Stroke-Based Autoencoders: Self-Supervised Learners for Efficient Zero-Shot Chinese Character Recognition

Round 1

Reviewer 1 Report

I reviewed an interesting paper related to 'OCR' for Chinese characters entitled "Stroke-Based Autoencoders: Self-Supervised Learners for Efficient Zero-Shot Chinese Character Recognition". Some comments and considerations.

1. The paper is overall well written, both from the methodological point of view as well as from the clarity of presentation.  The introduction section is quite detailed. I would only ask for celery outlining the main contributions of the paper at the end of the introduction.

2. Compared to introduction-related works is a bit weak. Not that it is not acceptable; however, there is room for improvement.  For instance, The authors do provide how they differ from the related works, however, they do not clearly mention what the issues with the related works are and what impact their approach (or the way they design their algorithm) has or could have on the issues raised.

3. Was there a dataset used? Usually, authors provide at least a paragraph on the nature of the dataset used.

4. Methodology and results are clear and well-written. The discussion is also well written. However, as with all works, there are limitations. I would encourage the authors to outline a couple of them. This will also make the future directions less artificial.

5. There is a line with future directions, but it is quite artificail. I would encourgae for the authors to outline some more tangible, related to what they wish to do with this approach and kwnoledge they have generated.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The paper describes a composite method with self-supervised pre-training for chinese characters embedding, and stroke-based recognition which allows to build a zero-shot recognition method. The method described by the authors outperforms the previously described methods of this class.

The article is well structured, the description of the method is quite comprehensible and the presentation of the results is clear.

My main content suggestion would be to mention total size of the proposed modesl (in terms of number of coefficients), as the issue of the models size and computation time is not discussed at all in the paper, thus at least if the authors provide the model sizes the reader would be able to get a rough idea without having to completely reproduce the work.

On page 9 (first paragraph) when discussing the results of seen characters recognition (presented in Table 2), the authors write "our SAE still out-perform other methods in a consistent way" - that is not, strictly speaking, true, since the authors' method did not outperform some character-based methods (given in references 45-47). It is better to write more clearly, that the method outperformed all stroke-based or few-shot methods, and some character-based methods, staying in the same general accuracy level. The fact that the proposed method did not outperform all character-based methods does not detract from the paper in any way.

I would suggest the authors do a native speaker's proofreading of the paper to avoid language mistakes, however the required editing is minimal, since the paper is easy to read and understand. Some typos I found during the reading are at the bottom of the comments.

Overall I think that the paper is well-written, the authors clearly present their results and contribution, and it would be valuable and interesting for the advancement of the field.

Some small typos I found:

  - Page 1, introduction, "characters that have not been appeared" -> "characters that have not appeared".

  - Page 1, introduction, and in Page 3, right column, in one of the references list the reference 14 is given twice.

  - Page 2, left column, "aotuencoder"

  - Page 3, right column, "structure remains to explore" -> "structure remains to be explored"

  - Fig. 3, "Tuning" - probably meant "Turning"

  - Page 4, "’口’ (month)" - probably meant "mouth"

  - Fig. 5 caption, "deocder"

  - Page 9, first paragraph, "our SAE still outperform" -> "out SAE still outperforms"

  - Minor formatting issue: sometimes the quotation marks are misused (closing quotes are used as opening ones)

  - Minor formatting issue: some spaces missing between the references and words. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx