Self-learning Compositional Representations for Zero-shot Chinese Character Recognition

1. To further validate the superiority and scalability of the proposed method, we recommend conducting Chinese character recognition experiments in a non-zero-shot setting.

2. The experimental results indicate that as the size of the training set increases, the performance of the model experiences a slight decrease. It might be beneficial for the paper to explore and provide further insights into this phenomenon.

3. The paper notes that models trained on Chinese characters have shown promising performance in Korean and Japanese. However, it would be beneficial to include comparative experiments with other methods to further demonstrate its superiority.

4. Minor issues. In Eq.7, \Lambda^k is not defined or explained. Furthermore, the number of slots on the x-axis in Figure 5 should be expressed in integer units.

Weaknesses. Despite the promising experimental results, several key issues require further clarification:

1. In the slot-matching stage, SCR-CCR needs to prepare 10 example images for each character class to be used as slot information for this class to help with similarity matching. However, it significantly compromises zero-shot learning (ZSL) principles, as unseen character classes should not have accessible example images. Thus, the fairness of the comparative experiments is questionable.
2. SCR-CCR matching mechanism resembles template-based methods, relying on auxiliary example images. This approach may lack practicality in real-world settings where many uncommon character classes exist, and obtaining enough example images can be challenging. In contrast, template images are more easily synthesized by computers.
3. In Section 4.2, Character Zero-Shot Setting, the authors indicate that the proposed method depends on learning compositional representations based on character differences, which echoes the issues mentioned in the second point.
4. In the last sentence of the second paragraph of the Introduction, the authors state that "....these methods may lack the flexibility to adapt to different scenarios and have poor generalization in practical applications." To support this claim, additional zero-shot character experiments in natural scenes, such as on the CTW dataset, would be needed to demonstrate the proposed method's practicality and generalization over prior approaches.

In the abstract, the authors claim the issues of radical based method and stroke based methods, however, I do not find enough analysis about this in the experiments, I hope the authors can compare and highlight these points in the experiments.

This paper follows the idea from object-centric representation, and leverage the slot-attention module in the zero-shot Chinese character recognition. Hence, I do not find any specific novelty in this area.

The claim that this learned structure of characters in an interpretable way, however, it seems the method can only detect different slots but no the combination order. Hence, how to know the structure of this character is not clear.

The experiments seem show the effectiveness of the method, but more details and analysis should add, like the following questions.

The paper is not written very clear, like the following questions.

1. The authors conduct zero-shot CCR experiments using the 500-2755 training dataset. However, it raises the question of whether training on all existing labeled Chinese character data could yield better results in a true zero-shot scenario. If traditional radical-based, stroke-based, or character-based methods perform well in such cases, is the introduction of an object-centric approach necessary? This is a significant setting issue in the current zero-shot CCR field, as genuine zero-shot capability should involve fully utilizing available open-source data and then addressing the unseen scenarios.
2. The motivation behind the paper is questionable. Both radical-based and stroke-based methods are forms of object-based approaches. While the proposed object-based method allows the model to learn to extract objects autonomously, it is unclear whether this method is truly superior and this motivation is unexplainable.
3. During the zero-shot inference phase, are the example images printed or of other seen styles (seen during the pretraining stage)? In Figure 2, it appears that handwritten characters of the same style as test image are used.
4. A critical concern is that the innovation in the proposed method rely on slot attention. Would a standard attention mechanism achieve similar or even better results? This aspect is crucial for an ablation study, as the interpretability of the current object-centric learning paradigm is relatively poor.
5. In the HWDB experiments presented in Table 1, the performance decreases as the amount of training data increases. This result raises doubts about the validity of the experiment and the method.
6. Why is the CCR-CLIP performance in this table significantly lower than what is reported in the original paper, which achieved results around 90?
7. In Table 3, comparisons with state-of-the-art methods from Japanese and Korean studies should be included, especially considering the low accuracy for Korean. Additionally, it needs to be demonstrated whether the proposed method is effective solely for Chinese or also achieves optimal results in other languages.