Deep Transfer Learning Model for Semantic Address Matching

Round 1

Reviewer 1 Report

 

General remarks

This article describes 1) the structure of a Deep transfer learning model for Chinese address matching 2) an experimental evaluation of this model on real addresses. Since this article is for the Earth Sciences and Geography section of the journal (i.e not an AI section) I think the Materials and Methods section is not, in part, appropriate. In particular, subsections 2.2.2 and 2.2.3 give too much detail on the machine learning model. Moreover, a large part of these subsections is an almost a verbatim copy of the original article on XLNet (ref [30]). So the authors should instead provide a brief high-level description of the XLNet model and its hyperparameters and then explain how they have adapted the model to their needs. The part of the article that describes the experimental results is scientifically convincing and well written. It provides a good example on how to use the XLNet model for a complex NLP task related to geographic information. In find the article too long for the informational and new contents it conveys. By reducing section 2 it would reach a more reasonable size.

Remarks on specific points

The introduction should provide more information about Chinese addresses for the non-Chinese reader, compare them to simpler schemes of the form street number, street name, city, postal code, country. L135. The “comparison operator” = ˙ must be defined. What does x = ˙ y exactly mean? x and y refer to the same geographic coordinates (up to which precision?) or x and y designate the same building or the same door of a building? Figure 1. Should be simplified. The word “rail” must be replaced by “stream” L156 The sentence means the opposite of what it should. Replace “which do not have delimiters” by “they do not have delimiters”. L163 Replace “. The sub-words” by “, the sub-words” to make one sentence starting with Although. L187 The term “sequence of address” is misleading. It is in fact a sequence of characters that represents an address L204 where → when L417 (Figure 6) The caption contains “... accuracy...” and the vertical axis label is “Precision” L429-33 word2vec or Word2Vec? L438 Remove “relatively” L460 (Figure-7) The bar chart is incorrect. The vertical axis must range from 0.0 (not 0.6) to 1.0 to be fair.

Author Response

1. This article describes 1) the structure of a Deep transfer learning model for Chinese address matching 2) an experimental evaluation of this model on real addresses. Since this article is for the Earth Sciences and Geography section of the journal (i.e not an AI section) I think the Materials and Methods section is not, in part, appropriate. In particular, subsections 2.2.2 and 2.2.3 give too much detail on the machine learning model. Moreover, a large part of these subsections is an almost a verbatim copy of the original article on XLNet (ref [30]). So the authors should instead provide a brief high-level description of the XLNet model and its hyperparameters and then explain how they have adapted the model to their needs. The part of the article that describes the experimental results is scientifically convincing and well written. It provides a good example on how to use the XLNet model for a complex NLP task related to geographic information. In find the article too long for the informational and new contents it conveys. By reducing section 2 it would reach a more reasonable size.

 

Response: Thanks for your insightful and helpful suggestions! Indeed, in Section 2 we dwelled too much on XLNet. Considering your comments, we have therefore merged the original section 2.2.2 and 2.2.3 into 2.3.2 and removed the description of the details of the XLNet implementation. The reduced section is shown in section 2.3.2. The adjustment of XLNet's key hyperparameters is described in detail in section 3.1.

 

2. The introduction should provide more information about Chinese addresses for the non-Chinese reader, compare them to simpler schemes of the form street number, street name, city, postal code, country. L135. The “comparison operator” = ˙ must be defined. What does x = ˙ y exactly mean? x and y refer to the same geographic coordinates (up to which precision?) or x and y designate the same building or the same door of a building? Figure 1. Should be simplified. The word “rail” must be replaced by “stream” L156 The sentence means the opposite of what it should. Replace “which do not have delimiters” by “they do not have delimiters”. L163 Replace “. The sub-words” by “, the sub-words” to make one sentence starting with Although. L187 The term “sequence of address” is misleading. It is in fact a sequence of characters that represents an address L204 where → when L417 (Figure 6) The caption contains “... accuracy...” and the vertical axis label is “Precision” L429-33 word2vec or Word2Vec? L438 Remove “relatively” L460 (Figure-7) The bar chart is incorrect. The vertical axis must range from 0.0 (not 0.6) to 1.0 to be fair.

 

Response: Thank you so much for your careful check! In the second paragraph of our introduction, we compare the address patterns of different countries and also specifically analyses the characteristics of Chinese addresses (page 2, lines 51-62).

