Comments from reviewers for submission verion of Kuroda (2009) at PACLIC 23

Review 1

Let me distinguish four models of linguistic knowledge: (1) a strongly grammatical model, according to which only atomic elements are stored in memory and complex expressions are "generated" online by the grammar, (2) a weakly grammatical model, according to which complex expressions are "generated" but some of them, in addition to the atomic elements, are stored in memory, (3) a strongly usage-based model, according to which all the encountered instaces of linguistic expressions are stored in memory, and there is no such thing as grammar, and (4) a weakly usage-based model, according to which the encountered exemplars are first stored in memory, but some abstract patterns are extracted from them and constitute a "grammar"; such an "emergent" grammar acquires its own life. Clearly, the traditional wisdom is (1), while people like Gary S. Dell, in dealing with speech error data, probably had (2) in mind. I have received the impression that the paper under review attempts to argue for (3), but the way I see it, what is presented in fact is rather (4), and, unfortunately no version of (3) would probably work (although it is methodologically or philosophically ideal).

First, as the author is aware, the proposed theory sees a sentence as a list of word/morpheme. Words could be encountered as single-word utterances, but bound morphemes could never be. (The author's segmentation on the lower left corner of p. 5 treats "-ed" as an atomic element.) Segmentation into bound morphomes, hence, presupposes a morphological grammar. Also note that, in order to reject the independent notion of ditransitive construction, the author had to appeall to a family of superlexical items including [_, _ Bill, a letter]. (I have failed to see what status the presence vs. the absence of a comma has.) However, the author must definitely be unhappy if this particular superlexical item, for example, gave rise to something like "Right after Bill, a letter arrived." For this superlexical item NOT to give rise to this non-ditransitive sentence, the bracketing should play a role; two words followed by "Bill" and "a letter" cannot be found within a single constitutent in "Right after Bill, a letter arrived." However, this does presuppose constituency! Of course, such a discussion does not exclude a usage-based model in general; it only argues against (3), but not (4).

Note that people's "intuition that human memory is unreliable" (p. 2) can be understood in two ways: (i) it may easily be lost, or (ii) it may not be lost, but its content is transformed in time. Cognitive psychologists have already documented that the impression (ii) is correct (as the author perhaps is already aware). That defitenitely must be relevant here; if a vast number of encountered exemplars are first stored in memory, it does not necessarily excluded the possibility that some abstract patterns (grammar) are extracted ("emerge") and specific exemplars get lost gradually. Also note that connectionist models are instances of (4) in this regard.

This leads me to wonder how the paper should be located in the field. The author says in footnote 5 that Cognitive Linguistics fails to "escape form this grammar-based view," but aren't Langacker (1988) and Bybee (2001), referred to in the main text, Cognitive Linguists? (Definitely Langacker and Lakoff are "the" Cognitive Linguists for outsiders.) Langacker's Cognitive Grammar is clearly an instance of (4). The main contribution of this paper seems not to be an advancement of a new conception of language but a new formalization (in the author's term, foundation) of (4) in terms of pattern lattice.

Based on the content, the paper can be divided into two parts: (A) arguments for (3)-(4) and against (1)-(2), and (B) a specific proposal of pattern lattice formalization. Although I am personally very sympathetic to (4), as opposed to (1)-(2), I have to say I find the author's attempt in (A) has failed.

