Masked Reviewer ID: Assigned_Reviewer_2 Review: Question Overall rating Accept Reviewer expertise Some familiarity Paper strengths (2-3 brief points) Knet provides useful building blocks for efficiently implementing standard machine learning algorithms. The concise implementations for various algorithms provide performance competitive with other existing frameworks. The gradient descent method used in Knet allows for applications to provide code in an already existing high-level language. Paper weaknesses (2-3 brief points please) The paper does not clearly describe the limitations of the approach taken by Knet. Are there loss functions used in practice which cannot be handled by autograd? Are there algorithms used in practice which do not use gradient descent? The paper does not clearly state what functions are implemented or supplied by Knet. Is the grad function used in Section 2 just the same grad function supplied by autograd in Section 3.1? Detailed Comments How are the KnetArrays used by an application? Are the code segments in Section 2 using KnetArrays, or would the caller of "train" just supply arguments of type KnetArray? It would be useful to see larger scale benchmarks in Section 2.6. In addition, these benchmarks should describe whether a CPU or GPU is being used, which models of those, with how much memory, and which versions of the various pieces of software were tested. The code examples in Section 2 are clear and concise, which is very good, but some of the details of how the functions behave may be lost to readers unfamiliar with Julia (e.g., the array slicing syntax). In particular, the loss functions in 2.5 could be explained in a little more detail. The paper could be more clear if ordered slightly differently. Section 3 describes the functions provided by the knet library, which are used by the example implementations in Section 2. It may make more sense to swap this order. In 3.2, you mention that KnetArrays use a custom memory manager to reuse garbage collected pointers. This may be an interesting detail to elaborate on and evaluate the effects of (though in a short paper you may be space limited). Also in 3.2, you state that the paper examples run up to 17x faster on a GPU than a CPU. This number needs to be contextualized with the model of the CPU and the GPU you are using. Masked Reviewer ID: Assigned_Reviewer_3 Review: Question Overall rating Weak reject Reviewer expertise Some familiarity Paper strengths (2-3 brief points) The automatic differentiation approach seems like it could be an interesting mechanism, assuming the set of programs it works on are well characterized. It's nice to see higher level constructs applied to traditionally low-level environments like GPUs. Additionally, the code segments look like something a reasonable person could write. Paper weaknesses (2-3 brief points please) The contribution and motivation are unclear. It reads like the major contribution is the porting of an automatic differentiation module to Julia, along with some syntactic sugar over a CUDA library, though I assume there is more to it than that. Though the automatic differentiation technique looks promising, blackbox program transformations are slightly fraught in that they can fail on otherwise reasonable programs if they encounter a construct the author did not expect. Detailed Comments Overall, I'm slightly confused about who the paper's target audience is. Based on the examples in the paper, Knet targets those who wish to design their own models, i.e. people with a fairly sophisticated ML background. Yet for those much of the code in section 2 is redundant, a compelling example would simply take one non-trivial model and compare it against the next best framework. (Caffe is probably not the best best example to compare against, the decision to use a declarative rather than algorithmic approach is usually more fundamental than the choice of which library gets used). Furthermore, it's ambiguous what is actually being benchmarked; footnote 1 says the testbed uses AWS gpu instances but lines 262-263 imply that only CPU code was actually tested. Considering the importance of GPU targets this needs to be addressed preferably with a GPU comparison. Additionally, since TensorFlow is designed to function in a distributed setting comparing against it invites questions as to Knet's scalability. Since such a comparison is likely necessary, it may be a good idea to head off the issue by determining at what size dataset such scalability is actually beneficial. Also, since the benchmarks use different models, it would be good to have an indication that Knet based ones maintain rough parity for test data as well; as is there's no indication that they provide correct results. Additionally while automatic differentiation looks nifty, a cursory glance at the python Autograd's documentation implies the technique only works for a limited subset of the language's features. For a blackbox program transformation to be useable it really requires clear distinction between which language features work and which don't, ideally one that is automatically checked. Lacking this forces your users to retain a deep knowledge of which constructs are valid and which are not, and intuit when incorrect outputs are not due to their own bugs, but because of a mismatch between what they wrote and what the Knet backed can handle. It may be that you're already providing this, but if so that should be mentioned in the paper. Similarly, If you are achieving auto-vectorization, either through clever use of operator overloading, or from Julia's toolchain, this is a significant achievement and needs to be made explicit. As written, I cannot how much effort, if any, is required to translate between the example models and performant versions.

