Relativized Minimality and similarity based interference

The goal of the Minimalist Program is to reduce the language specificity of FL (i.e. UG) to the absolute minimum, so as to reduce the Logical Problem of Language Evolution (LPLE) to manageable size. The basic premise  is that LPLE is more challenging the more linguistiky FL’s operations and principles are. This almost immediately implies that to solve the LPLE requires radically de-linguistifying the structure of FL.

There are several reasonable concerns regarding this version of LPLE, but I will put these aside here and assume, with Chomsky, that the question, though vague, is well posed and worth investigating (see here and here for some discussion). One obvious strategy for advancing this project is to try and reduce/unify well grounded linguistic principles of UG with those operative in other domains of cognition. Of the extant UG principles ripe for such reconceptualization, the most tempting, IMO and that of many others as we shall see, is Relativized Minimality (RM). What is RM to be unified with? Human/biological memory. In particular, it is tempting to see RM effects as what you get when you shove linguistic objects through the human memory system.[1] That’s the half-baked idea. In what follows I want to discuss whether it can be more fully baked.

First off, why think that this is doable at all? The main reason is the ubiquity of Similarity Based Interference (SBI) effects in the memory literature. Here is a very good accessible review of the relevant state of play by Van Dyke and Johns (VD&J).[2] It seems that human (in fact all biological) memory is content addressable (CA) (i.e. you call the memory in terms of its contents (rather than, say, an index)). Further, the more the contents of specific memories overlap, the more difficult it is to successfully get at them. More particularly, say that one accesses a memory via certain content cues, the more these cues overlap the more they “overload” the retrieval protocol making it harder to successfully get the right one. On the (trivial) assumption that memory will be required to deal with the ubiquitous non-(linearly)-adjacent dependencies found in language we should expect to find restrictions on linguistic dependencies that reflect this memory architecture.[3] VD&J review various experiments that show the effects that distracters can have on retrieving the right target when these distracters “resemble” the interacting expressions.

Given that memory will create SBIs it is natural to think that some kinds of dependencies will be favored over others by this kind of memory architecture. Which? Well ones in which the cues/features relating the dependents are dissimilar from those of the intervening elements. Graphically, (1) represents the relevant configuration. In (1), if non-adjacent X and Y need to be related (say there is a movement or antecedence dependency between the two) then this will be easiest if the cues/features relating them are not also shared by intervening Zish elements.

(1)  …X…Z…Y…

This should look very familiar to any syntactician who has ever heard the name ‘Luigi Rizzi’ (and if you haven’t think of either changing fields or getting into a better grad program). Since Rizzi’s 1990 work, (1), in the guise of RM, is standardly used to explain why WHs cannot move across other WHs (e.g. superiority and Wh-island effects) or heads over heads (the head movement constraint).

IMO, RM is one of the prettiest (because simplest) empirically useful ideas to have ever been proposed in syntax.[4] Moreover, its family resemblance to the kinds of configurations that induce SBI effects is hard to miss.[5] And the lure of relating the two is very tempting, so tempting that resistance is really perverse. So the question becomes, can we understand RM effects as species of SBI effects and thus reflections of facts about memory architecture?

Psychologists have been pursing a similar (though as we shall see, not identical) hunch for quite a while. There is now good evidence that VD&J reviews that encumbering (working) memory with word lists while sentences are being processed differentially affects processing times of non-local dependencies and that the difficulty is proportional to how similar the words in memory are with words in the sentence that need to be related. Thus, for example, if you are asked to keep in memory the triad TABLE-SINK-TRUCK while processing It was the boat that the guy who lived by the sea sailed after two days then you do better at establishing the dependency between boat and sail then if you are asked to parse the same sentence with fix in place of sail. Why? Because all three of the memory list words are fixable, while none are sailable. This fact makes boat harder to retrieve in the second fix sentence than the first sail sentence (p. 198-9).

Syntactic form can also induce interference effects. Thus, the subject advantage inside relative clauses (RC) (i.e. it is easier to parse subject relatives than object relatives, see here) is affected by the kinds of DPs present in the RC.  In particular take (2) and (3). The Subject Advantage is the fact that (2) is easier to parse.

(2)  The banker that praised the barber climbed the mountain 

(3)  The banker that the barber praised climbed the mountain

VD&J note that SBI effects are detectable in such cases as the Subject Advantage can be reduced or eliminated if in place of D-NP nominal like the barber one puts in pronouns, quantified DPs like everyone and/or proper names. The reasoning is that the definite descriptions interfere with one another, while the names, pronouns and quantifiers interfere with D-NPs far less.[6]

VD&J offers many more examples making effectively the same point: that the contents of memory can affect sentential parsing of non-local dependencies and that they do so by making retrieval harder.

