decisionLetter.rtf

{\rtf1\ansi\ansicpg1252\cocoartf1404\cocoasubrtf470
{\fonttbl\f0\fswiss\fcharset0 Helvetica;}
{\colortbl;\red255\green255\blue255;\red255\green255\blue0;}
\margl1440\margr1440\vieww16840\viewh14500\viewkind0
\pard\tx720\tx1440\tx2160\tx2880\tx3600\tx4320\tx5040\tx5760\tx6480\tx7200\tx7920\tx8640\pardirnatural\partightenfactor0

\f0\fs24 \cf0 24 Jul 2016\
\
Dear Dr. Goodman,\
\
Ref.: Ms. No. TICS-D-16-00152\
\
Thank you for your manuscript "Pragmatic language interpretation as probabilistic inference" and for your patience during the review process. I have now heard from two expert reviewers, whose reports you can find at the end of this letter.\
\
As you can see, both reviewers were quite positive about your submission and recommended acceptance after minor revision. From an editorial perspective, your manuscript is well written and was a pleasure to read. I therefore have few editorial comments to add to those of the reviewers', which I have listed at the end of this letter (just before the reviewers' comments).\
\
While you are performing revisions on your article, I also wanted to let you know of an opportunity for your article to offer a more engaging learning experience to our readers. As you might know, Trends reviews are valuable teaching tools and are heavily used in the classroom. Authors can further enrich and extend the content of their article by using the online "Interactive Questions" tool at http://authortools.elsevier.com/iq/verification. The tool allows you to create multiple-choice questions (usually 3-4) based on your review that will be displayed in an interactive quiz next to your published article. We encourage you to consider using this tool to maximize the educational value of your review.\
\
To get more information about the tool and see examples please go to the "Extra Features" tab here: http://www.cell.com/trends/cognitive-sciences/authors. To learn more about "Interactive Questions" please go to https://www.elsevier.com/books-and-journals/content-innovation/interactive-questions. If you have any questions or would like additional information about "Interactive Questions" please email tics@cell.com.\
\
As I am hoping to include your revised article in an upcoming issue, I would be very grateful if you could return the revised version of your manuscript by *07 Aug 2016*. I would also be grateful if you could acknowledge receipt of this letter and confirm the suggested resubmission date by sending a brief message to tics@cell.com.\
\
I would like to thank you once again for your article. If you have any questions, please don't hesitate to get in contact. I look very much forward to receiving and reading your revised manuscript.\
\
With all best wishes,\
\
Becca\
\
---------\
Rebecca Schwarzlose, PhD\
Editor\
Trends in Cognitive Sciences\
---------\
\
EDITOR'S COMMENTS:\
\
Noah and Mike - thank you for this thoughtful and beautifully written piece. Please consider the reviewers' comments and suggestions regarding the piece but also consider the need to keep the piece concise and focused. Revisions should be guided by the goal of improving the piece and not by appeasing reviewers. If you have any questions, please do let me know.\
\
I believe that the reviewer's request to discuss potential drawbacks of the RSA/uRSA approach at least briefly is a good one. \
\

\i I extended / revised the second paragraph of the discussion to address this. hopefully that\'92s enough because we don\'92t have room for more.\

