Rediscovering a hidden classic: Comrie (1999) on explaining some universals of coreference marking

The following language universal seems to be poorly known, even among aficionados of universals:

Universal 1

Coreference marking tends to be longer in more local domains, and shorter in more extended domains.

The correctness of this universal, due to Comrie (1999), has not been systematically documented, but there are excellent reasons to think that Bernard Comrie was right in his far-sighted claims. And in addition, I think he proposed the correct explanation (see below). (But note that Comrie talks about “degrees of markedness” (p. 341), not about “length of marking”; the context makes it clear that he does indeed mean the latter.)

By “local domain”, Comrie means the verb and its arguments, while the adjuncts are less local, and subordinate clauses are still less local. So the generalization of Universal 1 can be illustrated by English and Russian, because the longer anaphoric forms (himself, sebja) are obligatory in the most local domain, not possible in the most extended domain (subordinate clauses), with variation in intermediate domains (clause-level adjuncts and possessors). This is illustrated by his examples (19)-(29).











Comrie then goes on to show that Universal 1 also makes the right predictions for languages with three types of anaphoric constructions, e.g. Japanese (zibun zisin, zibun, and Ø) and Tamil (where there is additional verbal marking of coreference in the most local domain).

Japanese is well-known for having zero anaphora corresponding to English her/him/them, but Comrie then shows that “zero anaphora” is actually found also in English and Russian subordinate clauses, as in (44) and (49) – in the latter, only Russian allows it, and in the former, both languages also require a special form of the verb (the infinitive). But crucially, these cases obey the generalization of Universal 1.






Most linguists have not considered reflexive pronouns and infinitival constructions together, and have thus missed what looks like a correct generalization. And what is nice about it is that it can be explained as an evolutionary adaptation – an instance of efficient asymmetric coding, like so many other general aspects of grammar. The key observation is the following:

Observation 1

Coreference is less likely in more local domains, and more likely in more extended domains.

So the reason why we get longer anaphoric marking in an agent-patient domain (as I also noted in my recent LSA talk) is that it is simply not very likely that the patient is the same as the agent (If John saw someone, it is unlikely that it is himself that he saw). Languages are flexible, and over time they often develop an efficient coding system, with longer forms for less expected situations, and shorter forms (or zero) for more expected situations.

This insight is present in Comrie’s paper, but perhaps not quite transparently. Instead of “likelihood” (as in Observation 1), Comrie talks about “the normal situation”, or “the most natural situation”, or “the unmarked situation”. These are somewhat confusing terms (especially “unmarked”, because the word “marked” is also used in the sense of formal length), but a careful reading of this (beautifully written) paper makes it clear that Comrie means exactly this. Let me quote two passages:

“The normal situation in terms of our conceptualization of the world is for an action that includes more than one participant to have an agent acting on a patient that is a distinct entity from the agent. In other words, the most natural situation is for the argments of a predicate to be non-coreferential.” (Comrie 1999: 341)

“The unmarked situation is to have “referential continuity” (“topic cntinuity”). Specia marking is used to indicate deviation from this pattern…” (Comrie 1999: 343)

Comrie did not himself make any frequency counts, and there seems to be very little work in this direction (but see Haspelmath 2008; Ariel 2008 for some frequency counts of a few languages). In general, linguists seem to be more interested in meaning-form correspondences, even though form-frequency correspondences clearly provide the best explanations. Comrie himself is a bit ambivalent about the nature of his explanation: On the one hand, he recognizes that there is an overarching pattern, accounting both for reflexive marking and coreference in extended domains, but he also says that he provides a “cognitive explanation” for the local domain, and that the principle for the extended domain has a “discourse basis” (1999: 335). While “cognition” and discourse no doubt play an important role, it seems much simpler to say that all the patterns he discusses have an explanation in terms of efficient asymmetric coding.

Most work on reflexive pronouns comes from the generative tradition (of “binding theory”, e.g. Reinhart 1983; Reuland 2011), and many intriguing language-particular patterns have been found. However, I am not aware of any similarly simple and powerful generalization to have emerged from this literature. Reuland (2011: 186) invokes a principle called IDI (“inability to distinguish indistinguishables”) and says that it is “a general property of computations”, seen also in the Obligatory Contour Principle in phonology. However, how exactly the ungrammaticality of *Jane1 saw her1 follows from this principle (while at the same allowing Jane washed, where agent and patient are identical) is not easy to see in Reuland’s highly complex system.

So it seems to be worthwhile revisiting Comrie’s old paper and trying to build a larger theory of anaphoric forms and constructions on it.


Ariel, Mira. 2008. Pragmatics and grammar. Cambridge: Cambridge University Press.

Comrie, Bernard. 1999. Reference-tracking: Description and explanation. Sprachtypologie und Universalienforschung 52(3–4). 335–346.

Haspelmath, Martin. 2008. A frequentist explanation of some universals of reflexive marking. Linguistic Discovery 6(1). 40–63.

Reinhart, Tanya. 1983. Anaphora and semantic interpretation. London: Routledge.

Reuland, Eric J. 2011. Anaphora and language design. Cambridge, Mass: MIT Press.

