Weaknesses:

Q1.The paper could delve deeper into the details of how the CLIP and Transformer models are integrated.

R1: Thank you for your comments. We would added more enough detail on the process of how the CLIP and Transformer models are integrated as follows. In the feature extraction module, an image feature learning network that combines CLIP and Transformer with each other and a text feature learning network based on the attention mechanism of Transformer are used to obtain the feature representations of images and texts. R1: Thank you for your comments. More detail are added as follows.

Q2.Further explanation is needed to elucidate the advantages of the Graph Attention Network compared to other methods.

R2: Using GAT to predict labels have many advantages. Compared with labels during training that is static, predicted labels with GAT is dynamic, which can be updated according to specific tasks and data sets to model the relationship between labels explicitly and improve the representation of label features. In new samples, where there is no clear label, the label definition is not clear, or a sample contains multiple labels, the graph structure of GAT can help to infer the label of the sample, so as to obtain better inference and classification. Using a graph structure, GAT reduces the impact of label noise on training data by smoothing operations, while being able to predict labels and reduce label mismatches in original sample.

Q3.Some typos exist in mathematical formulas. For example: ① In Line 202, “U and V” should be “U and V” . ② "." should be "." in Eq. (14). ③ Many mathematical formulas lack necessary punctuations.

R3: Sorry for this, full text would be carefully checked to ensure that the typos in mathematical formulas. “U and V” would be changed with “U and V”, "." would be changed with ".".Punctuations of mathematical formulas would be added in our final manuscript.

Questions:

Q1. I suggest that the authors elaborate on the results of the experiments in the section of the results of the comparison of the data obtained in detail

R1: Thank you for your comments. More detail are added as follows. In order to further investigate the performance of our EGATH method, we used three evaluation metrics: mAP, PR curve and Top-k curve, as can be seen from the tables in the paper, we compared with the state-of-the-art seven methods on three datasets, the underlined values in the table are the best results among the compared methods, Table 1 illustrates the results of the mAP for the different lengths of the hash codes (16-bit, 32-bit, 64-bit). bit) of mAP, we can get the following observations Firstly, our method achieves better experimental results on the three datasets, with 3.4%, 3.8% and 12.4% improvement on MIR Flickr25K, NUSWIDE, and MSCOCO, respectively, compared to the best results DSPH and DCHMT among the compared methods, especially the effect on MSCOCO dataset has a significant enhancement, which is due to the following reasons The MSCOCO dataset with rich label information provides more contextual information for GAT, which effectively improves the accuracy of label prediction and the consistency of cross-modal features, and indicates that our method has a special advantage in dealing with data-rich datasets with diverse labels. Secondly, the corresponding PR curves can be seen in Fig. 2, from which we can observe that EGATH performs better and obtains higher precision scores at all recall levels. EGATH improves semantic differentiation and consistency across modalities through feature networks, and improves potential semantic associations through tag classification based on graph-attention networks, thus facilitating the cross-media retrieval performance. Finally, in Fig. 3 we compare EGATH with other methods under different Top-k search results, including 50, 100, 200, 500, 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, and 5000, and these experiments cover retrieval tasks ranging from small to large scale in order to evaluate the different methods at various retrieval scales. The experimental results show that the EGATH algorithm outperforms other existing algorithms for the cross-media retrieval tasks I→T and T→I under any top-K ordering, especially with higher MAP values for larger K values. This advantage is attributed to the fact that the EGATH method better preserves semantic information during the generation of hash codes and thus performs better in large-scale retrieval.

Q2. Could the graphs in Figures 2 and 3 in section 3.4 of the paper be improved?

R2: Sorry for this. We would improve Figures 2 and 3 in our final manuscript. The version that has been changed is in the PDF attachment.

Q3.The way of constructing matrices for labels mentioned in the paper increases the computational complexity, so in the method of this paper, constructing the adjacency matrix of labels, which involves steps such as thresholding, will this be more complex than the traditional method? And what is the difference of this approach?