\i0 \
\
In the absence of a discussion of alternate approaches or potential shortcomings of this approach, the article currently reads more as an Opinion than a Review. Although we have the option of converting to an Opinion in the process of the revision, I think the piece would be more useful and impactful if this topic were at least briefly addressed. \cb2 Please ensure that the revised main text does not exceed 4,150 words and 130 references\cb1  (if a Review) or contact me to discuss if this is insufficient.\
\
Please also remove embedded figures from the resubmitted manuscript and submit them as separate files. Figure captions should be included in the main manuscript file after the reference list.\
\
Please also note that if you keep any pre-print postings as cited sourses, these should be cited as Resources rather than References - requiring a separate Resources list at the end of the manuscript and a separate numbering/citation system indicated with superscript numbers. I would be grateful, however, if citation of pre-print postings can be minimized or eliminated entirely.\
\
Finally, I have entered minor editorial comments directly on a marked-up copy of your manuscript. You can access this copy in the following manner:\
\
i.Log into the Trends in Cognitive Sciences website http://tics.edmgr.com/\
ii. Go to "Submissions needing revision" which will take you to your submission.\
iii. Select "Action links" then "File inventory".\
The marked up copy is labelled "TiCS manuscript - editor's comments". The comments can be displayed by selecting the track changes function.\
\
Please let me know if you have any questions about the revision or anything else. Thank you again for the excellent article. I look forward to receiving the revised version and seeing it in the pages of TiCS!\
\
All the very best,\
Becca\
\
\
GENERAL INFORMATION FOR RESUBMISSION:\
\
\
1) Please ensure that your revised manuscript is accompanied by a separate list of all the major changes made and a detailed response to the comments made by the referees and the editor. For any comments that you did not address through modification of the manuscript, please explain why.\
\
\
2) When resubmitting your manuscript, please provide us with high resolution (at least 300 dpi), editable versions of your figures.\
If your figures were prepared in powerpoint, please send us the .ppt/.pptx file.\
High resolution, editable files are necessary for article production and we cannot proceed further without them.\
\
\
3) Please note that it is your responsibility to obtain permission to reproduce copyrighted material (i.e. figures, tables or excerpts that have been published online or in print) from the publishers of the original material. If you are uncertain how to obtain permission, please contact the editorial office.\
\
\
4) Please ensure that you acknowledge all sources of funding that have supported this work (in a separate 'Acknowledgments' section), as we are unable to publish errata for funding acknowledgment post-publication of your article.\
\
\
Once you have completed your revision and wish to submit your revised article:\
1. Please go to http://tics.edmgr.com/ and then "Submissions needing revision"\
2. Click on the "Action link" and then "Submit revision". Please submit a Word file of your text (i.e. not a pdf) including all box texts and figure legends, together with a document outlining the major changes made and your responses to the referees comments.\
IMPORTANT NOTE: IF THE TEXT OF YOUR MANUSCRIPT IS SUBMITTED IN ANY OTHER FORMAT OTHER THAN .DOC, .RTF (OR LATEX INCLUDING THE SOURCE FILE), YOUR SUBMISSION WILL BE RETURNED TO YOU AND THIS WILL DELAY SIGNIFICANTLY THE PROCESSING OF YOUR ARTICLE.\
3. All figures should be submitted as separate files and not embedded in the text.\
\
\
----\
\
REVIEWER COMMENTS:\
\
\
Reviewer #1: This is a very well-written, accessible and balanced overview of the rational-speech-act approach to language interpretation. It clearly introduces the main ideas and nicely leads to the state-of-the-art and near-future challenges of the approach. I recommend publication, but would offer a few suggestions for ways in which the paper might be improved here and there.\
\
* From the abstract: "they explain how complex phenomena like hyperbole and vagueness can arise." -> This formulation could be misleading, as it may suggest an account of why human languages have such things as hyperboles or vagueness. In my understanding, the RSA approach focuses on interpretation of hyperbolic or vague terms, but offers only limited insight into why, say, communication with non-hyperbolic or non-vague means should not be expected to drive out the hyperbolic, vague and less precise talk.\
\

\i done\

\i0 \
\
* Page 2, 3rd paragraph: "... when violated" -> This is a bit sloppy. Some distinguish two kinds of implicatures: based on conformity with maxims (e.g., scalar implicatures, ad hoc implicatures in the interpretation of referring expresssion, etc.) and based on intentional and ostensive flouting of maxims (e.g., in sarcasm or irony). Maybe, to stay clear of all this, just say that "maxims are involved in the computation of implicatures"?\
\

