1. Lack of comparison with recent existing SOTA methods in Table 1, e.g., PIMNet, PARSeq.
2. Lack of evaluation of model performance in actual real-world scenarios.
3. No mention of the ability to generalize to other languages.

1. In your training process, the initial predictions from your encoder are supervised, which means after many epochs of training, the encoder's predictions will become very close to the target, with only a few errors. In such a case, the decoder may only imitate the encoder's output and fail to learn anything. Did you encounter this issue, and how do you ensure that both the encoder and decoder are learning effectively?
2. As mentioned in the first point, your method calibrates the predictions after the initial prediction. However, the paper lacks the experimental results that show the difference between the initial predictions and the final calibrated predictions. If you can demonstrate that the encoder’s predictions had several errors but the decoder corrected them, it would prove the effectiveness of your two-stage method. Otherwise, your two-stage method may not fundamentally differ from a conventional end-to-end approach.
3. Your masking strategy is innovative but somewhat confusing to me. In previous works, masking is used to hide valid information to add challenges to the model’s predictions. However, your approach replaces the masked character with the ground truth, which, as I understand it, seems to merely reduce the learning difficulty in the early stages of training. The paper does not explain in detail why this method helps the model learn the relationships between characters, despite making such a claim.
4. A small question: if the encoder's predicted sequence length differs from the ground truth (which happens often), how do you align them when calculating the cross-entropy loss for the decoder’s predictions?

1）The progressive sampling proposed in this paper is, in strict terms, still a form of masking technique. It utilizes the self-attention interaction between character embeddings of predicted and unpredicted characters, which is not novel in the context of scene text recognition. 2) Regarding Section 3.2, "Decoupled Non-Autoregressive Decoder," there seems to be an issue with the attention formula in equation (3). It should be $Attention(q,k,v)=softmax(\frac{qk^T}{\sqrt{d_k}})v$. Additionally, the authors mention that $qk^T$ can be expressed as $c_{q}^{T}X + p_{q}^{T}X_p$. However, what purpose does the multiplication of the two position matrices serve, and what is the significance of performing cross-attention between them? 3) Although the authors describe their model as relatively simple, the model structure and training procedures suggest otherwise. It involves a two-stage training process, introduces additional character positional information, and requires prediction outputs at each decoder layer. This approach incorporates a considerable amount of extra information, such as character positional data and a pre-trained CTC decoder for initializing inputs to the non-autoregressive decoder. From this perspective, the proposed method does not seem to be a straightforward model, despite achieving significant performance improvements over other non-autoregressive approaches. 4) While the authors compare their method with numerous text recognition techniques in Table 1, there is a lack of comparison with more recent text recognition technologies. It is unclear when the authors wrote this paper, but the latest method they compare against appears to be a paper published in 2022, making such experimental comparisons less convincing. Moreover, the authors only conducted experiments on some existing datasets, without addressing more challenging datasets like Union14M-Benchmark in this paper.

The paper presents some interesting ideas, and the results appear competitive. However, the presentation requires significant improvement. In its current form, the paper is barely readable, and certain sections require educated guesses to fully comprehend.

For example, Section 3 does not clearly describe the proposed method. Figures 1 and 2 lack explanations for the notations used, leaving readers to infer meanings. Additionally, many mathematical symbols are used without clear definitions—such as the domain for $x$ and $y$ in Equation 1, or the exact definition of $PE(x, y)$. Furthermore, some equations lack rigor and seem arbitrary in notation; for instance, in Equation 5, the term $\log p(y_t | y_{<t}, X; \theta)$ should use a more precise notation, such as $y_1, y_2, \ldots, y_{t-1}$, and $y$ itself should be defined.

• The authors will use the center coordinates of characters to encode the position information, so it seems that the character-level annotations will be involved. It may burden the annotations of the ground truth.
• The novelty of this paper is somewhat incremental. It seems that the proposed method introduces a progressive learning strategy and embeds the position of the character center into queries based on a non-autoregressive paradigm. In effect, some curriculum learning from easy examples to hard examples is explored, while some hard example learning strategies are also proposed. I think the authors should illustrate the advantage of the proposed progressive learning strategy compared with similar learning strategies on hard examples.
• It is a common operation that adds character-level position information to the queries. The authors should illustrate the difference between the re-designed decoder structure and the existing decoder networks. It seems that the authors adopt the core techniques of the decoder in the scene text spotting to the scene text recognition.
• The authors do not report the parameters of the proposed network.
• In Table 1, the authors only compared the methods before 2022. Some latest works are missing.

• The related works compared in this study are all from before 2022, lacking comparisons with the latest research, such as:

  [1] OTE: Exploring Accurate Scene Text Recognition Using One Token, CVPR 2024
  [2]  Multi-modal In-Context Learning Makes an Ego-evolving Scene Text Recognizer, CVPR 2024
  [3] Bridging the Gap Between End-to-End and Two-Step Text Spotting, CVPR 2024
  [4] SPTS v2: Single-Point Scene Text Spotting,TPAMI 2023
  

• This approach requires character and character position annotations, which are often unavailable in real-world data. Especially for irregular text, precise character positioning is difficult to obtain and costly to annotate. Addressing this limitation would improve applicability.

• The datasets chosen for comparison are relatively simple, with high accuracy already achieved by many methods. To better validate the method’s effectiveness, more challenging datasets, such as Total-Text or Union14M [5], could be considered. [5] Revisiting Scene Text Recognition: A Data Perspective, ICCV 2023

1. Limited Exploration of Trade-offs in Progressive Sampling: The current ablation study outlines improvements from progressive sampling, but it lacks details on how varying the sampling strategy impacts training convergence time, model complexity, and potential overfitting on challenging characters.
2. Insufficient Comparison with Broader State-of-the-Art Non-Autoregressive Techniques: Although the paper compares NAText with several autoregressive and non-autoregressive models, the non-autoregressive comparison group could be expanded to include more recent advancements in non-autoregressive Transformers and efficient decoding strategies beyond the baseline models presented.
3. Ambiguity in Efficiency Gains for Real-World Applications: While the model demonstrates significant speed improvements in experimental settings, it is unclear how these gains translate to various real-world applications, particularly in resource-constrained or edge computing environments.
4. Positional Query Design Clarification: Although the paper introduces a redesigned positional query with character center encoding to enhance 2D feature aggregation, the impact and implementation details of this component could be clarified.
5. ack of Qualitative Analysis and Error Case Discussion: The paper would benefit from including qualitative examples or visualizations of both successful and challenging cases for NAText.