Grammar framework features that are not supported cross-linguistically

Question

There are quite a lot of grammar frameworks postulated since the last century, like MP, LFG, RRG, RCG, MTT to name a few. I like reading about languages, but a lot of publications about languages are couched along such grammar frameworks, and for someone like me who has no academic background in Linguistics and has not read a lot of linguistics literature, its difficult to evaluate which grammar framework has more explanatory power, or which of their features are robust cross-linguistically.

It's probably contentious to ask which is the best, so what I would ask instead is, if we could come up with one particular feature, characteristic or attribute in a specific grammar framework that is invalidated cross-linguistically and which grammar framework addresses it best. It would be ideal if in the reply we could specify which languages and language features are counter-evidences of which grammar framework features. This is to avoid closing this question for not being constructive (soliciting "debate, arguments, polling, or extended discussion"). I'm thinking along the lines of responses like this and this.

As much as possible, please limit each answer to one feature per grammar framework, unless a set of grammar frameworks share this feature.

Also, if its possible, can this be made a community wiki?

As an example, this excellent discussion about binary branching of all syntactic structure in MP is a very good example and is the one that made me think of this question.

---

EDIT: In light of @prash's comment, what I am looking for are falsifiable "predictions" that flow out of these grammar frameworks/formalisms, in a Popperian sense. What cross-linguistic data could prove a formalism's feature to be true or false as against other formalism? Instead of looking for data that support such formalism, one should look for data that falsifies such. In the binary branching of all syntactic structure discussion, MP posits a binary nature of syntactic structures which Tim Osborne show is untrue. The weight always goes to nonconforming data.

Even if all formalism can accommodate all cross-linguistic data, applying Occam's razor can tell us which formalisms have the same explanatory power in spite of fewer assumptions.

That by itself is a count against some formalisms. Actually, I don't know much about these things, I was just following the discussions and very much interested in knowing other features which fall under the same category.

Answer 1

I agree with the premise of the question that the first task of a scientist is to devise a framework that makes potentially false predictions, and thus that looking for false predictions is good scientific practice. It is important to note though, that this game is better played towards one's own favored scientific theory, if only because this game is embarrassingly easy to play already for the most advanced of the exact sciences (just ask a physicist what dark matter and dark energy, which are supposed to constitute some 96% of the Universe, are supposed to be) and so becomes almost inane for a field in its infancy, like linguistics (anyone with a theoretical framework in linguistics should be able to produce on demand ten phenomena the framework cannot explain at all at present, otherwise the framework is simply not serious). Criticisms of other theoretical frameworks can be valuable, but should be done by proposing competing analyses, not by pointing out defects in actually existing ones (so to continue the analogy, by explaining the motions of galaxies and galactic clusters if one does not believe in dark matter). With that in mind...

Core assumptions of minimalist syntax

Usual core assumptions and goals in minimalist syntax are as follows.

1) There exists an early stage of syntactic computations (narrow syntax) the objects of which are constructed by repeated and recursive applications of a binary operation (Merge) in a bottom-up fashion. 2) The only operation beyond Merge allowed on these narrow syntactic objects is the operation Agree (that is features co-valuation) and both Merge and Agree are subject to strict locality conditions. 3) There exists specified and definite domains of computation (phases) the completion of which triggers transfer of the narrow-syntactic object somewhere else (the interfaces) with no return possible. 4) Everything else is interface conditions.

Potentially false predictions following from these assumptions

There are tons of them, because the assumptions above are extremely counter-intuitive, in fact worryingly so. Here are a few examples.

Sensitivity to hierarchical rather than to linear order.

It follows from 1) that narrow syntactic objects are necessarily unordered binary trees. Hence, minimalist syntax predicts that the relevant structures for syntax are those prevailing in binary trees (for instance C-command) and not the most obvious linear ordering of the words in the sentence.

With the caveat that no language wears its analysis on its sleeve, if one could find a language for which, say, negation was licensed by a phenomenon taking place at the third word of the sentence, or in which adverbs modify the closest verb in the linear order rather than the closest one in the hierarchical order (the infamous Instinctively, eagles that fly can swim example of Chomsky), or in which cataphoric pronouns were distinct from anaphoric pronouns, or in which questions words were in situ with no intervention effects, or in which words were systematically interpreted where they are pronounced, or for which binding of reflexives depended on linear precedence, or if there existed a V3 language in the same sense that there are V2 languages, then minimalist syntax would be dealt a serious blow.

Successive cyclic movement

Because narrow syntactic objects are constructed from the bottom-up (extremely weird), words appearing at the left-periphery of the sentence sometimes (often) have to be merged far deep in the hierarchical structure. Because computations is done by chunks, they then have to move at least out of each chunk to reach their final destination (weird). And finally, because the only way for them to move is to being remerged (extremely weird), the potentially false (in fact prima facie seemingly absurd) conclusion that sentences are riddled with copies of words all the way from their departure point to their final stop ensues.

So the existence of a language immune to island effects, or parasitic gap effects, or obligatory versus optional control effect, or without reconstruction effect would deal a serious blow to minimalist syntax. Beside, minimalist syntax would not have lasted 5 minutes (in fact would not have been born) if said copies were not occasionally found in some languages and constructions precisely where they were hypothesized to be.

Independence from the interfaces.

This is obvious, but well-worth being pointed out nevertheless. Minimalist assumptions affirm the independence of syntax from the interfaces. If there existed a language whose syntactic rules depended on phonology or for which, for instance, island effects would show up only for meaningless or unclear questions, then minimalist syntax would disappear.