R3: Constructing matrices for labels does increase computational complexity than traditional method, which is relatively small. However using GAT to predict labels have many advantages. Compared with labels during training that is static, predicted labels with GAT is dynamic, which can be updated according to specific tasks and data sets to model the relationship between labels explicitly and improve the representation of label features. In new samples, where there is no clear label, the label definition is not clear, or a sample contains multiple labels, the graph structure of GAT can help to infer the label of the sample, so as to obtain better inference and classification. Using a graph structure, GAT reduces the impact of label noise on training data by smoothing operations, while being able to predict labels and reduce label mismatches in original sample.

Weaknesses & Questions:

Q1.Some grammar, spelling, and symbol errors need carefully check.

R1: Sorry for this, we will carefully check to ensure that there are no grammatical, spelling, and symbol errors in the final version.

Q2. For the title, ‘Network Hash’ seems uncommon. Please check it

R2: Sorry for this,‘Network Hash’would be revised with ‘Hashing Network’ in the final version.

Q3. It is suggested to check the use of mathematical symbols. For example, it is difficult to distinguish vector, matrix, elements, etc in the paper.

R3: Sorry for this, we will carefully check carefully checked to ensure that the difference between vector, matrix and elements in the final version.

Q4. In (10), where is $i$?

R4: Sorry for this. $i$ is the foot mark of $\hat{S}$ in Eq.(9) and also $\hat{S}$ in Eq.(10).Correct form would be revised in our final manuscript.

Q5. It is suggested to add an Algorithm table to summarize the training/optimization process.

R5: Thank you for this insightful comment. Algorithm table has been added in PDF attachment with Table 2.

Q6. Figure 4 is too small to look.

R6：Sorry for this, Figure 4 would be revised in our final version. The version that has been changed is in the PDF attachment.

Weaknesses:

Q1. the paper does not provide enough detail on some specific components of the framework

R1:We would add more enough detail on the framework. we propose a new end-to-end cross-modal hash retrieval based on graph attention network, which contains three main components: feature extraction, GAT-based classification module and hash learning module. Firstly, in the feature extraction module, an image feature learning network that combines CLIP and Transformer with each other and a text feature learning network based on the attention mechanism of Transformer are used to obtain the feature representations of images and texts, and secondly, in the GAT-based tag prediction classifier, we integrate the co-occurrence information of tags into word embeddings, specifically, calculate the co-occurrence frequency of the labels and construct the adjacent matrix, take the initial feature vector and adjacent matrix of the labels as inputs, mine the implicit semantic information present in the labels through the dynamic attention mechanism, and compute the attention weights between the labels, so as to weight the average and update the embedding of the labels, and the updated embedding of the labels not only contains the information of the labels themselves, but also integrates the relational information between the labels, which is used to guide the feature learning of images and texts. Finally, the hash code loss function and optimum strategy enable the hash code to retain more semantic information, thus obtaining a discriminative and compact cross-modal uniform hash representation.

Q2. Some of the baselines could be added to this paper as appropriate

R2: We have added four new baselines. The experiment result shows that our proposed method has much better performance than the four baselines.

[1] Cross-domain transfer hashing for efficient cross-modal retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 2024

[2] Structures awarefine-grained contrastive adversarial hashing for cross-media retrieval. IEEE Transactions on Knowledge and Data Engineering, 2024

[3]Cross-modal generation and pair correlation alignment hashing. IEEE Transactions on Intelligent Transportation Systems, 24(3):3018-3026,2023

[4]Unsupervised multimodal graph contrastive semantic anchor space dynamic knowledge distillation network for cross-media hash retrieval. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), pages 4699-4708

Questions:

Q1. The authors need add some details about the graph attention network module, how to process the components.

R1: We would added the following detail in our final manuscript.

Firstly, the GloVe vectors and adjacency matrices of the labels are taken as inputs, and the information of the neighbour nodes is aggregated through the attention mechanism of GAT, which maps the input features to a new feature space by linearly transforming the input features using the weighting matrix. Secondly, each pair of labels is spliced and the attention score is computed to enable the model to assign higher weights to important node information, which in turn updates the feature representation of each label by weighted aggregation of information from neighbour nodes using the attention score. Finally, the updated label features are fused with the input features (image or text features), and the similarity is calculated by matrix multiplication to obtain the final category prediction.

Q2. In this paper, does the use of CLIP and Transformer and GAT models involve high model complexity? Does this affect the training performance of the model.

