Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. The paper proposes a behavior–rule inference model based on a hyponymy–hypernymy knowledge tree integrated into K-BERT. The introduction of a tolerance mechanism and the multi-label classification framework shows clear innovation and contributes meaningful insights to knowledge-enhanced reasoning models.
2. The improvements over the baseline K-BERT, including extended context length, comparative studies of different knowledge graphs, and the use of a Mask-Transformer mechanism, are technically sound and logically well justified.
3. The experiments cover multiple aspects such as loss-function comparison, hyperparameter sensitivity, and the influence of different knowledge trees. The results are consistent with the theoretical expectations and demonstrate the robustness of the proposed approach.
4. Some related work such as [FRT-DETR: faster real-time end-to-end detector for industrial surface defects based on transformer] should be introduced to improve the literature review.
5. The manuscript is well organized, following a logical progression from theoretical foundations to experimental analysis. Figures and explanations are detailed, making the technical workflow easy to follow.
6. The proposed model has promising applicability in domains such as legal document reasoning and behavioral decision support. The study expands the scope of integrating knowledge graphs with pretrained language models in domain-specific reasoning tasks.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

In computer science and mathematics, clear definitions are essential for following arguments. Therefore, it is important to define all key formal and technological concepts, even if they are well-known among specialists in natural language processing (NLP) or BERT models. This includes terms such as "gloss," "CLS," and "SEP."

In the text, particularly around equations 3.1 and 3.2, the methods of formal languages can assist in formalizing string manipulations, which represent the content of the Directed Acyclic Graph (DAG) and the spanning tree.

It is not clear whether transitioning to a spanning tree would result in information loss. Has this been measured or investigated? Is the potential information loss considered acceptable?

Equation 3.4, 3.5, 3.6:

 

It can be confusing that indices are sometimes presented in superscript and other times in subscript. It is likely that this notation does not have a specific meaning, as in the case of tensor calculus.

The definitions for the expression and its components, "𝑄𝑖+1, 𝐾𝑖+1, 𝑉𝑖+1," are not available. Therefore, Equation 3.4 cannot be interpreted in a mathematical sense. The right side appears to be numeric, but the left side is not defined, making the notation unclear.

Additionally, Equation 3.5 uses indices in superscript format, while Equation 3.4 uses them in subscript format.

"𝑊_𝑞 , 𝑊_𝑘 , and 𝑊_𝑣 are trainable model parameters."

It is unclear which trainable parameters are involved. Are they the weights in the nodes of the neural network or hyperparameters?

"ℎ_𝑖 represents the hiding state; however, the concept and variable are not defined. Is it a numeric parameter or a vector? How is it calculated?"

"m is the field of view matrix computed from the observation layer."

Equation 3.3 defines the element of M (supposedly) a matrix.

It is assumed that m=M in Equation 3.5.

Equation 3.5. The left-hand side seems to be a numeric value, the value of the i+1 dimension of the vector outputted by the softmax vector function. However, the right side cannot be determined. M is a matrix and added to an undefined structure. Matrix addition can take place by elements between matrices, but the left side of the addition sign is an undefined structure. Matrix addition produces a matrix, dividing by a numerical parameter (√d_k)  produces a matrix, softmax input argument is a vector or a real number, numeric parameter..

The softmax function can accept either a vector or a real number that represents the value of the ith dimension of that vector, provided it is defined and formulated for each dimension.

In the case Equation 3.5, probing the dimension of the argument of the softmax function fails.

Neither d_k nor T are defined

T is defined and used in (https://assets.amazon.science/96/81/c7cbc852412084faa3b51ab87132/integrating-noisy-knowledge-into-language-representations-for-e-commerce-applications.pdf?utm_source=chatgpt.com,

Samel, K., Ma, J., Wang, Z., Zhao, T., & Essa, I. (2023). Integrating Noisy Knowledge into Language Representations for E-Commerce Applications. 2023 IEEE International Conference on Big Data (BigData), 548–553. https://doi.org/10.1109/BigData59044.2023.10386615)

 

There are no citations and copyright information for Figures 2, 3, and 4. Even when using Creative Commons licensing, accurate citations are required.