\i reworded as suggested\'85\
\

\i0 \
* Page 2, 3rd paragraph again: In a recent like-minded overview article, Franke & J\'e4ger (Probabilistic Pragmatics, Zeitschrift f\'fcr Sprachwissenschaft, 2016) argue explicitly that RSA-style pragmatics is to be welcomed because it does away with maxims. Their paper defends RSA-style pragmatics against current standard approaches in linguistic pragmatics and shares many of the conceptual points raised in favor of RSA-style approaches in this manuscript. Possibly, this review might benefit from engaging with or acknowledging this paper.\
\

\i added cite at end of second section.\

\i0 \
\
* Page 4, before or around the introduction of the literal listener: The paper motivates the reasoning architecture in terms of "avoidance of an infinite recursion". It may be even more illuminating (for some readers) to point out that even if we wanted "infinite recursion" (e.g., for common knowledge or equilibrium), we would like to have a way of bringing semantic (conventional) meaning to bear on pragmatic reasoning. This was (historically) the main motivation for iterated best response models in the game-theoretic pragmatics tradition. So, starting with a literal listener can also be motivated by the desire to anchor pragmatic reasoning in the semantics, even if we then loop ad infinitum. In other words, pragmatic reasoning is assumed, by all of these models, to start from the (possibly purely hypothetical) assumption that speech has a conventional meaning and that this is the starting point for deliberating about what a speaker has meant on some occasion. Mentioning\
this perspective may add to the conceptual (not mere-technical) motivation of RSA-style approaches.\
\

\i done.
\i0 \
\
\
* Page 5, "to explain the communicative use of sentence fragments and prosodic stress (Bergen et al., 2016)" -> The authors should really know much better than me, but this kind of thing can easily go unnoticed. So, please double-check and don't be offended if I ask whether the reference to this paper is the right one in this context. As far as I remember, said paper does not deal with sentence fragments or prosodic stress. Other work by Leon Bergen (and co-authors) does, as far as I know.\
\

\i done.\

\i0 \
\
* Page 6, first paragraph that ends here: If space permits, it might be useful for the reader if the paper was a little more explicit about how RSA-models relate to (quantitative) data. While the "link hypothesis" might be totally obvious for those who have seen RSA-based papers or presentations, it might otherwise be missed. So, on page 6 it might help to just state that the production probabilities defined by P_S on page 3 correlate well with the average answers from the task. Similarly, for the listener.\
\

\i done.\

\i0 \
\
* Page 8: The caption of figure 2 could possibly do with a little more explanation as to who the nice picture is to be interpreted. E.g., what's exactly happening inside the thought bubbles of the literal listeners? Where is a state in which both affect and prize needs to be communicated (as mentioned in the text)?\
\

\i done, i think.\

\i0 \
\
* Page 9, the discussion of the hyperbole model: It did not become sufficiently clear what drives the explanation of the hyperbole effect. Is it only the fact that $1,000 are implausibly expensive? Or is it additionally an assumption that, given such a ridiculous price, the speaker will want to talk about affect (because of some obvious emotional reaction)? The latter would probably have to be implemented in the production probabilities, but then it would be much less of an elegant explanation. My fading memory of the model is that it is only the former. If so, the review might stress this fact and make it more clear what the mechanism is, in intuitive terms.\
\

\i done, more or less. could use another pass?\

\i0 \
\
* Page 9: "Linguistic descriptions, especially adjectives, are both context-sensitive and\
vague." -> Maybe hedge to avoid ruffling of feathers? For example: "Many ... descriptions, especially certain adjectives, ..."? Some hardliner might say that "married" is neither context-sensitive (beyond a trivial temporal parameter) nor vague.\
\

\i done.\

\i0 \
\
* Page 9-10: The paragraph on 'embedded implicatures' misses an explication of what the joint inference is about, i.e., in this case uncertain lexical meanings.\
\