============================================================================ COLING 2016 Reviews for Submission #383 ============================================================================ Title: Learning grammatical categories using paradigmatic representations: Substitute words for language acquisition Authors: Mehmet Ali Yatbaz, Volkan Cirik, Aylin Küntay and Deniz Yuret ============================================================================ REVIEWER #1 ============================================================================ --------------------------------------------------------------------------- Reviewer's Scores --------------------------------------------------------------------------- Relevance: 2 Originality: 2 Technical correctness / soundness: 2 Readability and clarity: 2 Meaningful comparison: 2 Substance: 2 Impact of ideas: 2 Impact of resources: 1 Overall recommendation: 2 Reviewer Confidence: 3 --------------------------------------------------------------------------- Comments --------------------------------------------------------------------------- This paper suggests that performance of the prediction of a syntactic category of a word in CHILDS corpora performs better with the model of the pattern with a slot (i.e. described as "a * b" in the article) than using a model separating the pattern heads ('aX') and the tails ('Xb') (i.e. described as "aX+Xb"). The performance of "a*b" being better than for "aX + Xb" in prediction seemed obvious to me in the presented experiment, since as in Figure 1, the information used for prediction in the former is larger than for the latter. In this sense, the comparison performed in this paper is not a "fair comparison". Maybe I misread, but if so, then authors must make sure to convince the readers that they are conducting a fair comparison. Therefore, I did not understand what claim is made in this article and my judgement with respect to COLING presentation unfortunately has to be reject. Apart from this main point, there are controversial issues related to this article. 1. The authors claim that "a*b" corresponds to "paradigmatic" representation, whereas "aX+Xb" to "syntagmatic", but uses of these two terms had better be reconsidered. In linguistics, paradigmatic and syntagmatic has the origin in F.de Saussure's definition. Calling "a*b"/"aX+Xb" as paradigmatic/syntagmatic seems too wide. 2. The authors refer to EMNLP-coNLL paper (the last reference), which has similar keywords in the title. If this refered paper is the authors' paper, then since the review process must be double blind, the authors should disclose the author names. If not, it must be shown what exactly is new of this paper compared with the reference. 3. The authors must explain more about their exact intent of this work. Do the authors want to engineer the CHILDES text? Then, the authors must explain the application needs and situate their work among previous work in the domain of engineering. Do the authors want to say from this paper, that children must be processing grammar through "a*b"? Then, the authors must explain the relation of their experiment how the experiment could relate to a cognitive claim. 4. Above all, I wonder whether this paper is appropriate for a presentation in the conferences of ACL community. Majority of the references are in the cognitive scientific domain. If behind this paper lies a cognitive question, then, it had better that they seek for a discussion in a conference dedicated in cognitive science domain. ============================================================================ REVIEWER #2 ============================================================================ --------------------------------------------------------------------------- Reviewer's Scores --------------------------------------------------------------------------- Relevance: 5 Originality: 3 Technical correctness / soundness: 4 Readability and clarity: 5 Meaningful comparison: 4 Substance: 3 Impact of ideas: 3 Impact of resources: 1 Overall recommendation: 4 Reviewer Confidence: 4 --------------------------------------------------------------------------- Comments --------------------------------------------------------------------------- ## General comments This paper presents an interesting study about the use of paradigmatic presentations for word category learning in the setting of natural language acquisition. The authors compare two models of word representations -- syntagmatic (where each word is represented based on its neighbors), and paradigmatic (where the word is represented using substitute words). The experiments presented here clearly show the benefit of paradigmatic representations, and this is a very interesting contribution to the field. However, I'm having a hard time understanding the dichotomy presented between these two representations: both seem to rely on context-based substitutability criteria, since even the paradigmatic presentation constructs the substitute words via use of n-grams which in turn are created by looking at the neighboring words of each target word. A related point is that there is a slight misrepresentation of Harris's distributional hypothesis in the example presented in the Introduction: in Harris's view, substitutability arrives through the use of all (total) environments of the words/morphemes; it is therefore "unfair" to compare the substitutes which have been created from a very large collection of environments (n-grams) with the local context of these two sentences. ## Specific issues/comments The first sentence of your introduction would find an entire section of the field in disagreement, namely researchers who subscribe to the constructivist approach who believe that children learn rules about individual words and only later form