Martin Haspelmath: Many thanks, John, for reading my paper “Explaining grammatical coding asymmetries” so carefully as a reviewer for a journal. We have known each other for a long time, so there is no reason not to have this discussion in the open. I have learned a lot from your three books (1994; 2004; 2014), and your work has been a constant inspiration. So what do you think about the theory of frequency-induced coding efficiency that I summarize in that paper?
John Hawkins: The paper is a succinct summary of a lot of compelling and convincing empirical work in morphology, morphosyntax and syntax, carried out by you and colleagues, many of the details for which are spelled out in a large number of cited publications. The compelling part of the explanation involves the impressive number of form-frequency correspondences, many of which have been pointed out and documented for the first time by you, making this a substantive addition to the language typology literature. It is useful to have this summary in one convenient place.
More importantly, Greenberg’s original work in 1966 pointing out a number of these form-frequency correspondences (discussed under the rubric of “feature hierarchies” and “markedness”, which you critiqued insightfully in your 2006 paper) has still not had the impact it deserves within the language sciences as a whole, and it is important that the broader and more up-to-date discussion of these phenomena that we find in your work be added to the database of the language sciences and accessed by those who need it. With the recent explosion of interest, within psycholinguistics and computational linguistics in particular, in Zipf’s Law of Abbreviation, frequency effects and also predictability, the time is ripe for these cross-linguistic patterns to assume a more central role in the discipline, and your empirical patterns are among the most extensive and impressive that we have in this area.
Martin: Many thanks, that is nice to hear. One critical comment that I frequently get is that the frequency differences themselves are in need of explanation.
John: Well, there are all kinds of different reasons for the greater or lesser frequency of the different forms you discuss (singular vs. plural, present vs. future, animate vs. inanimate, male vs. female, etc.), including social biases in gender roles, communicative biases in favor of animates, real-world frequencies, the greater certainty of the present, etc, etc. You mention these, but as I pointed out earlier and as you agree, it isn’t really necessary, in psycholinguistics, in discussions of Zipf, or in language typology, to give a detailed elaboration of these ultimate causes. We can just start with the frequencies themselves and derive predictions from them, for e.g. preferred meaning assignments to ambiguous words and phrases as Maryellen MacDonald does, or indeed for all discussions of frequency effects in production and comprehension studies, and for shorter forms in language typology as Greenberg and you do. So this potential problem for the ultimate explanation of shorter forms in coding, that frequency effects are themselves derived from something else, is something that we can, and do, all live with.
Martin: Good, but you also have some serious points of criticism. As I said, my thinking has been building upon your insights, but my focus has mostly been on the typological generalizations, and not so much on making connections with findings from psycholinguistics – a field that I do not know all that well, and that you know much more about.
John: For all the apparent simplicity of these form-frequency correspondences, and the cleanliness and generality of the correlations, the explanation for them is not at all straightforward nor well-understood, and it is because of this that you get yourself into some difficulty in the paper. Indeed, there is a danger that the paper, as you package the discussion and explanation at the moment, will not have the impact it deserves, despite its (for me) brilliantly intuitive and compelling nature.
Martin: OK, so how could the message of the paper be improved?
John: I think that your discussion of predictability in the paper is not satisfactory. In (3), you give the causal chain as “frequency of use -> predictability -> shortness of coding”. But the whole emphasis in psycholinguistics has shifted in the last few years from a preoccupation with parsing (and “integration” of incoming material in Ted Gibson’s sense) and production to prediction and its role in facilitating both comprehension and production. There are all kinds of issues being addressed and experiments being conducted and debates going on about how exactly prediction interacts with integration and working memory constraints, etc, and there have been some useful general papers written by e.g. Gina Kuperberg and Florian Jaeger and also Fernanda Ferreira asking what exactly we mean by prediction. I think that if you appeal explicitly to “predictability” as mediating between frequency of use and shortness of coding, you can’t just systematically ignore the work and insights of a whole field within the language sciences whose primary mission is to figure out precisely what predictability means and what role it plays in language use…
Martin: Well, I wrote the paper in such a way that it is easy to understand for any psycholinguist (though not addressed specifically to them, as you did in your 2007 paper), and I would hope that psycholinguists can look at what I say and try to relate their work to it. I do not want to be forced to take a stand on the debates in psycholinguistics. Moreover, if their mission is to “figure out precisely what predictability means”, it doesn’t sound as if they have come to definitive conclusions yet.
John: … and when one talks about one of these differentially coded forms being more “surprising” than another, you can’t completely ignore Roger Levy and his collaborators, who have developed a whole theory of “surprisal”. I feel that if you are going to invoke predictability at all, then you have to at least acknowledge the existence of this field and grapple with it in some way, because prediction is far more complex and nowhere near as well-understood as you or we would like, just like “markedness” and “iconicity” in linguistics are not straightforward primitives as you yourself have shown (in your 2006 and 2008 papers).