The inaccurate mathematical formulation makes doubts that the implementation of the algorithm would be correct. 

For this reason, the displayed experiment cannot be considered convincing.

The mathematical fundamentals and proposed engineering approaches must be accurate, and the utilized literature and sources should be cited correctly.

The implemented algorithms should be demonstrated through pseudocode to verify the correctness of the underlying mathematical approach.

Syntax:  

 

Minor errors and typographical mistakes, such as "O(n^2)" and "(n-1)^2." Note the designation of exponents. 

"sample "[CLS]vi is a vj[SEP]" is constructed as an input
to the model for one relation between the words v_j and v_j." --- Indexes!

"ℎ_𝑖 is the hiding state of the ith
mask self-attention block".  Is it a superscript or a subscript?

 

Comments on the Quality of English Language

Some sentences should be improved to make it clearer.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper applies K-BERT with knowledge trees to legal rule inference, achieving 78.85% accuracy on divorce cases. While technically competent, it falls short of journal standards due to overclaimed contributions, narrow experimental scope, and weak analysis of its core knowledge integration component.

1. You’re essentially doing hyperparameter tuning on existing architecture. Extending context from 256→512, swapping CE for BCE loss, and using RoBERTa instead of BERT aren’t research contributions, they’re engineering decisions. What’s genuinely novel here?
2. Testing on one self-built dataset (652 training samples) from a single legal domain is insufficient. Where are the cross-domain experiments? Comparisons with modern legal AI baselines?
3. Table 4 shows your domain-specific knowledge graph barely outperforms generic HowNet (83.7% vs 81.7%), and no knowledge at all sits right in the middle (82%). If knowledge injection is your main contribution, why such marginal gains? What knowledge actually helps, when does it hurt, and why doesn’t domain expertise matter more?
4. Your “tolerance”metric is poorly defined and non-standard. How does dynamic tolerance differ from Precision@K or F1? Why not use established metrics?
5. Single runs with no error bars, confidence intervals, or significance tests won’t cut it. Where’s the reproducibility information? Training time?
6. Figures 1-2 are generic architecture diagrams that add nothing.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I have no further concerns.

Author Response

Response 1: We thank the reviewer for this positive feedback. We have taken the opportunity to further improve the manuscript based on the general suggestions provided in the evaluation. Specifically, we have:
Enhanced the introduction to include more background and relevant references.
Clarified the research design and methodology.
Improved the presentation of results and conclusions.
These revisions are detailed in the "Questions for General Evaluation" section above and are marked in the revised manuscript.
4. Response to Comments on the Quality of English Language
Point 1:"The English could be improved to more clearly express the research."
Response 1: We have thoroughly revised the manuscript to improve the English language quality. The text has been polished for clarity, grammar, and fluency by a professional language editing service. Key changes include:
Restructuring complex sentences for better readability.
Correcting grammatical errors and inconsistent terminology.
Ensuring precise and academic phrasing throughout.
All revisions are highlighted in the re-submitted manuscript.   

Reviewer 2 Report

Comments and Suggestions for Authors

Mathematical denotation problems: 
"of no more than O(n2)" incorrect formula.

" (n-1)^2 negative relations" is not according to the mathematical standards.

Equations 3.4,3.5,3.6, were copied from the paper without citations:

"Liu, W., Zhou, P., Zhao, Z., Wang, Z., Ju, Q., Deng, H., & Wang, P. (2020, April). K-bert: Enabling language representation with knowledge graph. In Proceedings of the AAAI conference on artificial intelligence (Vol. 34, No. 03, pp. 2901-2908)."

"𝑊_𝑞, 𝑊_𝑘 , and 𝑊_𝑣 are trainable model parameters" is copied without explanation in this paper on how the training will happen.

Q _i+1, K_i+1, V_i+1, are not defined (There is one K as notation for the Knowledge base). It is dubious that, without an exact mathematical formulation, the underlying data structures can be implemented correctly, especially the softmax function. 