abstract syntactic categories. After the example presented an introduction, you mentioned that "the high probability substitute reflect both semantic and grammatical properties". As many papers in the distributional semantic network area have shown, this is not necessarily the case especially if you select a very small window from which you create your context representations (something which applies to the n-grams you use in this work), in which case they mostly reflect syntactic properties. How do you justify the claim that the Redington et al. method "lacks completeness"? There are a few missing references from the part of speech induction literature that would be interesting to include: Alex Clark's 2003 "Combining distributional and morphological information for part of speech induction", Christodoulopoulos et. al 2011 "A Bayesian Mixture Model for PoS Induction Using Multiple Features" and Blunsom and Cohn's "A Hierarchical Pitman−Yor Process HMM for Unsupervised Part of Speech Induction". In the input corpora that you used, did you remove the punctuation marks? They tend to provide a very strong distributional cue and are sometimes explicitly removed to better simulate the linguistic input of children. It is unclear what the contribution of Figure 1 is. Is it theoretically better or worse to have more units as input to the model? Is there any evidence to support these theoretical claims? Given the size of the training corpus for the n-gram model, I would suggest that the comparison to the syntagmatic models which are trained on significantly fewer words seems problematic. I would've liked to see a per-tag graph of the number of substitutes for each target word, since I would think that for some cases of word classes (interjections) this number would return a lot of noise. It would be better if you describe the connection between the results that you get in this work and those of St. Clair et al. 2010. ============================================================================ REVIEWER #3 ============================================================================ --------------------------------------------------------------------------- Reviewer's Scores --------------------------------------------------------------------------- Relevance: 4 Originality: 4 Technical correctness / soundness: 4 Readability and clarity: 4 Meaningful comparison: 3 Substance: 3 Impact of ideas: 3 Impact of resources: 1 Overall recommendation: 4 Reviewer Confidence: 3 --------------------------------------------------------------------------- Comments --------------------------------------------------------------------------- This paper describes a comparison of two approaches to POS tagging, one (syntagmatic) in which tags are assigned based on ngrams, and one (paradigmatic) in which tags are assigned based on possible lexical substitutions. The paper reports the latter (paradigmatic) gives higher accuracy. Overall it's an interesting result, but I have some reservations about how generally the results may be interpreted. First, if this model is to be taken as an acquisition model, as Section 1 seems to describe, I'm surprised the model uses supervised learning. Second, given this model is a supervised POS tagger, the accuracy results seem rather low compared to state-of-the-art sequence model taggers. This is potentially a problem because the gains due to paradigmatic modeling may not turn out to be significant in the presence of a more robust syntagmatic baseline. Minor comments: The last term in equation 3 should be "P(w_{n-1}|w^{n-2}_0)" (the zero is missing). Figures 2 and 3 should use bar graphs, since the slopes of lines between columns isn't interpretable. "patter(n)s" p. 5. You may want to cite work on syntagmatic and paradigmatic parser acquisition by Simon Dennis at U. Adelade in Australia.

COLING 2016 review: Title: CharNER: Character-Level Named Entity Recognition Authors: Onur Kuru, Ozan Arkan Can and Deniz Yuret ============================================================================ REVIEWER #1 ============================================================================ --------------------------------------------------------------------------- Reviewer's Scores --------------------------------------------------------------------------- Relevance: 5 Originality: 3 Technical correctness / soundness: 3 Readability and clarity: 4 Meaningful comparison: 4 Substance: 3 Impact of ideas: 3 Impact of resources: 3 Overall recommendation: 4 Reviewer Confidence: 4 --------------------------------------------------------------------------- Comments --------------------------------------------------------------------------- The paper proposed character based NER for languages with word segmentation. The character-based tagging is proposed previously. Their contribution is to include LSTM models for the character-based tagging settings. However, the result is fair for the targeted languages. In my opinion, the method may be promising for the languages without word segmentation such as Chinese and Japanese, since the word segmentation error affects the score of NER. ============================================================================ REVIEWER #2 ============================================================================ --------------------------------------------------------------------------- Reviewer's Scores --------------------------------------------------------------------------- Relevance: 5 Originality: 3 Technical correctness / soundness: 4 Readability and clarity: 4 Meaningful comparison: 4 Substance: 3 Impact of ideas: 3 Impact of resources: 1 Overall recommendation: 4 Reviewer Confidence: 5 --------------------------------------------------------------------------- Comments --------------------------------------------------------------------------- This paper presents a character-based model for named-entity recognition based on bidirectional LSTMs. There is very recent research that tries to do something similar, in NAACL-16 (Lample et al.) and ACL (Ma and Hovy, for example) with the exact same motivation: remove external resources, such as gazetteers and use a character-based approach to achieve high results. This should not invalidate the paper, though. The main difference with previous research (mentioned above) is that this model examines a sentence as a sequence of characters and outputs a tag distribution for each character. They later use transition matrices that only allow tags consistent with the word. The results are nice, however they are not the best at all. They are very good compared to systems that do not use external resources including word embeddings, however it should be a requirement to report results provided by the other systems without external resources (see Lample et al. for example) In Table 6, you present results for Ma and Hovy and Lample et al. and you include them in the "External" row, as far as I know they only use word embeddings (if they do)... I think that you should incorporate "word embeddings" to the caption, otherwise readers might think that they use gazetteers. Some missing references: Two EMNLP-15 papers that presented interesting results for tagging, parsing and language modeling by using character-based embeddings. - Wang Ling; Chris Dyer; Alan W Black; Isabel Trancoso; Ramon Fermandez; Silvio Amir; Luis Marujo; Tiago Luis Finding Function in Form: Compositional Character Models for Open Vocabulary Word Representation -Miguel Ballesteros; Chris Dyer; Noah A. Smith Improved Transition-based Parsing by Modeling Characters instead of Words with LSTMs This paper is also worth mentioning, since it also produces an entire character sequence for sentences: Bhuwan Dhingra; Zhong Zhou; Dylan Fitzpatrick; Michael Muehl; William Cohen Tweet2Vec: Character-Based Distributed Representations for Social Media Minor comment: Lample et al. do some more than LSTM-CRF, they also presented a shift reduce algorithm that exploits character-based embeddings. ============================================================================ REVIEWER #3 ============================================================================ --------------------------------------------------------------------------- Reviewer's Scores --------------------------------------------------------------------------- Relevance: 5 Originality: 3 Technical correctness / soundness: 4 Readability and clarity: 5 Meaningful comparison: 4 Substance: 4 Impact of ideas: 4 Impact of resources: 1 Overall recommendation: 4 Reviewer Confidence: 3 --------------------------------------------------------------------------- Comments --------------------------------------------------------------------------- Very interesting work. It clearly shows that a deep bidirectional LSTM architecture combined with a Viterbi decoder effectively finds language specific features for NER. Applying the algorithm to Languages written without space characters, e.g., Chinese and/or Japanese, may be interesting. ================================= EMNLP 2016 review: Title: CharNER: Character-Level Named Entity Recognition Authors: Onur Kuru, Ozan Arkan Can and Deniz Yuret Instructions The author response period has begun. The reviews for your submission are displayed on this page. If you want to respond to the points raised in the reviews, you may do so in the box provided below. The response should be entered by 17 July, 2016 (11:59pm Pacific Daylight Savings Time, UTC -7h). Response can be edited multiple times during all the author response period. Please note: you are not obligated to respond to the reviews. Review #1 Appropriateness: 5 Clarity: 4 Originality: 3 Soundness / Correctness: 4 Meaningful Comparison: 5 Substance: 4 Impact of Ideas / Results: 4 Impact of Accompanying Software: 3 Impact of Accompanying Dataset / Resource: 1 Recommendation: 3 Reviewer Confidence: 5 Comments This paper presents a named entity recognizer in which the entire sentence is encoded as a sequence of characters, and a bidirectional LSTM is used to make predictions. Unlike previous (and recent) approaches, such as Lample et al. 2016 that presented character-based representation of words and then an LSTM/bidirectionalLSTM/stack-LSTM on top of that. This model is similar to the tweet2vec model recently accepted at ACL2016, even though it tries to solve a different task. They examine a sentence as a sequence of characters and outputs a tag distribution for each character.This model,as Lample et al., has the potentialities of being language independent and they apply it cross-lingually. The motivation and goals are also similar to Lample et al. (remove external features such as gazzetteers etc, and still achieve high results) Figure 3 does a great job summarizing the entire paper. In order to avoid things like J o h n w o r k s P O O O G G G G O they use a decoder as Wang et al. 2015, that applies a transition matrix, at the end they output the entire sequence. They make a good comparison with related work, but since some of the models are freely available, I'd expect that the authors run them in the languages without public results (such as Arabic or Turkish) In table 5 you should definitely differentiate between systems that use gazzetteers, and neural models that use pretrained word embeddings. They are not the same thing and how it is presented it might confuse the reader. This is an interesting paper, but it lacks a bit of novelty given all the previous work that already demonstrated the usefulness of characters and sequential models for NER. Minor comments: Missing ref (?) in related work.