So, features matter, both syntactic features and “semantic” ones (and, I would bet other kinds as well).

What are the relevant dimensions of similarity? Well, it appears that many things can disrupt, including grammatical and semantic differences. Thus, the “semantic” suitability of a word on the memorized word list and the syntactic features that differentiate one kind of nominal from another can serve to interfere with establishing the relation of interest. [7]

Friedmann, Belletti and Rizzi (FBR) (here) reports similar results, but this time for acquisition. It appears, for example, that subject relatives are more easily mastered than object relatives, as are subject vs object questions. FBR discusses data from Hebrew. Similar results are reported for Greek by Varlokosta, Nerantzini and Papadopoulou (VNP) here. Moreover, just as in the processing literature, it appears that DPs interfere with one another the more similar they are. Thus, replacing D-NP nominal with relative pronouns and bare whs (i.e. what vs what book) eases/eliminates the problem. As FBR and VNP note, the subject advantage is selective and, in their work, is correlated with the syntactic shapes of the intervening nominal.[8] The more similar they are, the more the problems caused.

So, at first blush, the idea that RM effects and SBI effects are really the same thing looks very promising. Both treat the shared features of the interveners and dependents as the relevant source of “trouble.” However (and you knew that knew was coming, right?) things are likely more complicated. What’s clear is that features do make a difference, including syntactic ones. However, what’s also clear is that not only syntactic shape/features matters. So do many other kinds.

Moreover, it is not clear which similarities cause problems and which don’t. For example, the standard RM model (and the one outlined in FBR) concentrates on cases where the features are identical vs when they overlap vs when they are entirely disjoint. The problem with relative clauses like (3) for example, is that the head of the relative clause and the intervening subject have the exact same syntactic D-NP shape and the reason subbing a pronoun or name or quantifier might be expected to mitigate difficulty is that the subject intervener only share some of their features thereby reducing the minimiality effects. So in the case of RCs the story works as we expect.

The problem is that there are other data to suggest that this version of RM delivers the wrong answers in other kinds of cases. For example, a recent paper by Atkinson, Apple, Rawlins and Omaki (here) (AARO) shows that “the distribution of D-linking amelioration effect [sic] is not consistent with Featural Relativized Minimality’s predictions…” (1). AARO argues that carefully controlled rating methods of the experimental syntax variety show that moving a which-NP over a which-NP in Spec C is better than moving it over a who (i.e. (4) is reliably better than (5)). This is not what is expected given the featural identity in the first case and mere overlap in the second.[9]

(4)  Which athlete did she wonder which coach would recruit

(5)  Which athlete did she wonder who would recruit

IMO, superiority shows much the same thing. So (6) is quite a bit better than (7) to my ear.

(6)  I wonder which book which boy read

(7)  I wonder which book who read

Once again, the simple syntactic version of RM suggests that the opposite should be the case. If this is so, then there is more than just structural similarity involved in RM effects.

This, however, might be a good thing if one’s aim is to treat RM effects as instances of more general SBI effects. We expect many different factors to interact to provide a gradation of effects, with syntactic shape being one factor among many. The AARO data suggests that this might indeed be correct, as are the parallels between the VD&J parsing data and the FBR/VNP acquisition data. So, if AARO is on the right track, it argues in favor of construing RM effects as kinds of SBI effects, and this is what we would expect were RM not a grammatically primitive feature of FL/UG, but the reflection of general memory architecture when applied to linguistic objects. In other words, this need not be a problem, for this is what one would expect if RM were just a species of SBI (and hence traceable to human memory being content addressable).[10]

What is more problematic, perhaps, is settling what “intervention” means. In the memory literature intervention is entirely a matter of temporal order (i.e. if Z is active when X and Y are being considered it counts as an “intervener,” i.e. roughly speaking if Z is temporally between X and Y then Z intervenes), while for RM the general idea is that intervention is stated in terms of c-command (i.e. Z intervenes between X and Y if X c-commands Z and Z c-commands Y) and this has no simple temporal implications. Thus, the notion explored in the memory literature mainly explores a “linear” notion of intervention while RM relies on a structural notion and so it is not clear that RM effects should be assimilated to memory effects.

However, I am not currently sure how big a problem this might be. Here’s what I mean.

First, much of the literature reviewed in VD&J involves English data where linear and hierarchical intervention will be the same. We know that when these two are pulled apart in many cases it is hierarchy that matters (see the discussion of the Yun et. al. paper here. It shows that the Subject Advantage persists in languages where linear order and hierarchical order go in different directions).

Similarly, Brian Dillon (here) and Dave Kush (in his unavailable thesis; ask him) show that hierarchical, not linear, intervention is what’s critical in computing binding relations.

Of course, there are also cases where hierarchy does not rule, and linear intervention seems to matter (e.g. agreement attraction errors and certain NPI licensing illusions are less sensitive to structural information, at least in some online tasks, then hierarchical restrictions suggest they should be).[11]

So both notions of proximity seem to play a role in language processing. I don’t know whether both have an effect in acquisition (but see note 11). So, does this mean that we cannot unify RM effects with SBI effects for they apply in different kinds of configurations?  Maybe not.  Here’s why.

Memory effects arise from two sources: the structure of memory (e.g. is it content addressable, rates of decay, number of buffers, RAM etc.) and the data structures that memory works on. It is thus possible that when memory manipulates syntactic structures that it will measure intervention hierarchically because linguistic objects are hierarchically structured. In other words, if phrase markers are bereft of linear order information (as say a set theoretic understanding of phrase markers entails) then when memory deals with these it will not be able to use linear notions to manipulate them because such objects have no linear structure. In these cases, when task demands use memory to calculate the properties of phrase markers, then RM effects is what we expect to see: SBIs with c-command determining intervention.  Of course, sentences when used have more than hierarchical structure and it is reasonable to suppose that this too will affect how linguistic items are used.[12] However, this does not prevent thinking of RM as a memory effect defined over PM-like data structures. And there is every reason to hope that this is in fact correct for if it is then we can treat RM as a special case of a more general cognitive fact about us; that we have content addressable memories that are subject to SBI effects. In other words, we can reduce the linguistic specificity of FL.

Maybe an analogy will help here. In a wonderful book, What the hands reveal about the brain (here), Poizner, Bellugi and Klima (PBK) describe the two ways that ASL speakers with brains damage describe spatial layouts. As you all know, space has both a grammatical and a physical sense for an ASL speaker. What PBK note is that when this space is used grammatically then it functions differently than when it is used physically. When used physically, stroke patients with spatial deficits show all the characteristic problems that regular right hemisphere stroke patients show (e.g. they use only half the space). However, when signing in the space (i.e. using the space linguistically) then this particular spatial deficit goes away and patients no longer neglect half the signing space. In other words, depending on how the space is being used, linguistically or physically, determines what deficits are observed.

Say something analogous is true with memory; when used in computing grammatical properties, intervention is hierarchical. When used otherwise, linear/temporal structure may arise.  Thus what counts as an intervener will depend what properties memory is being used to determine. If something like this makes sense (or can be made to make sense) then unifying RM effects as SBI effects with both related to how human memory works looks possible.[13]

Enough. This post is both too long and too rambling. Let me end with a provocative statement. Many at Athens felt that Darwin’s Problem (the LPLE) is too vague to be useful. Some conceded that it might have poetic charm, that is was a kind of inspirational ditty. Few (none?) thought that it could support a research program. As I’ve said many times before, I think that this is wrong, or at least too hasty. The obvious program that LPLE (aka Darwin’s Problem) supports is a reductive/unificational one. To solve LPLE requires showing that most of the principles operative in FL are non-linguistically specific. This means showing how they could be reflections of something more cognitively (or computationally or physically) general. RM seems ripe for such a reanalysis in more general terms. However, successfully showing that RM is a special case of SBI which is grounded in how human memory operates will take a lot of work, and it might fail. However, the papers I’ve cited above outline how redeeming this hunch might proceed. Can it work? Who knows. Should it work? Yup, the LPLE/DP hangs on it.[14]

---

[1] The first person I heard making this connection explicitly is Ivan Ortega Santos. He did this in his 896 paper at UMD in about 2007 (a version published here). It appears that others were barking up a similar tree somewhat earlier, as reviewed in the paper by Friedmann, Belletti, Rizzi paper discussed in what follows. The interested reader should go there for references.

[2] Julie Van Dyke and Brian McElree (VD&M) wrote another paper that I found helpful (here). It tries to zero in on a more exact specification of the core properties of content addressable memory systems. The feature that they identify as key is the following:

The defining property of a content addressable retrieval mechanism is that information (cues) in the retrieval context enables direct access to relevant memory representations, without the need to search through extraneous representations (164).

In effect, there is no cost to “search.” Curiously, I believe that VD&M get this slightly wrong. CA specifies that information is called in virtue of substantive properties of its contents. This could be combined with a serial search. However, it is typically combined with a RAM architecture in which all retrieval is in constant time. So general CA theories combine a theory of addressability with RAM architecture, the latter obviating costs to search. That said, I will assume that both features are critical to human memory and that the description they offer above of CA systems correctly describes biological memory.

[3] In other words, unifying RM with CA systems would effectively treat RM effects as akin to what one finds in self-embedding structures. These are well known to be very unacceptable despite their grammaticality (e.g. *that that that they left upset me frightened him concerned her).

[4] My other favorite is the A-over-A principle. Happily, these two can be unified as two sides of the same RM coin if constituents are labeled. See here for one version of this unification.