Figure 3., 4. . 5. , 6. are copied from the paper without permission for copyright: 
Liu, W., Zhou, P., Zhao, Z., Wang, Z., Ju, Q., Deng, H., & Wang, P. (2020). K-BERT: Enabling Language Representation with Knowledge Graph. Proceedings of the AAAI Conference on Artificial Intelligence, 34(03), 2901-2908. https://doi.org/10.1609/aaai.v34i03.5681

The experiment is not designed clearly. The stated case study is based on corpora of divorce cases. The classification is not defined for the assessment of the results.

In the case of natural language processing, the human should be in the loop to evaluate the goodness or accuracy of the end result. 
The mentioned metrics can be applied in various fields of machine learning; however, their effectiveness in natural language processing is questionable. While they may measure model performance, they are not suitable for system verification or validation.

 

Comments on the Quality of English Language

Some sentences should be improved to make it clearer.

Author Response

Response 1: We thank the reviewer for identifying these mathematical notation issues. We have corrected all mathematical expressions throughout the manuscript:
Changed "O(n2)" to proper mathematical notation "O(n²)"
Corrected "(n-1)^2" to "(n-1)²"
Conducted a full review of all mathematical expressions to ensure consistency with standard notation
These corrections appear in Sections 3.2, 3.4, and 4.3 of the revised manuscript

Comments 2:  "Equations 3.4,3.5,3.6, were copied from the paper without citations: Liu et al. (2020)"
Response 2: We sincerely apologize for this oversight in citation practices. We have:
Added explicit citations to all equations derived from K-BERT: "Equations 3.4-3.6 follow the formulation in Liu et al. [1]"
Included the complete reference in the bibliography:[1] Liu, W., Zhou, P., Zhao, Z., Wang, Z., Ju, Q., Deng, H., & Wang, P. (2020). K-BERT: Enabling Language Representation with Knowledge Graph. Proceedings of the AAAI Conference on Artificial Intelligence, 34(03), 2901-2908.
These additions appear in Section 3.4 of the revised manuscript

Comments 3: "Wq,Wk and Wv are trainable model parameters' is copied without explanation in this paper on how the training will happen."
Response 3: We have expanded the methodology section to provide a detailed explanation of the training process:
Added a new subsection "3.5 Training Procedure" describing the optimization process
Explained how Wq,Wk and Wv are initialized and updated through backpropagation
Included details about the Adam optimizer, learning rate scheduling, and loss function
Comments 4: "Q i+1, Ki+1, Vi+1, are not defined (There is one K as notation for the Knowledge base). It is dubious that, without an exact mathematical formulation, the underlying data structures can be implemented correctly, especially the softmax function."
Response 4: We have addressed this concern by:
Clearly defining all mathematical symbols in a new "Notation" subsection (3.1)
Differentiating between Knowledge base (now denoted as KG) and Key matrix (K)
Providing detailed mathematical formulation of the softmax function implementation
Adding pseudocode in Appendix A to demonstrate the implementability
Comments 5: "Figure 3., 4., 5., 6. are copied from the paper without permission for copyright: Liu et al. (2020)"
Response 5:We apologize for this copyright oversight and have,All charts have been properly cited.
Comments 6:"The experiment is not designed clearly. The stated case study is based on corpora of divorce cases. The classification is not defined for the assessment of the results."
Response 6:We added a manual evaluation system to ensure the reliability of the experimental conclusions.
Comments 7:"In the case of natural language processing, the human should be in the loop to evaluate the goodness or accuracy of the end result. The mentioned metrics can be applied in various fields of machine learning; however, their effectiveness in natural language processing is questionable."
Response 7:We have significantly enhanced our evaluation methodology:
Added human evaluation with domain experts in section 4
4. Response to Comments on the Quality of English Language
Point 1:"The English could be improved to more clearly express the research."
Response 1: We have thoroughly revised the manuscript to improve clarity and readability:
Engaged professional editing services for language polishing
Restructured complex sentences for better comprehension
Ensured consistent terminology throughout the document
All revisions are highlighted in the track-changes version   

Reviewer 3 Report

Comments and Suggestions for Authors

accept

Author Response

Response 1: We thank the reviewer for this positive assessment of our manuscript's language quality. We have maintained the high standard of English expression throughout the manuscript and have performed additional proofreading to ensure linguistic precision and clarity. 