We have added an explanation of the comparison operator as follows (page 5, lines 187-189):” The operation objects on either side of the comparison operator refer to the same real-world object with the same coordinates.”

We have simplified Figure 1 and replaced " rail " with "stream" (page 6, lines 210).

We have replaced “which do not have delimiters” by “they do not have delimiters” (page 6, lines 214-215).

We have replaced “. The sub-words” by “, the sub-words” to make one sentence starting with Although (page 6, lines 222).

We have replaced the term “sequence of address” by “address record” in our paper.

We have replaced “accuracy” by “precision” in line 425.

We have replaced “Word2Vec” by “word2vec” in our paper.

We have removed “relatively” in line 460.

We have modified Figure 7 so that the vertical axis range starts at 0.

Reviewer 2 Report

At some points in the paper, the term 'semantic' is too general.  A number of technologies and methods that analyze semantics are presented, but the types of semantics may differ at each stage. For example, 'spatialization' of address elements based on neighboring text positions; geographic properties; natural language pragmatics, classification and inclusion; or algorithmic semantics. A better clarification of the nature of the semantic meaning that is derived from a form or model would help tied the stages of the research better to each other and to clarify the difference between semantics and syntactics of the data processing.  Readers may infer semantic differences between stages, but by using the word 'semantic' too often, it hides the nature of the meaning at a particular stage. 

Author Response

1. At some points in the paper, the term 'semantic' is too general. A number of technologies and methods that analyze semantics are presented, but the types of semantics may differ at each stage. For example, 'spatialization' of address elements based on neighboring text positions; geographic properties; natural language pragmatics, classification and inclusion; or algorithmic semantics. A better clarification of the nature of the semantic meaning that is derived from a form or model would help tied the stages of the research better to each other and to clarify the difference between semantics and syntactics of the data processing.  Readers may infer semantic differences between stages, but by using the word 'semantic' too often, it hides the nature of the meaning at a particular stage.

 

Response: Thank you for pointing out this problem in manuscript. In order to clearly illustrate the different meanings of "semantic" in this paper, we explain the different stages of "address semantics" in Section 2.2. The added text is as follows: (page 5, lines 193-198) “In addition, due to the many different representations of the same location, we believe that it is not possible to achieve a correct match without processing from a natural language understanding perspective. Therefore, in our study, "address semantic understanding" refers to the textual understanding of the address corpus, while the "address semantic reasoning" used for address matching is based on the spatial relationship reasoning of addresses.”

Reviewer 3 Report

In this paper, the authors worked on a very interesting research topic and introduced an address-matching approach based on a pre-training fine-tuning model to identify semantic similarities between various addresses. They pre-trained the address corpus to enable the Address Semantic Model (abbreviated as ASM) to learn address contexts unsupervised and then build a labeled address-matching dataset using the address-specific geographical features, allowing the matching problem to be converted into a binary classification prediction problem.

But unluckily I have noticed that the paper needs extensive changes as it is not properly written, and it looks like notes so here it is suggested for the authors to do major changes to it and resubmit it.

Author Response

Dear reviewer,

Thank you for your helpful comments about our manuscript “Deep Transfer Learning Model for Semantic Address Matching.” Following the insightful comments and helpful suggestions, we have revised the manuscript, and the revised parts have been highlighted as yellow. Our point-by-point responses to the reviewers’ comments are as follows.

1. In this paper, the authors worked on a very interesting research topic and introduced an address-matching approach based on a pre-training fine-tuning model to identify semantic similarities between various addresses. They pre-trained the address corpus to enable the Address Semantic Model (abbreviated as ASM) to learn address contexts unsupervised and then build a labeled address-matching dataset using the address-specific geographical features, allowing the matching problem to be converted into a binary classification prediction problem.

But unluckily I have noticed that the paper needs extensive changes as it is not properly written, and it looks like notes so here it is suggested for the authors to do major changes to it and resubmit it.

 

Response: We feel sorry for the inconvenience brought to the reviewer. Considering the Reviewer’s suggestion, we have re-written Introduction according to the Reviewer’s suggestion, and have restructured the full paper. In addition, the paper has been polished by an established expert in English. We also have carefully checked and improved the English writing in the revised manuscript. Please check if the amended version meets the standard of English presentation.