Martin: This sounds as if you suggest that I should take a stand on debates going on in an adjacent field that I could not possibly grasp fully. I find it hard enough to understand the things that non-psycholinguists have said about reflexive pronouns, differential argument marking and inverse constructions. The generative work on these topics is often very hard to understand, but when I understand it, it is often clearly misguided. So how do I know that the more technical work of the psycholinguists is on the right track? As far as I know, nobody is questioning the basic Zipfian insight that frequently expressed meanings have short forms because they are highly expected or predictable, whatever the details of predictability in language processing.
John: Let me be more specific: To the extent that we understand prediction to refer (as you seem to intend throughout most of your paper) to whatever cognitive consequences and benefits there are that derive from the frequency of use for a given form in all previous communications, then predictability in this sense is neither necessary nor sufficient to explain many shorter forms. It is not necessary because the immediate context can lead to shorter forms, regardless of the frequencies in all previous communications (just as meaning assignments to ambiguous words like port can lead to different preferences in a context about wine than their overall frequencies would lead us to expect, as shown by Maryellen MacDonald’s work).
Martin: I would say (admittedly in an intuitive sense of “predictability”) that the immediate context makes some meanings more predictable. When I’m watching football with my friends, then words like “referee” and “Mbappé” are more predictable than words like “reviewer” and “passé simple” (these latter words would be more expected in a linguistics context).
John: You cite Piantadosi et al.’s (2011) study on predictability, and you include both context and (overall) frequency to the left of the arrow leading to predictability in the causal chain in your Figure 1 (p.20). But a lot of these context effects have been discussed in the psycholinguistic and linguistic literature in terms of “accessibility” (e.g. in Mira Ariel’s work on referent accessibility), or in terms of greater “activation” (by Rick Lewis). If something is recently mentioned, and a short pronoun like it succeeds in identifying it for the hearer, then it is not necessarily predictable in any clear sense of the term that reference is going to be made to that entity again, or that it will have that particular referent.
Martin: Why not? Take a sentence like “Yesterday I watched a game of Paris Saint Germain. Initially, I enjoyed it.” Yes, I know Ariel’s work on accessibility, and she would say that the personal pronoun it can be used because the game (and the football club Paris Saint Germain) is highly accessible, having been mentioned in the previous sentence. But it is also easier to predict that I will talk about the game in the next sentence. Discourse tends to have high topic continuity, and we typically talk about the same things that we talked about just before. (For example, I can predict that your answer will make reference to predictability and related concepts.)
John: OK, but if this is also to be subsumed under predictability, we need some discussion of the common mechanism that can be claimed to underlie both the long-term frequency effects from all previous communications and the effects of the current discourse, which are often at variance with one another. The current accessibility-based idea is that a short pronoun can be used on this occasion because it is easy and efficient to do so, given the shared knowledge of speaker and hearer and the degree of activation and accessibility of the referent to both. But are accessibility and overall frequency synonymous with, or part of, an overarching notion of “predictability”? If so, we’re back to the central question: what exactly is meant by predictability?
Martin: Yes, as I said, I find that it makes very good intuitive sense to say that both contextual predictability (accessibility) and frequency-based predictability are kinds of predictability. I don’t have anything to contribute to explaining the meaning of the word “predictability”, but I see no problem with saying that both long-term frequency knowledge and current discourse awareness leads to greater predictability. If I’m known as a long-time fan of Paris Saint Germain, you will not be surprised to hear me talk about my club, and I will simply say “PSG” (or even reduce my articulation of Mbappé’s name: [bape] instead of [əmbape]). And if I just talked about linguistics, you will not be surprised when I use abbreviations like “NP” or “indexing” (instead of “person indexing of arguments”).
John: I think that a further problem is that predictability, in any sense, is not sufficient to explain shorter forms. There are some nice examples of this in the generative literature on deletions and ellipsis phenomena. For example it has been observed by Joanna Nykiel (among many others) that a sentence like I’m going to a wedding can be followed by a longer “sluicing” PP fragment To whose wedding? or by a shorter NP Whose wedding?. But the sentence Give it back, can be followed only by the longer “sprouting” fragment PP To whom?, not by the shorter NP *Whom?, even though the complement to-phrase (give it back (to someone)) is “implicit” and predictable from previous mention of the verb give. Yet precisely in a context when the complement is not mentioned like this, but predictable in a reasonable and intuitive and psycholinguistically verified sense of this term, the shorter counterpart cannot be used.
Martin: True, such cases are not immediately expected on an efficiency view of grammar.