Title: Transfer Learning for Low-Resource Neural Machine Translation Authors: Barret Zoph, Deniz Yuret, Jonathan May and Kevin Knight Instructions The author response period has begun. The reviews for your submission are displayed on this page. If you want to respond to the points raised in the reviews, you may do so in the box provided below. The response should be entered by 17 July, 2016 (11:59pm Pacific Daylight Savings Time, UTC -7h). Response can be edited multiple times during all the author response period. Please note: you are not obligated to respond to the reviews. Review #1 Appropriateness: 5 Clarity: 4 Originality: 4 Soundness / Correctness: 4 Meaningful Comparison: 4 Substance: 4 Impact of Ideas / Results: 4 Impact of Accompanying Software: 1 Impact of Accompanying Dataset / Resource: 1 Recommendation: 4 Reviewer Confidence: 4 Comments Neural machine translation (NMT) performs much worse than phrase-based or syntax-based statistical machine translation (SMT) for low-resource language pairs. This paper adopts transfer learning to improve the low-resource NMT. An NMT model is first trained on a rich-resource language pair, then transfer some learned parameters to the low-resource NMT model. In this way, the low-resource NMT obtained comparable performance as SMT. Although the idea is simple, experiments on four low-resource language pairs proved its efficiency. Moreover, a detailed and nice analysis is included in the paper, discussed the effect of the different rich-resource languages, fixing different part of the model, learning curve, etc. The section about key idea (section 3) lacks a bit of clarity, I understand the whole idea after read section 4 and 5, maybe you need to make it clearer what you exactly do transfer learning for NMT in section 3. A small question is related with this, do you use the same learning rate to train the parent model and child model? If not, how do you set the learning rate? Review #2 Appropriateness: 5 Clarity: 5 Originality: 3 Soundness / Correctness: 4 Meaningful Comparison: 4 Substance: 4 Impact of Ideas / Results: 3 Impact of Accompanying Software: 1 Impact of Accompanying Dataset / Resource: 1 Recommendation: 3 Reviewer Confidence: 4 Comments The paper presents a transfer learning approach for low-resource NMT. The idea is extremely simple: training a full-scale parent model is used to initialise a child model with limited resources. The approach seems to be surprisingly effective and the authors convincingly present empirical proof for the use of this method. The final model is still below the performance of the baseline syntax-based SMT model in three out of four language pairs but outperforms the baseline when used for restoring n-bet lists. It seems surprising that the randomised mapping of word embeddings to the low-resource language word types actually works well enough. It would be interesting to know how much they are changed during training and how much they differ from training without parent model (if that can be quantified in some reasonable way). In the analyses part, the authors present a est in which they improve the mapping with dictionary-based assignment which sounds much more intuitive. But the use of this initialisation is minimal and seems to fade out completely in later training epochs. I'm still puzzled why there shouldn't be any effect of more appropriate mappings. Is there a problem with the dictionary-based assignments? Another question is whether this approach would also work the other way around. Could you fix the source language embeddings when translating from English into low-resource languages and would you expect the same kind of improvements? Did you already try this and would you have anything to report in the reverse direction? This would be more interesting then some of the ablation tests (like 5.2 which seems to argue for the same as 5.1) In any case, the discussion about the effect of parent language pair is limited as this does not systematically evaluate the effect using many language pairs and language families with various kinds of relationships. I have also a question about the syntax-based model used as baseline. What is the target syntax you are using in your string-to-tree model and how does it compare to state-of-the art settings? 26 BLEU for the English-Frech system sounds low to me but what is the test set and how does it compare to state-of-the-art? Another question would be about evaluations beyond BLEU scores. What does happen to the actual translations and what do humans think about them? The limits of automatic evaluation metrics are well know and the authors should discuss translation quality in other terms as well. Review #3 Appropriateness: 5 Clarity: 5 Originality: 3 Soundness / Correctness: 4 Meaningful Comparison: 5 Substance: 4 Impact of Ideas / Results: 4 Impact of Accompanying Software: 1 Impact of Accompanying Dataset / Resource: 1 Recommendation: 4 Reviewer Confidence: 3 Comments The paper presents a method for transferring parameters from an NMT model learned on a large data set (fr-en) to a low resource translation task with the same output language (ha-en, tr-en, uz-en and ur-en). Part of the transfer process is selecting which parameters to fine-tune to the low resource translation task. Although I'm not convinced that the presented method is the best way to leverage the larger data set in a low resource translation setting, I strongly believe that this method needs to be presented to the community. In the prose you say that French' would be a good parent language for French, and I agree; in the experiment it seems like you flipped them so that French is the parent language, and French' is the child language. I don't know whether this makes any difference in your model at all, but please be consistent. In Table 7 you report ablation tests, measuring model perplexity and BLEU on the dev set. I'm left with the question of how these two quantities correlate with BLEU on an unseen test set. This question comes up again in the dictionary initialization experiments (Figure 4), where perplexity is again reported, but there is no way to tell how well it correlates with translation performance. Minor comments s3p2: "employing w a separate" -> "employing a separate" Fig1: the colors are really hard to see on a regular office laser printer print-out. Fig2: I'm guessing that eg "Source Word" and "Source Word" are different source words, and that they may be close to each other, but please use proper indexing to clarify exactly how.