[5] In fact, you might enjoy comparing the pix in VD&J (p. 197) with (1) above to see how close the conceptions are.

[6]  VD&J do not report whether replacing both D-NPs with pronouns or quantifiers reintroduces the Subject Advantage that replacing barber eliminates. The prediction would seem to be that it should on a straightforward reading of RM. Thus, someone who you/he saw should contras with someone who saw you/him in roughly the same way that (2) and (3) do. VNP (see below) report some data suggesting that quantifiers might pose separate problems. VD&M reporting on the original Gordon & Co studies note “their data indicate that similarity-based interference occurs when the second noun phrase is from the same referential class as the first noun phrase, but it is reduced or eliminated when the noun phrases are from different classes (158).” This suggests that the SBI effects are symmetric.

[7] The scare quotes are here for in the relevant examples exploit a “what makes sense” metric, not a type measure. All the listed expressions in boat-sail and boat-fix examples are of the same semantic type, though only boats are “sailable.” Thus it is really the semantic content that matters here, not some more abstract features. VD&M review other data that points to the conclusion that there are myriad dimensions of “similarity” than can induce SBIs.

[8] VD&J cites the work of Gordon and collaborators. They do not link the abatement of SBIs to syntactic shape but to their semantic functions, their “differeing referential status.” This could be tested. If Gordon is right, then languages with names that come with overt determiners (e.g. Der Hans in German) should, on the assumption that they function semantically the same as names in English do, obviate SBIs when a D-NP is head of the relative. If MR is responsible, then these should function like any other D-NP nominal and show a Subject Advantage.

[9] This is the main point. Of course there are many details and controls to worry about which is why AARO is 50 pages rather than a paragraph.

[10] This might create trouble for the strong RM effects, like moving adjuncts out of WH islands: *How did you wonder whether Bill sang. This is a really bad RM effect. This is correct and the question arises why so bad? Dunno. But then again, we currently do not have a great theory of these ECP effects anyhow. One could of course concoct a series of features that led to the right result, but, unfortunately, one could also find features that would predict the opposite. So, right now, these hard effects are not well understood, so far as I can tell, by anyone.

[11] Note the very tentative nature of this remark. Are there any results from language processing/production/acquisition that implicate purely linear relations? I don’t know (a little help would be nice from you in the know). The NPI stuff and the agreement attraction errors are not entirely insensitive to hierarchical structure. Maybe this: VNP cite work by Friedman and Costa which shows that children have problems with crossing dependencies in coordinate structures (e.g. The grandma₁ drew the girl and t₁ smiled). The “crossing” seems to be linear, not hierarchical. At any rate, it is not clear that the psycho data cannot be reinterpreted in large part in hierarchical terms.

[12] However, from what I can tell, pure linear effects (e.g. due to decay) are pretty hard to find and where they are found seem to be of secondary importance. See VD&J and VD&M for discussion. VD&J sum things up as follows:

…interference is the primary factor contributing to the difficulty of integrating associated constituents…wit a more specialized role arising for decay…

[13] One other concern might be the following: aren’t grammatical restrictions categorical while performance ones are not? Perhaps. Even RM effects lead to graded acceptability, with some violations being much worse than others. Moreover, it is possible that RM effects are SPI effects that have been “grammaticized.” Maybe. So RM is a grammatical design feature of G objects so that such objects mitigate the problems that CA memory necessarily imposes. I have been tempted to this view in the past. But now I am unsure. My problem lies with the notion “grammaticization.” I have no idea what this process is, what objects it operates over, and how it takes gradient effects and makes them categorical. At any rate, this is another avenue to explore.

[14] There are some syntactic implications for the unification of RM and SBI effects in terms of the structure of CA memory. For example, if RM effects are due to CA architecture then issues of minimal search (a staple of current theories) are likely moot. Why? Well, because as observed (see note 2), CA “enables direct access to relevant memory representations, without the need to search through extraneous representations.”

In other words, CA eschews serial search and so the relevance of minimal search is moot if RM effects are just CA effects. All targets are available “at once” with none more accessible than any others. In other words, no target is further/less accessible than any other. Thus if RM special case of CA then it not search that drives it. This does not mean that distance does not matter, just that it does not matter for search. Search turns out to be the wrong notion.

Here’s another possible implication: if decay is a secondary effect, then distance per se should not matter much. What will matter is the amount of intervening “similar” material. This insight is actually coded into most theories of locality: RM is irrelevant if there is only one DP looking for its base position. Interestingly, the same is true of phase based theories, for structures without two DPs are “weak” phases and these do not impose locality restrictions. Thus, the problems arise when there are structures supporting two relevant potential antecedents of a gap, just as a theory of minimality based on SBI/CA would lead one to suppose.