Round 3

Reviewer 2 Report

Comments and Suggestions for Authors

 

There is a minor copying error in equation 3.5; the "T" is a superscript indicating the transposition of the matrix and should be corrected.

The responses are acceptable and have improved the paper. Copyright issues have been observed.

Author Response

1. Summary
We sincerely thank the reviewer for their careful assessment and positive feedback on our manuscript. We greatly appreciate the reviewer's recognition of our revisions and the constructive comments regarding the technical details. We have thoroughly addressed all points raised, and the specific revisions are detailed below.
2. Questions for General Evaluation	Reviewer’s Evaluation	Response and Revisions
Does the introduction provide sufficient background and include all relevant references?	Can be improved	We have expanded the introduction to include a discussion of the suggested related work (e.g., FRT-DETR) and added formal definitions of key terms.
Is the research design appropriate?	Can be improved	We have enhanced the description of our research design, particularly regarding the knowledge graph transformation and its justification, in Sections 3.2 and 3.3.
Are the methods adequately described?	Can be improved	The methodology has been substantially revised with greater detail, especially concerning mathematical formulations, variable definitions, and the Mask-Transformer mechanism (Sections 3.4–3.6).
Are the results clearly presented?	Can be improved	The results section has been restructured for clarity, and statistical measures (mean ± std) have been added to demonstrate robustness (Section 4).
Are the conclusions supported by the results?	Can be improved	The conclusions have been revised to more conservatively and directly reflect the experimental outcomes (Section 5).
Are all figures and tables clear and well-presented?	Can be improved	All figures and tables have been reviewed and optimized for clarity. Source citations and copyright information have been added to Figures 2, 3, and 4.
3. Point-by-point response to Comments and Suggestions for Authors
Comments 1: "There is a minor copying error in equation 3.5; the 'T' is a superscript indicating the transposition of the matrix and should be corrected."
·        Response 1:We thank the reviewer for this precise technical correction. We have corrected Equation 3.5 accordingly. ·        Location of Revision:​ Section 3.4, Equation 3.5 ·        Original Equation:​ ·        S^{i+1} = softmax( (Q^{i+1} * K^{i+1}.T + M) / sqrt(d_k) ) ·        Revised Equation:​ ·        S^{i+1} = softmax( (Q^{i+1} (K^{i+1})^T + M) / sqrt(d_k) ) ·        Change Made:​ The matrix transposition notation has been corrected from K^{i+1}.Tto the proper mathematical superscript notation (K^{i+1})^T.
Comments 2:"The responses are acceptable and have improved the paper."
Response 2: We are very grateful for the reviewer's positive assessment of our previous revisions. We are pleased that our efforts to enhance the manuscript have been recognized.
Comments 3: "Copyright issues have been observed."
Response 3: We take this comment seriously. In addition to the previous revisions, we have conducted a final comprehensive check to ensure full copyright compliance. Actions Taken: Figure Sources:​ We have re-confirmed that all figures are either original creations or properly adapted and cited. Source citations have been double-checked in all figure captions (e.g., "Adapted from..."). Equation Citations:​ We have verified that all equations inspired by or adapted from prior work (especially from the K-BERT paper) are now accompanied by appropriate citations [26]. Text Recycling:​ We have ensured that all background and methodological descriptions are written in our own words, with proper referencing provided for concepts and ideas from other sources. We confirm that the manuscript now fully adheres to academic copyright standards.
4. Response to Comments on the Quality of English Language
Point 1:The English could be improved to more clearly express the research.
Response 1: The manuscript has undergone comprehensive professional proofreading to improve grammatical accuracy, syntactic clarity, and overall fluency.
5. Additional clarifications
Inspired by the reviewer's meticulous attention to detail, we have performed a final proofreading of the entire manuscript, focusing on: Technical Notation:Ensuring consistency and correctness of all mathematical symbols. Reference Formatting:Verifying the completeness and consistency of the reference list. Language Polishing:Minor grammatical refinements for better clarity.