8 thoughts on “Rediscovering a hidden classic: Comrie (1999) on explaining some universals of coreference marking

  1. [Continuing the discussion started by David Pesetsky] What correlates with short coding is frequency of use (an independently established, measurable factor that causes hearer expectations), as partially documented in my 2008 paper, and also in related work by Mira Ariel, as well as Petra Hendriks and colleagues. In Comrie’s paper, this was left a bit vague (he uses “normal” for what I call “frequent”), but I think it can be made more precise – probably in a variety of ways, and these need to be discussed more in the future. Perhaps the most striking contrast is between “She(1) saw her-self(1)” and “She(1) took her(*self’s)(1/2) bag”, apparently due to the fact that coreference/binding is much more expected for possessors. One may opt for a different p-linguistic structural analysis (in terms of c-command or similar), but there is a strong cross-linguistic trend here, which can hardly be expressed by cross-linguistic structural notions. (Incidentally, do you have an explanation why Russian has Ivan i ego sestra ‘Ivan and his sister’, rather than Ivan i svoja sestra? It seems that c-command would suggest that the former should be impossible, because Russian otherwise requires svoj when the possessor is in the c-command domain of the antecedent. For this reason, I have always been skeptical of asymmetric descriptions of coordination.)

    • Do you mean “less expected situations”, as you wrote in your blog, or “frequency of use”, as in your comment — and if the latter, frequency of use of what?

      (Concerning “Ivan i ego/*svoja sestra”, c-command alone doesn’t license reflexives in Russian, of course — and “svoj” is subject-oriented, more or less. So my guess is that will rule out the reflexive, whatever the right theory of subject orientation is. But in a language where possessive reflexives are not subject-oriented, I wonder if one gets the same data.)

      I will stop now — but thank you for the conversation.

  2. You write: “Most work on reflexive pronouns comes from the generative tradition (of ‘binding theory’ […], and many intriguing language-particular patterns have been found. However, I am not aware of any similarly simple and powerful generalization to have emerged from this literature.”

    Actually, a one-way version of a longer/shorter generalization was actually discovered and reported in the generative literature of the 1980s by Dong-Whee Yang (1983) and by Pierre Pica (1987) — that monomorphemic reflexives may permit long-distance binding, but polymorphemic (phrasal, in their view) reflexives do not.

    But I would have thought that an additional “simple and powerful generalization” that emerged early on from the study of reflexive and non-reflexive anaphora in the generative literature is that the distribution of these forms does not have anything to do with co-argument status (e.g. being agent and patient of the same predicate), despite initial appearances. The easiest way to see this is by looking at raising constructions that create configurations in which an argument of an embedded predicate assumes a new grammatical relation in a higher clause, and the distribution of pronoun vs. reflexive reflects the surface grammatical relation, not “the verb and its arguments”, e.g. in English “Mary seemed to herself (*her = ‘Mary’) to be likely to solve the problem.”, “Mary considers herself (*her = ‘Mary’) to have solved the problem” (vs. *Mary believes that herself has solved the problem).

    The converse demonstration can also be made: co-arguments where the preference for a reflexive over a non-reflexive predicted by the Comrie/Haspelmath approach is reversed. In Icelandic, dative-subject/nominative-object verbs exemplify this. The subject and object may corefer if the object is a non-reflexive pronoun, but no reflexive is possible (Everaert 1991).

    There’s lots of debate about the details, as you know. For example, the treatment of these phenomena in the LFG and HPSG literature, as far as I know about it, involves the extension of notions similar to “co-argument” to raising configurations. But I can’t see how “more expected situation” vs. “less expected situation” can possibly explain such differences once one looks across verb classes and configurations. Anaphora into the subject position of a finite clause vs. a non-finite clause, for example, describes the same situation. I don’t know for sure, but I wouldn’t be surprised to learn that we can produce pairs of DAT-NOM and NOM-ACC/OBL predicates in Icelandic that describe similar situations as well but show opposite distributions of dedicated reflexive form vs. pronoun. I’m not an anaphora expert, but as far as I can tell, wherever one looks, the distribution of reflexive vs. non-reflexive forms tracks formal properties like case and syntactic location (whatever mysteries remain about the details), not situational expectedness.

    P.S. In what way does Russian reflexive “sebja” count as “longer” than “eë” ‘her’, “ego”, ‘him’? Two open syllables, both have onsets (since the pronouns start with [j], transliteration notwithstanding). Should I be focusing on “ix” ‘them’? What am I missing?

    Everaert, Martin. 1991. Nominative anaphors in Icelandic: morphology or syntax? In Issues in Germanic syntax, ed. Werner Abraham, Wim Kosmeijer, and Eric Reuland, 277–307. Amsterdam: Benjamins.

    Pica, P. (1987), ‘On the nature of the reflexivization cycle’, in McDonough and B. Plunket (eds.), Proceedings of NELS 17, University of Massachusetts, Amherst, pp. 483–499.

    Yang, D.-W. (1983), ‘The extended binding theory of anaphors’, Language Research 19, 169–192.