\i done.\

\i0 \
\
* Page 10, last paragraph on the page on "processing": In my opinion, this was noticeably the weakest part of the paper. There is no (strong) argument here that "RSA potentially suggests a new take on pragmatic language processing", or at least I do not see it. When the paragraph concludes that "RSA is consistent with modern psycholinguistic theories that emphasize interactive, incremental processing" that may be so, but consistency is still a far shot from deriving actual processing predictions from RSA. Take one of the central findings from the cited Degen & Tanenhaus paper: subjects who choose to say that "some" is false in an "all" situation take longer to reach that decision than those who choose to say true. How would RSA explain that? -- Now, I do firmly believe that RSA can make interesting predictions about processing, but it is not as easy and clear as this paragraph suggests. I would recommend a more careful reformulation, possibly integrating "processing" in the\
box on "outstanding questions" possibly in relation to "algorithms".\
\

\i i tried to revise this paragraph to make a less strong claim, while still describing the different spirit of processing RSA inspires.
\i0 \
\
\
* Page 18 on utils in case of speaker uncertainty: The paper makes it sound as if we now look at "expected utilities" while previously (in the main text) we did not. This might be confusing since the paper speaks of "the utility [the speaker] expects to gain" on page 3. Maybe adapt the formulation? Actually, I believe that both the standard RSA-model and the Goodman-Stuhlm\'fcller model can be defined in terms of the same (non-expected) utility function based on KL-divergence between speaker and listener belief. If the speaker knows the true state, that reduces to surprisal, if not, it reduces to the Goodman-Stuhlm\'fcller utilities up to an additive constant, which then cancels out under the soft-max function, so that the resulting utterance probabilities are equivalent.\
\

\i i like our formulation in this paper better.\

\i0 \
\
Reviewer #2: This review looks at recent work applying the Rational, Bayesian Probabilistic (RBP) approach to issues that have been traditionally addressed in Pragmatics, following on work by Grice. Perhaps among the many domains that RBP applications have been explored, Gricean pragmatics is among the most promising. This is so since it is widely (though not universally) assumed that effects of meaning in conversation are derived by assuming that interlocutors are rational and that the speakers have commonly known overarching goals, like being co-operative, or relevant. This specific review mostly focuses on work in the Rational Speech Act framework, pioneered by the co-authors. Given the progress being made among the group who work in this framework, I feel that a TiCS review is timely. However, I have some critical comments and suggestions which are set out below.\
\
1.      Improve critical example: The key  illustration for the RSA approach (relating to Fig. 1) is not particularly convincing. The descriptions used in the experiment are not naturalistic. Typically, one would say, 'just glasses' rather than, 'glasses', and maybe also, 'glasses and hat' for the character with hat. In the absence of context that severely limits the use of words, the experimental results do not clearly generalise to the everyday cases of quantity implicature (or specificity implicature). It would help to replace this or find an additional example. I think that should not be too difficult.\
\

\i not going to change example, but added a sentence addressing the restriction on utterances.\

\i0 \
\
2.      The paper is written by two key players using the RSA approach and cites work mostly by people that work in a close knit group around themselves. This is all fine by me since the approach is new. However, because it is relatively new, there has been little time for critical discussion of that approach from outside of the RSA group to emerge. The upshot is that the authors are left with formulating their own critical commentary. To an extent they've done a good job with that (but see 3. below). They could however address critical discussion of the general RBP approach  in this review, in the absence of published critical work on RSA theory. E.g. Jones & Love (BBS, 2011) and Marcus and Davis (Psych Science, 2013).\
\

\i don\'92t want to cite those, because they are stupid. but should acknowledge the alternatives more clearly. (ok, fine, i cited jones and love\'85.)\

