Root allosemy and affix selection

I've never posted a proper "blog entry" of this kind before, but I thought I would try at least as a way to get some thoughts down while they are still fresh in my mind. As some of you know, I gave a talk on allosemy at NELS on Friday. Some of you raised some really interesting questions, during the Q&A and later. I'll put the handout in the dropbox folder in case anyone wants to have a look. 

Yesterday, two days after my talk at NELS, someone asked me a very interesting and hard question that relates to some of the things we have been discussing in this group. It's something I have thought about many times before, but I think that the range of approaches we might consider could be different now. The question was what to do about cases where an overt verbalizing or nominalizing affix seems to affect root allosemy. These kinds of cases are fairly well-known, and have been discussed by Dave Embick, Hagit Borer, etc.. These are things like cover vs. coverage, form vs. formation, government vs. governance, etc. 

One thing I should say from the start is that I don’t think these kinds of examples affect the main point I was making about the relationship between morphology and semantics. The semantic distinction between RN, SN, and CEN readings (although I have a slightly different semantic typology in my book) is not reflected in the affixes that realize the property of “being a noun”. Likewise, the distinction between causative, inchoative, activity, etc., is not reflected in the choice of verbalizers. So from the perspective of “being verb/noun” syntactically and morphologically, the semantic “ways of being a verb/noun” seem to be blind to the morphological choices. 

But there is a real issue here, which is that if we take the derivational affixes to be allomorphs of v and n, as I and many others generally have, and the information about that choice is not available at LF, then how do we get a state of affairs where the root seems to be sensitive to the choice? 

Here I think we could divide up the approaches into two basic kinds: Approach A would say that information about, say, the distinction between -ance and -ment actually is available in the syntax and thus the semantics. Roots can then be sensitive to that in the normal way. Approach B would insist that the distinction is not available in the syntax, so sensitivity of the root must be something else. 

The approach to derivation that we have been considering lately, where overt categorizers are really roots + category, would be a version of Approach A, and it solves the problem at hand almost immediately. But there are some other issues that might make one reluctant to go this way. First, the basic issue of the many-to-many relations that led me to adopt the allosemy approach to verbalizers and nominalizers in the first place would lose a lot of its bite. Now -ize, -ify etc., are systematically semantically vacuous roots. We don’t want to say that they are subject to the allosemy that gives you CAUSE vs. BECOME, because then we are back to the “many-to-many” problem. (All the roots would have the same set of allosemes.) Moreover, now we have no direct morphological marking of being a verb—the categorizers that are doing the semantic work are all null. 

Another possible consequence of this approach is that we don’t have an easy way to distinguish between derivation and compounding. This is ok—and in some sense probably correct. But in chapter 6 of my 2023 book I analyzed a “verbalizer” -væða that really does look like a bound root compound head. It has the property that it cannot occur with semantically vacuous v because of the way compounds are interpreted. The idea is that the semantics recognizes the structure of a compound and computes a relation between the head and non-head, which would fail if the head is semantically empty. This derives the fact that -væða verbs cannot be idiosyncratic, but have to be either CENs or true, compositional “result” nouns that are built on the meaning of the verb. So adopting the root approach to verbalizers/nominalizers means that this kind of analysis is lost—or at least, something else would have to be said to get it to work out. 

Another version, without roots, would be to say that, say, -ize and -ify really do realize different v heads, but they just aren’t on the dimensions that matter for our previous “semantic flavors” of v. The problem is that by and large I don’t know what kinds of v heads those would be. In some cases, we might point to things like -age as deriving mass nouns, which is true a lot of the time. But there are enough exceptions to force us back into the allosemy/many-to-many problem (e.g. carriages, packages, bandages, etc.). Even if worked sometimes, it doesn’t seem likely to work nearly enough to solve the problem. 

So let’s think about Approach B, where the information about the affix choice is not in the syntax. One solution is that consider that apparent “doublets” have different structures. So, for example, government doesn’t have a natural CEN reading, whereas governance does. So perhaps government is really root derived and governance is really verb-derived. Then we just need to allow VI rules that can look like:

n <—> -ance / [v √GOVERN v ] _ 

n <—> -ment / √GOVERN _

Perhaps many of the other cases can be accounted for like this. Partial generalizations like with -age could get a treatment like:

n <—> -age / { [{√1, √2, … } Num[–count]], √3, √4} __

Messy… lots of assumptions needed. But it’s a logically possibly approach. 

What if we can’t give a unique structure to all cases where affix choice seems to affect root meaning? These are things that Dave writes about in detail in his handout “Approaching Polymorphy”. One solution is that it’s all an illusion—that the difference between cover and coverage is not actually a matter of root allosemy—all the readings are available for both as far as the grammar is concerned. But our usage habits lead us in one direction or another most of the time, so our judgments are skewed about what is possible. 

A more extreme version of this might say that all root allosemy is a product of the encyclopedia, specifically in the sense that PF, context, everything, can feed it. There are some DM models that assume this, and Alec’s early DM paper “A late note on late insertion” essentially goes this way. I am reluctant to go this way if for no other reason than the fact that it seems to blur all the lines that we have worked so hard to draw carefully over the years: the distinction between idioms and allosemy, the “Wheel of Fortune” corpus, and so on. In my view, too much is lost if we assume that root meaning can be sensitive to anything, and if we adopt that view, then we lose all hope of discovering something interesting that this puzzle might be leading us to. 

Back to Embick’s approach, which I find to be the most interesting version of this overall approach, the basic idea is that the affix choice helps you navigate root meaning, but does not directly condition it. If I understand right, it is basically the other way around. You choose the root meaning you have in mind first, and then that choice might influence the realization of the categorizer. This would be sort of like the boundary where we determine that vP is an “agentive event” and so Voice gets the agent alloseme. The difference is that that is entirely in semantics. In this case, we want some kind of marker that says what alloseme of the root was chosen, so that at the next cycle the insertion of the categorizer at PF can be sensitive to that. So e.g.: 

[ √COVER n ] —> interpret root; add PF-visible marking

[ √COVER₁ n ] —> VI 

n <—> -age / √COVER₁ _

n <—> -Ø / √COVER₂ _

I do worry about this version (which is not exactly what Embick says) having the ability to “pass off” too much information to PF. Without at theory of what meanings can be “marked” and what that means, we risk expanding our descriptive power too much.

I also haven’t mentioned the other big part of the picture that Embick brings up, which is the fact that things like confusal, while odd/unacceptable, are closer to being possible than things like bended.

In the end, I am not sure where I currently stand on these things. Since working on the topic, I have wanted to stick to the assumptions within the world of Approach B, and mostly try to cover as much ground as possible within the “different structures” approach, plus some version of the “illusion” approach (i.e, free variation plus contextual habits). This was boosted by situations where, for example, I heard someone use coverage to refer to tupperware covers. I don’t think that I have been fully convinced that I need to vacate this position entirely, mostly because I don’t see an obvious alternative that strikes me as a net gain. Alec’s recent work has me seriously considering the “all derivational morphemes are roots” approach, because there the net gain might be larger, if it helps us to develop a principled understanding of “paradigmatic participial derivation” (if I might call it that). But it didn’t occur to me until I was asked yesterday that I am maybe not as comfortable with my default “answer” (or perhaps “response”) to this question as I used to be.  I welcome any thoughts any of you might have!