    • Many thanks, David, for these comments! I knew about Pica, and I cited him in my 2008 paper (, but I’m glad to hear about Yang (1983). Pica tries to give an explanation, but I think the functional explanation is more general and easier to test.

      Now you rightly point out that the generalizations for English and Icelandic are not predicted – that is absolutely true, so we need p-analyses in addition to the g-explanations proposed by Comrie and myself. I think the reason we made progress over the Reinhart-Chomsky approach (and also over Reuland and associates) was that we do not conflate language-particular generalizations (= p-analyses) with explanations at the general level of Human Language. Conceptually, the p-/g-distinction is clear, but in practice, Chomsky and his followers have constantly played it down, and have tried to carry over analyses from one language to another. This makes it very difficult to see what the universals are, as we discussed before (e.g. here:

      So what I’m saying is that for English, your proposed p-analysis may be right because of the English Raising construction, and for Icelandic, Everaert’s proposed p-analysis may be right because of the Dative Subject construction. But these are language-particular facts that do not immediately tell us anything about Human Language – for this, we need universals (because general linguistics must be based on universals And what I claim about reflexive constructions is that whatever is universal about the reflexive forms (voice markers and reflexive pronouns) can be explained with reference to expectation-based efficient coding (

      And yes, you are probably right that “the distribution of reflexive vs. non-reflexive forms tracks formal properties like case and syntactic location” – so these may indeed be somehow part of human linguisticality (or biological capacity for language). But this claim needs to be formulated precisely and in a testable fashion – for example, we must make sure that “reflexive/non-reflexive” is not *defined* in such a way that the outcome is certain. So we must say clearly what kinds of logically possible languages are excluded by our theories and predicted not to occur (or to occur rarely). This is what I’m currently working on.

      P.S. I usually operationalize length as number of segments (standing in for articulatory effort), so indeed, “sebja” (4 segments) is not longer than “ee [jijo]” (4 segments). If one looks at enough languages (as one must), it is not necessary to go into more detail. It may or may not be a coincidence that a more detailed picture would reveal less effort for “ee” than for “sebja” in this particular detail. For illustrative purposes, this example is thus not ideal, but also not completely wrong.

      • Thank you for your reply. I don’t see how your p- vs. g.- issue is relevant to evaluating the strength of the proposal you favor. Is there any language at all for which situational expectedness can be shown to offer a superior explanation for the distribution of pronoun vs. reflexive (or shorter vs. longer anaphoric device) to a grammar-based account — a language in which it can be shown that the comings and goings of a reflexive form correlates with the expectedness of the relevant situation?

        Imagine a language, for example, in which “Mary saw her” describes a situation involving two distinct individuals, while if Mary is watching a movie where she unexpectedly appears in a crowd scene, and therefore need a reflexive predicate, we would say “Mary saw herself” — but if we added “… in a mirror”, the judgments reverse. You can of course see other individuals reflected in a mirror, but the expected situation is one in which you see yourself. Do we ever find such a thing?

        And if as you agree, the Russian examples don’t really fit the longer/shorter side of your generalization, what are we left with? There is a distribution to be described for Russian “sebja” vs. its non-reflexive counterparts, and if longer/shorter is irrelevant and the relevance of situational expectedness cannot be shown, how does Russian support your version of Comrie’s proposal at all?

        • No, there’s no single language for which the efficiency theory offers “a superior distrbutional explanation”. The latter is p-linguistics, and this is not my concern. As I said, I think each language has its own building blocks, and I’m not making claims about particular languages. Reinhart may well have been right about English. The problem is with the claim that Reinhart’s account of English is necessarily relevant to other languages, or that it is even related to innate properties of the human mind (which would be necessary for it to be relevant to stimulus poverty considerations).

          The claim is exclusively about general trends of coding, especially coding length (there also seems to be a general trend of role rank affecting the direction of binding, so that in most languages you can’t say “*Himself hilled John”; I haven’t made claims about that general trend).

          Expectation is not usually directly reflected in grammatical coding (though some languages have mirative markers of various sorts), but that’s not what I’m talking about. I’m talking about an evolutionary explanation of the efficient coding that we observe: In general, in this and in many other situations (see, shorter coding tends to be used for more expected situations (if the coding is asymmetric). This seems to explain many phenomena that others have tried to explain by innateness-based proposals, but these are undermotivated once one considers the simpler efficiency-based theory.

          Russian does not directly “support” the theory, but it illustrates it. The theory is supported by a claimed cross-linguistic trend (and can be refuted by showing that the claimed trend is illusory), not by any single language. (Polish would have been a better language to illustrate it, because “siebie/ją” does show a length difference.)

          • Thanks again, but still genuinely not following. What data supports your “g-linguistic” claim? Limiting ourselves to anaphora/reflexiviation phenomena, what independently establishable metric of expectedness correlates better with long vs. short coding of referential links or binding than alternatives?

    • Concerning monomorphemic reflexives and long-distance binding, my work-in-progress on Balinese with my co-author Shannon Bryant indicates that the morphologically complex anaphor “awak-ne” (self-POSS.3SG) may always be read long-distance, whereas the morphologically simpler anaphor “awak” (self) cannot. This is also potentially the case for Turkish and Hungarian. Here is a link:

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.