakers. This problem is illustrated by the ambiguity of (4): (4) Deep ambiguity: I saw the batter. (batter = subject or object of ) Speakers are aware that the batter may be interpreted as the logical subject of ( the ter of the ball ) or the logical object of ( by a badly pitched ball, for example). Yet there is no grouping of the overt elements into different constituents that can explain this difference of meaning. (5), a typical surface tree representation of (4), illustrates this point: 2
(5) Surface tree for (4): S P I S saw P the batter A more abstract syntactic representation of the grammatical relations is therefore needed to capture the deep ambiguity of (4). Chomsky has conceived of this abstract representation as a deep structure, an initial structure (or structures) which undergoes transformation into the overt perceptible form. Example (4) would thus have two deep structure sources, illustrated schematically in (6), which would account for the ambiguity: (6) a. I saw [ the batter (something]) ] b. I saw [ (something) the batter ] (6b) would undergo a syntactic transformation that causes it to resemble the surface form of (6a): (7) Passivization ( movement): I saw [ the } batter {{} (be) (by something) ] The idea behind the transformational analysis shown in (7) is that a word in one position is related to another in a nonadjacent position in the tree structure. As shown in (8) it seems that not one tree, but two trees are necessary to characterize the relation a deep, underlying, or initial tree structure and a surface, derived, or final tree structure. 3
(8) Underlying and transformationally derived trees for (4): S S P P the batter the batter e Just as deep grammatical ambiguity arises when different underlying structures correspond to the same surface structure, so grammatical equivalence can arise when two distinct surface structures correspond to the same underlying tree structure. This is illustrated by the phenomenon of clitic movement in French, shown in (9). (9a) and (9b) are equivalent in their grammatical relations in both the pronoun meaning us is the object of the verb meaning know : (9) Deep equivalence: Marie nous connaît. Marie us knows Marie knows us. Marie ne connaît que nous. Marie not knows but us Marie knows only us. The clitic in (9a) is assumed to be derived from the same underlying structure as (9b) by the transformation of clitic movement shown in (10): (10) Clitic movement: Marie nous }{{} connaît It is characteristic of the transformational approach that in both of these types of cases deep ambiguity and deep equivalence the underlying tree structure is assumed to correspond rigidly and invariantly to the semantic roles of the participants in the eventuality denoted by the verb. For example, the 4
logical object of the verb the one who is or known always appears in the postverbal direct object position of the deep structure tree, whether or not it is ultimately expressed on the surface preverbally as a passive subject or a cliticized pronoun. In effect, lexical roles are represented in the form of an underlying syntactic tree. This syntacticizing of lexical semantics is the core assumption that is abandoned in the newer family of nonderivational grammatical theories. 2 Imperfect-Correspondence Model How are such phenomena, which have long been taken to motivate transformational grammar, explained within the newer family of nonderivational grammatical theories? Let us take lfg (Kaplan and Bresnan 1982) as a concrete example. The most important idea is that lexical roles are directly mapped onto surface syntactic positions without the mediation of deep or underlying syntactic positions. This is possible because lexical roles are seen as having a distinct structure in their own right, independent of syntactic tree structure. The two types of structure are not rigidly related, but allow for a flexible correspondence. In lfg the correspondence between lexical roles and syntactic expressions is decomposed into two parts, as shown in (11): (11) Mapping in lfg: i. There is a flexible (imperfect) correspondence between lexical roles and grammatical relations. ii. There is a flexible (imperfect) correspondence between grammatical relations and overt syntactic tree structures. (11i) is illustrated in (12). For the active verb the subject is the ter, represented by the first argument role x. For the passive verb the subject is the one, represented by the second argument role y. 5
(12) Flexible lexical mapping of roles to relations: active: passive: < x y > < x y > roles subj (obj) (by obj) subj relations As shown in (13) both the active and the passive allow the non-subject argument to be omitted; hence the ambiguity: (13) Notation for lexical forms (optional arguments omitted): active: < subj x, 0 y >, passive: < 0 x, subj y > Note that in both active and passive lexical forms x = the ter and y = the thing, while the grammatical relations borne by these semantic roles vary lexically: y is the object in the active and the subject in the passive. Thus in example (14), the ambiguity follows from the lexical choice of an active or passive lexical form of the verb form : (14) I saw the batter. (the batter = subject of active or passive ) No transformational movements are necessary: the ( the batter ) always remains in its surface position of the complement verb and is given whichever lexical role (x = the ter or y = the thing) the verb lexical associates it with, constrained only by the completeness and coherence of the result. In sentence (14) both the active and the passive lexical choice make for complete and coherent analyses, given the omissibility of the object of the active verb and the (by object) of the passive verb. The syntactic functions of the phrases can therefore be read right off the surface structure tree: 6