First, let me discuss the author's arguments on p. 2:

  1. noncompositional semantics: This argument is far from convincing. Here, Barwise and Perry's (1983) distinction between "meaning" and "interpretation" is useful. An interpretation of a sentence is an aspect of the (physical or mental) world that the sentence describes. The "meaning" is something that regulates such an "interpretation" relation. Barwise and Perry argue that the non-compositionality of "interpretation" does not argue against the compositionality of "meaning." Now, it is "interpretation" that counts in machine translation (MT). And it is already a textbook-level thesis in cognitive psychology that our understanding ("interpretation"), particularly at the discourse-level, largely comes from non-linguistic sources, and MT researchers in the 1980's were trying to implement non-linguistic knowledge, obviously a hopeless endevour (if it is to be done by human hand). The failure of traditional, grammar-based MT and the success of usage-based statistical MT, then, can naturally be interpreted with respect the availability of non-linguistic sources; each usage reflects non-linguistic knowledge in an implicit way. This interpretation fits nicely with the compositionality of natural language semantics and non-linguistic knowledge; it is just that more information is involved than a linguist can deal with.
  2. bounded productivity: (Somehow the discussion of this comes before (ii) in the text.) The "grammar" people can simply claim that such an aspect is beyond their scope; the compotence/performance distinction is still alive, whatever it actually means, and without convincingly arguing against that distinction, this argument has no force.
  3. endless variation: As far as I can see, no specific problem this poses for "grammar" people is given in the text.

Thus, the author's attempt in (A) seems to have failed. In addition, the author has failed to deal with various syntactic as well as phonological and psycholinguistic, observations that the "grammar" people have alluded to as evidence for sophisticated constituency; long-distance dependency is only a tip of the iceberg. For example, intonation has been observed to be affected in a systematic way by (syntactic and/or phonological) constituency. Perceptual processing has also been observed to be affected by syntactic constituency, etc. Even within an introductory syntax, we can easily find claimed evidence for constituency besides long-distance dependency (e.g., coordination). A usage-based model of the (4) kind could in principle deal with such phenomena, because it could assume that constituency "emerges" from remembered exemplars. However, the exact details of how are left unclear yet, as the author points out, and if the author admits (4) as opposed to (3), probably the author's theory could deal with them in principle; but this in turn means that the author's theory has also left the details unclear and is in the same vein as what the author would call unformalized/unfounded usage- based models already out there.

The author might want to spend some time studying phonology and phonetics. The memory-based or usage-based idea in the study of linguistic sounds is far older than Port (2007); it began with work on perception by people like Goldinger and Keith Johnson, and it has already gained wide, it not universal, acceptance. (I'm not sure if I remember it correctly, but I thought Goldinger's claim was once formalized in Adaptive Resonance Theory.) And, as noted above, claimed evidence for constituency abounds in phonology too. Also the author might want to note Janet Pierrehumbert's argument that type- (as oppposed to token-) frequencies of words affect phonetic manifetations of phonological patterns.

The author might also want to study cognitive/experimental psychology. For example, in the left half of p. 3, the author mentions anecdotes in his/her memory limit discussion. But it would be very suprising to find anecdotes in the main discussion of a psychology paper addressing memory issues. Perhaps the author can find in the psychological literature those kinds of evidence in favor of his/her position, which do count as scientific evidence as opposed to anecdotes.

Finally, given that the paper is likely to be published in the proceedings, I would strongly recommend that (i) the author use a spell-checker, and (ii) ask a reliable native speaker of English to proofread. For example, on p. 8, "t" was dropped from "against." And, although the author's English is good in general, many grammatical and conceptual mistakes are found. The grammatical mistakes include the omission of necessary articles, and some examples of the conceptual mistakes include the following. For example, in the first section of 1.4, the author says that the VMT "calls for" problems. But when something X calls for something Y, Y is something desired, while problems are headaches; the author should have said something like: the VMT "creates problems that call for solutions." Let me also say that the author's way of citing literature was somewhat irritating to me: for example, on p. 1, he/she says: # Note, incidentally, that here is another typo. However, in linguistics, psychology, and even in the format of the Association for Computational Linguistics, this should read: The family name goes between the parentheses only when the cited literature as a whole is within the parentheses, as in:


Barwise and Perry (1983). Situations and Attitudes. MIT Press.

Review 2

I believe that this paper has merits enough to be presented at PACLIC23. I think that the approach proposed in this paper does have better performance than the previous ones.

Review 3

No comments.