Title: Why Neural Translations are the Right Length Authors: Xing Shi, Kevin Knight and Deniz Yuret Instructions The author response period has begun. The reviews for your submission are displayed on this page. If you want to respond to the points raised in the reviews, you may do so in the box provided below. The response should be entered by 17 July, 2016 (11:59pm Pacific Daylight Savings Time, UTC -7h). Response can be edited multiple times during all the author response period. Please note: you are not obligated to respond to the reviews. Review #1 Appropriateness: 5 Clarity: 5 Originality: 4 Soundness / Correctness: 5 Meaningful Comparison: 5 Substance: 4 Impact of Ideas / Results: 3 Impact of Accompanying Software: 1 Impact of Accompanying Dataset / Resource: 1 Recommendation: 5 Reviewer Confidence: 4 Comments This paper convincingly explains why NMT translations are the right length; they show the mechanism (that there are components in the vectors that keep track of length during encoding and decoding) explicitly with a very clear toy example, as well as the tendency in a real-world task. The paper is admirably clearly written and very accessible. Please keep in mind that some people still print papers out, and stick to a gray scale legible coloring scheme (blue and red are indistinguishable after the gray scale dimensionality reduction has been applied). s1p1: "covert that vector in a target sentence." -> "convert that vector into a target sentence." Review #2 Appropriateness: 5 Clarity: 4 Originality: 4 Soundness / Correctness: 4 Meaningful Comparison: 4 Substance: 4 Impact of Ideas / Results: 3 Impact of Accompanying Software: 1 Impact of Accompanying Dataset / Resource: 1 Recommendation: 4 Reviewer Confidence: 3 Comments This paper investigates the question of how neural MT models manage to produce output in the right length. Starting with a smal toy problem and proceeding to an actual neural MT system, the paper shows that (groups of) specific cells are "dedicated" to encoding sentence length. The paper is refreshingly different from many other NMT papers I've seen lately, in that it attempts to understand what's going on within the neural model, thus addressing a point of criticism that is occasionally brought forward against neural approaches to MT: that they are black boxes and no-one seems to care about what's going on inside. The paper is well written and generally easy to understand. What I'm missing most is a good motivation for addressing this question. What do we gain from knowing that neural models explicitly encode length? Also, is this behavior consistent across neural models? What happens if we start model training with different random initializations? A few minor quibbles: - The images in Figure 2 are two small. You've got plenty of room left (6 pages max.!) to provide bigger images. - Figures 4 and 5 should be tables, not figures. Review #3 Appropriateness: 5 Clarity: 5 Originality: 3 Soundness / Correctness: 5 Meaningful Comparison: 4 Substance: 4 Impact of Ideas / Results: 4 Impact of Accompanying Software: 1 Impact of Accompanying Dataset / Resource: 1 Recommendation: 4 Reviewer Confidence: 4 Comments One of the issues with NMT is the apparent opacity of the model. It is hard to know what is going in inside the block box. The authors start to peel back the curtain here by investigating the question of how NMT models output target sentences of the right length. By looking at both a toy auto-encoder with 4 units and a regular NMT model, they provide interesting insight and show that the model has one of a small handful of units devoted specifically to the token counting task. The paper is well written & easy to follow and the insights will be of interest to the EMNLP audience.