John: Similarly for VP anaphora (I’m not sure he will do it) and VP ellipsis (I’m not sure he will), we have the famous distinction pointed out by Jorge Hankamer and Ivan Sag that the use of VP ellipsis is much more limited, typically to textual antecedents only, and cannot be used to describe a visual scene in which someone is observed trying to carry out the action in question. In both cases, the event being described by the VP is clearly accessible, either in previous discourse or visually, so if accessibility is to be subsumed under predictability, then we will have again a situation where there is predictability but it is not sufficient to explain why shorter forms, including zero, are not used in numerous cases. The traditional explanation given for examples of this sort is that degrees of accessibility (in terms of explicit versus implicit mention, textual vs. situational reference, structural accessibility, etc) are responsible for the shorter versus longer coding, not predictability as such.
Martin: I think that the mistake in this reasoning is that one is asking a theory of efficient coding to explain parochial facts of English. This often seems possible, but in fact, all we can explain (in most sciences) is general trends. We seem to be able to explain facts of English grammar, but in reality we explain a general trend that an English fact happens to instantiate. I have not seen a general trend where we find the kind of disconnect between accessibility that you describe for English. VP-Anaphora and VP-Ellipsis are very peculiar to English, so this may just be a quirk. In the case of Give it back! *(To) whom?, there is a straightforward English-internal explanation for the impossibility of the answer without the pronoun, and maybe there is such an explanation for the anaphora/ellipsis cases as well. But this would not be relevant to general linguistics.
John: I still think that you cannot just lump context and frequency together (as in your Figure 1 with arrows leading to predictability). The context effects point to greater or lesser ease of processing and to efficiency, in referent identification, VP-interpretation, etc, and these are prima facie very different in kind from the accumulated long-term frequency effects from all previous discourses. Calling it all “predictability” is not obviously correct.
Martin: Yes, it is not obvious to me that I am right, but it is also not clear to me what is wrong about lumping context and frequency together. I find it intuitively plausible that I can abbreviate the forms I use when I use them frequently, and that I can abbreviate them when I just used them in the previous context. This seems to happen all the time with letter abbreviations (like PSG or NP), and there seems to be quite a bit of evidence that this is also the case with articulation. More frequent words are articulated with less effort (e.g. time is shorter than thyme, Gahl 2008), and words that were used just before are also articulated with less effort.
John: In my general discussion of “Minimize Forms” in Hawkins (2004: 38-49) and elsewhere in that book, I tried to tease apart the many factors that lead to more minimal forms for lots of grammatical and lexical phenomena that have been discussed in the literature, and I summarize how language users enrich these forms in efficient communication (cf. esp. pp. 44-45). I don’t think that calling it all “predictability” is right or helpful.
Martin: Yes, your “Minimize Form” principle in that book (and also in your earlier 2004 book) is similar to my approach, but it does quite a bit more work, and I did not understand all of it fully. So my proposals are more modest, but of course broadly compatible with your principle (and in line with your general efficiency perspective), and I think that my claims could easily be shown to be wrong. But if the empirical predictions hold up, then we need an efficiency explanation of SOME sort – it seems that we agree on this. Let’s continue the debate on what exactly the explanation will be on some other occasion. And many thanks again for reading and discussing my paper!
Ariel, Mira. 1990. Accessing noun-phrase antecedents. London: Routledge.
Gahl, Susanne. 2008. Time and thyme are not homophones: The effect of lemma frequency on word durations in spontaneous speech. Language 84(3). 474–496.
Greenberg, Joseph H. 1966. Language universals: With special reference to feature hierarchies. The Hague: Mouton.
Haspelmath, Martin. 2006. Against markedness (and what to replace it with). Journal of Linguistics 42(1). 25–70. doi:10.1017/S0022226705003683.
Haspelmath, Martin. 2008. Frequency vs. iconicity in explaining grammatical asymmetries. Cognitive Linguistics 19(1). 1–33.
Hawkins, John A. 1994. A performance theory of order and constituency. Cambridge: Cambridge University Press.
Hawkins, John A. 2004. Efficiency and complexity in grammars. Oxford: Oxford University Press.
Hawkins, John A. 2007. Processing typology and why psychologists need to know about it. New Ideas in Psychology 25: 87-107.
Hawkins, John A. 2014. Cross-linguistic variation and efficiency. New York: Oxford University Press.
Kuperberg, Gina R. & T. Florian Jaeger. 2016. What do we mean by prediction in language comprehension? Language, Cognition and Neuroscience 31(1). 32–59. doi:10.1080/23273798.2015.1102299.
Piantadosi, S. T., H. Tily & E. Gibson. 2011. Word lengths are optimized for efficient communication. Proceedings of the National Academy of Sciences 108(9). 3526–3529.