R2: Constructing matrices for labels does increase computational complexity than traditional method, while this doesn’t affect the training performance of the model. Actually, using GAT as a label classifier to predict labels can provide more context information. Moreover, combining the predicted labels with features of the extracted image and text can enhance the representation ability of feature representation, at the same time, improve the consistence and matching effect between images and text.

Q3.The author need add comparative experiments on Graph Attention Networks to further illustrate the importance of this module.

R3:We conducted further tests for this module in the added comparison experiments, specifically using the GASKN model from the 2024 ICDE conference. The experimental results show that the addition of the graph network module significantly improves retrieval accuracy, which validates the effectiveness of graph networks in improving cross-modal retrieval performance.

Weaknesses:

Q1. The contribution is not significant. Actually, some hashing methods have explored the use of CLIP, Transformer or graph attention network.

R1: Thank you for your comments. The main contributions of our work are as follows. By combining CLIP and Transformer technologies, an end-to-end architecture is implemented that significantly improves the model’s ability to capture global features of multi modal data, thus ensuring semantic consistency of images and text. Different from other existing works, where CLIP and Transformer are both used for text and image to extract features. While we utilize CLIP coupled with transformer to extract features for image data, and feature extraction via a transformer for text, which can realize lightweight network. Predicted labels is used to combined with the feature extracted from the feature modules of image net and text net to enhance feature representation, which also reduce the noise in the original labels during training. We utilize GAT as a label classifier to explore the hidden information in the label to predict labels, which can directly model the label graph to dig the correlation between labels and has higher flexibility than other label classification using preset weights for feature processing. Our EGATH was evaluated on three widely recognized benchmark datasets: NUS-WIDE, MIRFlickr25K, and MS-COCO, proving that it has obvious advantages in performance. Experimental results on these datasets show that our EGATH is superior to the current advanced cross-modal hash methods.

Q2. The results are not convincing. Only one method proposed in 2023 is used as a baseline for comparison. To my knowledge, many methods of this field are proposed in each year. The authors should compare the proposed method with more recent works.

R2: Thank you for your comments. We have added four new baselines, which includes the GASKN in 2024 ICDE, CDTH in 2024 IEEE TSVT, SCAHN in 2024 IEEE TKD, and CMGCAH in 2023 IEEE TITS. The experiment result shows that our proposed method has much better performance than the four baselines. More detailed are as follows.

Table 1 Additional experimental data

REFERENCE

[1] Fengling Li, Bowen Wang, Lei Zhu, Jingjing Li, Zheng Zhang, and Xiaojun Chang. Cross-domain transfer hashing for efficient cross-modal retrieval. IEEE Transactions on Circuits and Systems for VideoTechnology, 2024.

[2] Meiyu Liang, Yawen Li, Yang Yu, Xiaowen Cao, Zhe Xue, Ang Li, and Kangkang Lu. Structures awarefine-grained contrastive adversarial hashing for cross-media retrieval. IEEE Transactions on Knowledgeand Data Engineering, 2024.

[3] Weihua Ou, Jiaxin Deng, Lei Zhang, Jianping Gou, and Quan Zhou. Cross-modal generation and paircorrelation alignment hashing. IEEE Transactions on Intelligent Transportation Systems, 24(3):3018–3026,2023.

[4] Yang Yu, Meiyu Liang, Mengran Yin, Kangkang Lu, Junping Du, and Zhe Xue. Unsupervised multimodal graph contrastive semantic anchor space dynamic knowledge distillation network for cross-media hash retrieval. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), pages 4699–4708.IEEE, 2024.

Question:

Q1. The authors use GAT to predict labels. Actually, the samples have labels during training. Is it necessary for the model to predict the labels? In addition, from the framework, I cannot understand how the model makes use of the predicted labels.

R1: Thank you for your comments. Compared with labels during training, the predicted labels can explore the information hidden in labels during training to enhance feature representation. Moreover, labels during training have a lot of noise and mismatches, while predicted labels can effectively handle noise to avoid the mismatches. The model integrates the co-occurrence information from the labels during training into the word embeddings to form the predicted labels (text and images). Then, model combined the predicted labels with the feature extracted from the feature modules of image net and text net. The whole process can not only bridge the information gap between text and image but also enhance the deep characteristics of them.