\i0 \
\
On this score, I personally think it is worth highlighting that the RSA framework is capable of modelling different kinds of account of pragmatic phenomena and provide a basis on which we can compare their effectiveness. This point is already kind of there in the review, where the authors acknowledge that different assumptions about speaker utility can be made. However, the authors seem to imply that all aspects of speaker utility mentioned (also in the box) would be in play at all times. But the pragmatics literature contains arguments for or against certain of the mentioned aspects. To give one example, approaches to scalar implicature differ as to whether informativeness relative to a conversational purpose is key (a la Grice or the neo-Griceans), whether a comparative notion of relevance (a la Ben Russell or Sperber & Wilson) is key or their own idea of 'specificity' - i.e.  informativeness to the observed state of affairs is right. When combined with, say, economy,\
these assumptions make different predictions (relative to the way that economy is implemented in the model).\
\
In sum, the authors' review suggests that the RSA model just needs to incorporate ever more assumptions about speaker utility and also speaker uncertainty in order to account for the data. But it seems clear that different combinations of assumptions yield different predictions. In effect, the RSA approach is really a framework for comparisons among sets of proposals. As has been said before in the critical work mentioned (e.g. Jones & Love), when this is recognised, this is the strength of RBP approaches, otherwise it can be perceived as a weakness.\
\

\i i really agree with this point. tried to work it in to the GD\'85
\i0 \
\
\
3.      The problem of alternatives: Following on from the above point, the authors cryptically mention the problem of how alternative utterances are computed in the outstanding questions box. This is, or closely relates to, the so-called Symmetry Problem for Grice and his followers. I feel this perhaps needs a bit more space, maybe even a box of its own. The SP is a great challenge to Grice and one that has no clear solution, in spite of quite a lot of work. Bergen et al. (2016) claim to have a solution to SP but they don't really. First, they allow arbitrary sets of alternatives. Second, even if all possible alts are allowed into a model of the context, there are outstanding problems - such as those described by Trinh & Haida in a recent paper. The SP is one motivation for many linguists to give up on the Gricean program for at least scalar implicatures and look to non-Rationalist approaches (such as the grammatical approach of Chierchia, Fox and colleagues). Thus I feel it\
would be a more representative review if the authors could acknowledge the problem in more detail and maybe sketch the prospects of a solution from within the RBP framework.\
\

\i i think this is too technical to address in detail here\'85. \cb2 maybe we can at least acknowledge ongoing discussion of SP somewhere?\cb1 \

\i0 \
4.      Comparison with non-Rationalist approaches. As mentioned, I think this review could do a good job of showing how the RSA framework can provide a space for comparing and evaluating broadly Gricean approaches to pragmatic phenomena. However, I feel that some space could be given over to a mention of non-Rationalist approaches,  such as the Grammatical Theory of scalars (see point 3). What others are on offer? How are comparisons to be made?\
\

\i yes. added a bit in GD. \

\i0 \
5. Maybe better use of the boxes could be made. The box on referential expressions relates to the first point above. I contains an assumption about the felicity of the examples related to figure 1. These utterances are not felicitous except  in special circumstances. Maybe change or remove this box?\
\

\i i don\'92t really understand what is being suggested, so i\'92m ignoring this.\

\i0 \
\
\
[Other things from becca\'92s file, to be dealt with:]\
\
\pard\tx720\tx1440\tx2160\tx2880\tx3600\tx4320\tx5040\tx5760\tx6480\tx7200\tx7920\tx8640\pardirnatural\partightenfactor0
\cf0 go through becca's highlighted terms and add to glossary or simplify.\
\

\i did this. i simplified several places. didn\'92t add too many to glossary.
\i0 \
\pard\tx720\tx1440\tx2160\tx2880\tx3600\tx4320\tx5040\tx5760\tx6480\tx7200\tx7920\tx8640\pardirnatural\partightenfactor0
\cf0 \
\
\cb2 Please also remove embedded figures from the resubmitted manuscript and submit them as separate files. Figure captions should be included in the main manuscript file after the reference list.\
\cb1 \
\cb2 get box numbering right.\
\cb1 \
\cb2 trends and OQ as separate files.\
\cb1 \
}