A less trivial potentially false (in fact, again, prima facie absurd) conclusion occurs when this principle is taken conjointly with morphology: as soon as one attributes some overt morphology phenomenon to hierarchical position in minimalist syntax, then one has to assume that all languages exhibit (possibly abstractly) the same morphology effect in the same hierarchical position.This would be proved false, and put minimalist syntax in serious trouble, if for instance a language was found in which licensing of DP by nouns and by verbs were identical.

Answer 2

An answer that produces a list of clear criteria that could be used to test the predictive accurracy of grammar frameworks would be difficult. The grammar frameworks listed in the question generally all have components that can be accessed to make predictions, and if a given grammar framework is lacking in a specific area (it makes no predictions in that area), it can always be augmented in one way or another.

Neverthelesss, I think I can provide one major criterion. In the list of grammar frameworks mentioned in the question, worth noting is that two are constituncy-based (MP and LFG), one of them is dependency-based (MTT), and the other two are (to my knowledge) not so explicit about the dependency vs. constituency distinction (RRG, RCG).

In a number of my papers (such as in the one linked to in the addendum to the question), I attempt to call out all constituency-based theories by pointing to the diagnostics that are commonly employed in syntax textbooks to identify syntactic structure, i.e. to the tests for identifying constituents (e.g. topicalization, clefting, pseudoclefting, proform substitution, answer fragments). I emphasize that all constituency-based frameworks make a terribly inaccurate prediction in this area. Here's an example:

(1) Fred likes music.

a. *Likes Fred music.                        - Topicalization

b. *It is likes that Fred music.             - Clefting

c. *What Fred does about Music is likes.     - Pseudoclefting

d. *Fred does so music.                      - Proform substituion

e. Q: What does Fred think about music?            
   A: *Likes.                                - Answer fragment 

Based upon these data, one can conclude that the verb likes alone is NOT a constituent, since all five of the tests agree that it is not a constituent. Yet all constituency-based grammar frameworks (really, all of them!), necessarily view likes as a constituent. In constituency-based systems, every single individual word is a constituent by definition. Hence constituency-based frameworks seem to be terribly flawed at a basic level.

Dependency-based frameworks, in contrast, make the correct prediction because they do not view the finite verb likes alone as a constituent. It is not a constituent because it dominates other words. Dependency-based frameworks are therefore making the correct prediction. The point is touched on at the end of this article https://en.wikipedia.org/wiki/Constituent_%28linguistics%29 and it is debated entertainingly at length on the talk page for that same article here https://en.wikipedia.org/wiki/Talk:Constituent_%28linguistics%29.

Returning to the question, this is a major criterion that can be used to assess grammar frameworks. The criterion concerns constituency tests and what they tell us about syntactic structure.The dependency-based frameworks destroy the constituency-based frameworks in this area.

My experience with this issue (which has accummulated over many years) is that constituency-based frameworks have two ways of dealing with the problem. They preferrably ignore it by ignoring dependency-based frameworks as much as possible. Dependency grammarians are all very clear about the dependency vs. constituency distinction, whereas many constituency grammarians seem hardly aware that dependency even exists. This situation is starting to change, however, mainly because of the increasing use of dependency for automated parsing.

On the rare occassion that a constituency grammarian is confronted directly with the issue and thus forced to respond, a common answer appeals to projection levels (minimal, intermediate, and maximal). The tests for constituents diagnose for maximal projections, not for minimal, or intermediate ones. This observation does indeed address the shortcoming to a large extent. The tests can identify only those constituents that are maximal projections (i.e. complete phrases).

But this is where Occam's Razor enters the picture, as suggested in the question. Dependency makes the correct prediction without the onerous distinction between projection levels. In other words, to overcome the shortcoming, constituency-based systems need a much greater theoretical apparatus, one that acknowledges much, much more syntactic structure and also distinguishes between types of constituents (minimal, intermediate, and maximal ones). According to Occam's Razor, one should, rather, cut out the extra structure and onerous distinction by assuming a dependency-based system instead.

In sum, one major criterion that directly provides what the question is asking for is the approach to constituent structure. The diagnostics for constituent structure can be used to evaluate grammar frameworks.

Answer 3

Linguists seem to love diagrams, but what kind of diagrams you give for language expressions bears only a distant relationship to any real differences among grammatical theories, if any relationship at all. Since you've been interested in binary branching as one way to distinguish formal grammatical theories, I thought you might be interested in the following account of just how trivial a matter this is, theoretically. You can find a way to write binary branching diagrams for context free phrase structure grammar (cfpsg) or any of its notational variants (including dependency grammar).

A sentence is generated by a cfpsg when there is a correct derivation for the sentence, and a derivation is a list of strings of terminal and nonterminal symbols, the first in the list being an initial symbol (typically S) and the last in the list a string of terminal symbols (say, e.g., a string of phonemes giving the pronunciation of a sentence in the language being described). It's a very simple theory. Each line of the derivation after the first is produced from the preceding line by using a phrase structure rule to replace some non-terminal symbol with a string of symbols. The last line of the derivation has only terminal symbols (phonemes, let's say).

A tree diagram of the usual sort giving the grammatical structure of a sentence can be produced from the cfpsg derivation in a mechanical fashion, so that the tree structure can be thought of as, essentially, an abbreviated cfpsg derivation. Trees are grammatical derivations.

One way to construct such a tree diagram is to make a branch at each application of a phrase structure rule. There is a previous line of the derivation and a phrase structure rule, and from these two things, the next line of the derivation is derived by applying the phrase structure rule to some non-terminal in the previous line. Two things in the tree are combined to form the next node up.

So there you have it -- binary branching. As you can see, it's a trivial formal maneuver. If you wanted more branches than just two, you could arrange for that by applying several phrase structure rules at once in the underlying phrase structure derivation. Could we ever find out anything interesting about language by playing games with diagrams? I don't see how. Diagrams are not theories.