(15) Direct mapping of surface syntactic positions into roles: S P I S saw Subject the batter P < subj x, 0 y > or < 0 x, subj y > This direct surface mapping is possible because there are no syntactic transformations; grammatical relations are never transformed one into another by any syntactic operation. In effect, a grammar is a direct correspondence mapping between a set of surface tree structures and a parallel set of lexical functional structures. The same surface structure tree can thus be directly mapped onto two different lexically induced feature structures, called functional structures or f-structures in lfg, as illustrated in (16) and (17). 7
(16) S P...... [ ] def +... subj pred batter pred < subj x, 0 y > x I S saw P the batter (17) S P...... [ ] def +... subj pred batter pred < 0 x, subj y > y I S saw P the batter In this way deep ambiguities can be explained without syntactic transformations. How are cases like clitic movement handled, where there is no lexical ambiguity? This is where assumption (11ii) comes in: there is a flexible correspondence between grammatical relations (such as Object) and tree structure positions as illustrated in (18) and (19): 8
(18) P ne connaît Object que nous (19) Object Cl nous connaît In French the Object relation may be expressed either postverbally in the verb phrase or preverbally as a clitic. The preverbal position is restricted to the category of clitics, while the postverbal position can be occupied by s of all types. In this way two different surface structure trees can be mapped onto the same lexically induced functional structure, as illustrated in (20) and (21). (20) P [ pred know < subjx, obj y > obj [ us ] ] ne connaît (21) Cl nous connaît que nous [ pred know < subjx, obj y > obj [ us ] ] Notice how this approach differs from the transformational conception of grammar, where the underlying tree structure is assumed to correspond rigidly and invariantly to the lexical semantic roles, and subsequently undergoes transformational derivation to achieve the variety of surface expressions actually 9
found. In the nonderivational, constraint-based approaches, in contrast, lexical roles map directly onto the variety of surface structures without transformations. Under the direct mapping, certain sets of surface expressions are grammatically equivalent. For example, the French preverbal clitic and postverbal pronoun are equivalent in that both correspond to the same lexical roles and both alternate with a subject pronoun under passivization. This equivalence is captured in lfg by classing both structures as members of the grammatical relation Object. In general, grammatical relations such as Object and Subject can be defined simply as equivalence classes of surface forms under the mappings (11i,ii). 3 Optimality Theory An obvious deficiency of the lfg approach to grammatical representation, based as it is on flexible correspondence mappings between parallel structures, is that there must be limits to flexibility. Not every grammatical relation can correspond to every lexical role, and not any tree position can correspond to any grammatical relation. And even if one can point to one or another farflung language that exhibits an unusual correspondence mapping, there are still strong generalizations in the form of crosslinguistic asymmetries. For example, there are a number of languages like English which have free pronouns and no clitic pronouns, and there are many languages like French which have both free and clitic pronouns, but there are few if any languages which have only clitic pronouns and no free pronouns whatsoever (Bresnan 1999). This is a classical markedness asymmetry of the sort observed by members of the Prague School (e.g. Jakobson 1984) and further documented in the domain of syntax by linguists of a typological and/or functionalist bent (e.g. Haiman 1985, Givón 1995). Greenberg (1966) in particular has studied these markedness relations in terms of inplicational universals and frequency counts. In early transformational theory the deep or underlying structure is the basic, unmarked structure, and complexity is added by transformational rules. Unfortunately, the lack of transparency in that framework between deep structures and overt perceptible constructions soon became so great that there has been little fruitful work attempting to explain markedness asymmetries in transformational terms. Greenbergian implicational universals, because they concern only surface structures, have attracted little interest within the transformational conception of Universal Grammar. 10
How can such typological asymmetries be captured in nonderivational theories like lfg in which both the clitic pronoun and the free pronfoun are surface forms equally basic in terms of generation (18) and (19)? We can solve this problem by viewing a lfg as a universal generator of candidate linguistic structures which are evaluated as to their markedness along multiple conflicting dimensions. A grammar is not then just a structure generator, as in the classical conception of generative grammar, but a means of selecting structures which optimize the harmony among multiple conflicting dimensions of markedness. It is this optimization which defines the limits of flexibility in the correspondence mappings between linguistic content (represented by functional structures) and linguistic form (represented by surface constituent tree structures). 11