Summary and Contributions: The paper proposes a new benchmark (gSCAN) for learning compositional, grounded natural language understanding and establishes baselines on 8 different tasks.
Strengths: Excellent work building on the SCAN benchmark, exhibiting appropriate understanding of natural language semantics, as well as appropriate ways of evaluating the performance of ANNs on NLU tasks.
Weaknesses: Given the current state of the art of the sub-field, there are no significant weaknesses. Obviously, the formal semantics sub-field of linguistics has systematically investigate many constructions in English and many other languages in a grounded and compositional way, and gscan is still very far from that.
Relation to Prior Work: Yes.
Summary and Contributions: The authors introduce a new benchmark, named gSCAN, for evaluating compositional generalization in situated language understanding. Their main contribution is the design of test sets that require different forms of linguistic generalization.
Strengths: 1. The authors introduce a new benchmark grounded SCAN (gSCAN) and evaluate eight types of compositional generalization in the states of a grid world accessible to the agent. 2. Among the eight different generalization splits, they compare the baseline and GECA methods with sequence matching accuracy as an evaluation metric. It is concluded that it is challenging to account for common natural language generalization phenomena with standard neural models, and that gSCAN could be used as a fruitful benchmark for developing models with more human-like compositional learning skills. 3. It will be of great help to the improvement of the generalization ability of the subsequent models in understanding natural language. This work also has a relatively large impact on the research work of the NeurIPS community in this area.
Weaknesses: 1. Although the author has done a comparative experiment between the multi-modal and the GECA methods on the proposed benchmark, the design and analysis of the experiment are confusing and difficult to understand. 2. Besides, there are lots of inconsistencies in the experiment section, such as the title of Table1 claim that “Models fail on all splits except C and F”, which is inconsistent with the data in rows C and F. The poor performance of the two methods on most split datasets could explain that advances are needed in neural architectures for compositional learning. 3. However, I am not quite sure why the author can draw the following conclusions based on the experiment and analysis of split A-I: artifacts in SCAN are not central to the nature of compositional generalization, gSCAN removes these artifacts by introducing more sophisticated semantics through grounding. I think that seems a little bit unconvincing.
Correctness: The author’s conclusion in the paper seems to be reasonable, and the empirical methodology is also correct in the general direction. However, there are some problems in the experiment design, which make the experiment analysis section difficult to understand, and the experimental results are also unconvincing to support the drawn conclusions.
Clarity: The writing of this paper is good and native, but there are logical inconsistencies, especially in the design and analysis of experiments. This makes it difficult for readers to understand, even if they are relatively familiar with the field.
Relation to Prior Work: The authors mention enough related works in the paper. The most important previous work is related to the SCAN benchmark and Good-enough compositional data augmentation (GECA) method. The author argues that SCAN lacks grounding, which severely limits the variety of linguistic generalizations it can examine. In addition, authors believe that GECA is unreliable in more realistic, grounded language understanding. And the difference between this work and the previous works is also clearly explained.
Additional Feedback: 1. I think that the authors should give more detailed description about the grid work of navigation-related tasks, which may be not familiar to most readers. 2. What does SD in the title of the Table 1 mean? 3. What does the <SOS> in the Figure 3 mean? 4. There are some subtle errors in Figure 3, such as the output of Bi-LSTM and Conv. Net should be connected to c_j^c and c_j^s respectively, right? 5. In Table 1, I think the Baseline mean the multi-modal NN method, and GECA corresponds to the previous method, right? Thus, the GECA should be the baseline method, right? 6. Are there some other works which also use the multi-modal NN to model this action sequence prediction task? If so, could you compare to those works? 7. I think language understanding is hard to define for the computer. Can sequence match accuracy be used to measure the ability of language understanding?
Summary and Contributions: This paper presents a benchmark for systematic generalization in grounded language understanding, which is an extension from SCAN with grounding to the object knowledge and states. The main contribution of this work is on the multiple new aspects of generalization into the benchmark in addition to the previous configurations. The experimental results show that both the baseline and a SOTA method on the SCAN have limitations in the newly introduced generalization aspects in this proposed benchmark.
Strengths: - This work presented multiple research problems that haven't been addressed by previous work on compositional generalization. - Very detailed discussions were given with the experimental results.
Weaknesses: - There is no significant contribution to the model itself. - The scope of the work is narrow restricted only to the SCAN environment. - Not convinced if the comparisons between just two models are enough to support the main claims of this work.
Correctness: High-level concepts make sense to me.
Clarity: I think this paper could be much further improved in terms of its clarity especially for readers who are not familiar with its previous work. I was also struggling to understand the details before I read the original SCAN paper.
Relation to Prior Work: This work cites many related studies with clear descriptions.
Summary and Contributions: Summary: the authors propose a new task that focuses on syntactic generalization that is grounded in the states of a grid world. The new benchmark facilitates new development for learning generalization through linguistically motivated rules. Contributions: This paper offers an interesting combination of compositionality and contextuality in natural language understanding. Previous work such as SCAN focuses mostly on compositionality aspect without external grounding to a state of the world. Through a grid world, the authors combine these two aspects for a grounded SCAN benchmark (gSCAN).
Strengths: The proposed benchmark extends the previous SCAN benchmark into a multimodal setting with a grounding state of the world. This new setting is challenging and might render previous methods less effective. For example, the permutation approach proposed in  might not be applicable in gSCAN because it would alter the action sequence. gSCAN is, hence, more realistic and requires stronger reasoning and generalization capabilities.
Weaknesses: One weakness of the paper is the synthetic setting of the dataset, e.g. 2D grid world, which is quite unrealistic to real world environment. Another limitation is the fixed hold-out attributes of test splits. For example, split B is limited to test whether the model can ground 'yellow squares' only and split C is limited to 'red squares' only. Therefore, the test results might be affected by data distribution bias in these specific hold-out attributes in the training split.
Correctness: I am a bit confused about the difference between split E and split C. In both splits, at training time, the model does not see a specific novel combination of attributes ("red square" for split C, and "small circle" for split E) and at test time, the model is challenged to ground these attributes. Is the difference between two splits established by color combination (split C) and size combination (split E)? Could you give a potential explanation on why GECA performs well in split C but not in split E?
Clarity: The paper is well written with clear motivation, technical details, and data construction process. The empirical results also show clear details of performance breakdown.
Relation to Prior Work: The authors describe previous work related to compositional generalization tasks. They explained the shortfalls of previous methods when applied into the new benchmark gSCAN with a grounding environment. It would also be useful if the authors can discuss in more detail (probably in a separate paragraph) on work that focuses on contextual generalization (one example would be ) and how the new benchmark is different from current work in this line of research. Since grounded/ contextual generalization is an important contribution of the work, similar related work should be highlighted more.
Additional Feedback: Is the difference between two splits established by color combination (split C) and size combination (split E)? Could you give a potential explanation on why GECA performs well in split C but not in split E? Web link in Reference  is not working.