The Role of Pred in Lfg+glue

In this paper, I argue that standard, co-descriptional glue semantics provides no clear and satisfactory role for the traditional PREDfeatures of LFG, due to the fact that the linear logic of glue semantics does the work of the Completeness and Coherence Constraints. But then I show that a reduced but significant role for PRED-features can be found in an alternative ‘Description-by-Analysis’ (DBA) formulation, proposed in Andrews (2007a). The DBA formulation is argued to be superior in various respects, and some constraints are proposed to cause the DBA approach to approximate some of the empirically justifiable aspects of the behavior of the co-descriptional formulation. The standard way to combine LFG with glue-semantics has been with a ‘co-descriptional’ architecture in which lexical entries introduce the usual grammatical features in the usual way, together with ‘meaning-constructors’ that account for the meanings, both of the PRED-feature associated with the lexical item, and any semantically intepretable grammatical features that it might introduce, either inherently or due to the inflectional morphology. Typical examples would be the following entries for the verb form went and the noun-form feet: (1) a. went :V, (↑PRED)= ‘Gomotion<(↑ SUBJ)>’, (↑TENSE)=PAST, λx.go(x) : (↑ SUBJ)e −◦ ↑p, λP.Past(P ) : ↑p −◦ ↑p b. feet :N, (↑PRED)= ‘Foot’, (↑NUM)=PL, λx.Foot(x) : ↑p, λP.Past(P ) : ↑p −◦ ↑p Co-description was introduced and motivated in Halvorsen and Kaplan (1988) as an alternative to the earlier (and overall more often used) ‘description-byanalysis’ (DBA) architecture, in which the f-structure is the primary input to the semantics. Although the norm in glue-semantics, co-description raises a puzzle with respect to the role of PRED-features, namely, why they are there at all. The problem is that, as pointed out in Kuhn (2001), the linear logic resource management employed in glue is in itself sufficient to account for the phenomena of Completeness, Coherence, and Predicate Uniqueness, which comprise the major special properties of PRED-features. This leaves us with no clear reason why these features couldn’t just be omitted from the lexical entries of (1). Even if absence of the PRED-features caused some And, independently developed for XLE (Crouch, p.c.), although no longer used. Using p ‘proposition’ for the type of propositions rather than the usual t, and a clearly oversimplified Priorian operator treatment for tense. See for example Halvorsen (1983), Wedekind and Kaplan (1993), Frank and Semecky (2004), Crouch and King (2006), Crouch (2006). subtle problems, putting them back in would still constitute an explanatory problem, since there isn’t any principle that requires LFG lexical entries to introduce PRED-values. If the benefits of co-description were sufficiently impressive, one could presumably deal with this issue, but I will first show that the original motivation for it is insufficient, and point out that it creates various problems, one of which was noted by Andrews (2007a). Then I will describe a DBA architecture for glue, and show it it provides a role for PRED-features. But this is not the same as in pre-glue LFG, since glue will be doing the work of Completeness and Coherence (but not Predicate Uniqueness). So the last step is to propose some constraints which will cause meaning-constructors in the DBA architecture to act in a way that is similar in certain empirically justifiable respects to standard PRED-features controlling Completeness and Coherence, but avoiding the problems with co-description. 1 Problems and Non-benefits of Co-Description The main proposed benefit of co-description was that it could make available for semantic interpretation information not present in f-structure (Halvorsen and Kaplan 1988:284, 1995 version). But this ignores the fact that, thanks to the inverse of the φ projection, anything accessible from c-structure is also accessible from f-structure. Andrews (2007b), for example, proposes constraints involving c-structure in a DBA glue framework. However, it might still be the case that co-description is the best approach, either for all, or only for some, kinds of linguistic phenomena. Here I will argue that it isn’t best for what would be traditionally regarded as the interpretation of features and lexical items (by contrast, co-description seems very well suited for the properties of information-structure, c.f. Mycock (2006)). Perhaps the most immediate problem, pointed out in Andrews (2007a), is that it becomes an accident that the occurrences of features and their traditionally ascribed meanings are quite closely correlated, with only limited exceptions, such as pluralia tantum, which I’ll discuss later. There would for example be nothing obviously wrong with a variant of (1b) in which the plural meaning-constructor was present but not the plural feature-equation. But this doesn’t happen, even with the exotic plurals that English is so fond of borrowing from other languages: (2) a. These seraphim are annoyed b. This seraph is annoyed c. *This seraphim is annoyed (plural meaning, singular syntax) “Every interpretation scheme based on description-by-analysis requires that all semantically relevant information be encoded in the functional structure.” But agreement, the main motivation for having features at all, leads to a further problem with the meaning-constructors. This is that one has to decide which of the various lexical entries introducing a given feature-value occurrence is the one that is introducing the constructor. Consider an Italian example such as: (3) (le the(FEM.PL) ragazze) girl(FEM.PL) vengono come(3.PL) The girls/they are coming If the subject is present, one would presumably want the noun to introduce the plural meaning-constructor, and the verb not to (since not all NPs are in positions where there is a verb to agree with them and provide their number constructors), but if the subject is omitted, then the verb would presumably be the provider of the constructor. It is certainly not impossible to come up with grammars that will work properly, but it involves delicate choices with considerable scope for stipulation, which it would be good to reduce to the greatest extent possible. Another problem resides in the overlapping powers and responsibilities of the PRED-features, with their argument-lists, and those of the meaningconstructors that refer to grammatical functions. This is that, although the PRED-features control what governable grammatical functions can and must appear, they no longer say anything about what their semantic contributions are, since this is done by the meaning-constructors. But, left unconstrained, meaning-constructors can do all sorts of peculiar things in the way of rearranging the semantics of the grammatical functions. Below, for example, (a) interchanges the semantic role of subject and object, while (b) creates an unspecified causee agent causative: (4) a. λPxy.P (y, x) : ((↑OBJ)e −◦ (↑ SUBJ)e −◦ ↑p)−◦ (↑ SUBJ)−◦ (↑OBJ)−◦ ↑p b. λPx.(∃z)(Cause(x, P (z, y))) : ((↑ SUBJ)e −◦ ↑p)−◦ (↑ SUBJ)e −◦ ↑p Without some further constraints, these meaning-constructors could be introduced by inflections or grammatical particles, thereby undoing the kinds of work people have been trying to accomplish with Lexical Mapping Theory and its competitors over the last several decades. The most obvious and direct solution to the overlap problem is to drop the PRED-features entirely, since, as noted above, the resource management provided by linear logic can do all of the syntactic work of the PRED-features, and of course the meaning constructors also take over their informal role of encoding the meaning. Therefore, the natural consequence of adopting codescription is to abandon PRED-features. This might of course be the right thing to do, but I will argue in the remainder of the paper that glue-byDBA would be a good thing to try first for certain aspects of semantic interpretation, especially, morphology and the lexicon. However, note that the use of meaning-constructors introduced by the PS rules, for example by Asudeh and Crouch (2002) and Sadler and Nordlinger (2008), is not implicated in any of the problems raised here, and is consistent with what I will be proposing.