Reviewer 4 Report

To find semantic commonalities across diverse addresses, the authors presented an address matching method based on a pre-training fine-tuning model. To enable the ASM to learn address contexts unsupervisedly, they first pre-train the address corpus. Using address-specific geographic features, they create a labeled address matching dataset that enables the matching problem to be transformed into a binary classification prediction problem. These are my opinions:

1- Why, background, what, and gap are absent from the introduction section. complete disregard for hierarchy. It is recommended that the authors rewrite this section. Additionally, the paper's structure is missing.

2-The work lacks important references.

3-Formulas must be appropriately cited if they are derived from other sources.

4-Don't leave any sections empty. Complete section 3 with concise sentences describing the subsections and what happened afterward.

5-Make the Discussion and Conclusion sections distinct. also mention the drawbacks of the method in the conclusion section.

Author Response

Dear reviewer,

Thank you for your helpful comments about our manuscript “Deep Transfer Learning Model for Semantic Address Matching.” Following the insightful comments and helpful suggestions, we have revised the manuscript, and the revised parts have been highlighted as yellow. Our point-by-point responses to the reviewers’ comments are as follows.

To find semantic commonalities across diverse addresses, the authors presented an address matching method based on a pre-training fine-tuning model. To enable the ASM to learn address contexts unsupervisedly, they first pre-train the address corpus. Using address-specific geographic features, they create a labeled address matching dataset that enables the matching problem to be transformed into a binary classification prediction problem. These are my opinions:

 

1- Why, background, what, and gap are absent from the introduction section. complete disregard for hierarchy. It is recommended that the authors rewrite this section. Additionally, the paper's structure is missing.

 

Response: We feel sorry for the inconvenience brought to the reviewer. Considering the Reviewer’s suggestion, we have re-written Introduction according to the Reviewer’s suggestion, and have restructured the full paper. We also clarify the organization of the sections of the paper in the last paragraph of the Introduction section.

 

2-The work lacks important references.

 

Response: We gratefully appreciate for your valuable comment. We have added a total of eight important references from address matching studies in recent years, namely references 53 to 60. The additions are on page 2, lines 95 to 99; page 3, lines 102 to 105; page 3, lines 112 to 113.

 

3-Formulas must be appropriately cited if they are derived from other sources.

 

Response: Thank for your comments. We have removed the description of the details of the XLNet implementation and reduced the number of formulas. For the remaining formulas, we have indicated the source and cite.

 

4-Don't leave any sections empty. Complete section 3 with concise sentences describing the subsections and what happened afterward.

 

Response: Thank for your comments. We restructured the paper to place both the results and the discussion in Section 3. We show the experimental results and discuss them in each subsection of Section 3, elaborating on what happened in each subsection and afterwards.

 

5-Make the Discussion and Conclusion sections distinct. also mention the drawbacks of the method in the conclusion section.

 

Response: Thank you for pointing out this problem in manuscript. We have made the discussion and conclusion sections distinct by placing the discussion in Section 3 and the conclusion in Section 4. Based on your suggestions, we have also discussed the shortcomings of our study in the Conclusions section, with the following added text (page 15, lines 481-484):“However, this study has some limitations. 1)we treated each address record as a sentence, which results in the hierarchy of addresses being ignored; 2) our experimental area was limited to a city, which ignored the ambiguity of place names. ”

Reviewer 5 Report

The paper proposes a novel, transfer learning based solution for semantic address matching. 

The paper focuses on chinese address model which is not commonly known and should be explained or referenced in the paper in order to be more comprehensive. for instance it is not clear weather the position of a character in a string affects the meaning or not and what is the goal of permutations

 

The authors do not cite any of their previous work in the similar field so the path for the research presented is not clear. 

Several equations miss the explanation of the symbors. i.e. eq 1 what is E and p ..

 

Evaluation - it is not clear if the evaluation was performed over the same dataset? If so, what is the implementation of the methods used?, if not, reference the results used.

Literature cites rather old solutions while dealing with the modern problem. Only a few references have been published in the last 2 years. More recent literature should be included in overview

Author Response

Dear reviewer,

Thank you for your helpful comments about our manuscript “Deep Transfer Learning Model for Semantic Address Matching.” Following the insightful comments and helpful suggestions, we have revised the manuscript, and the revised parts have been highlighted as yellow. Our point-by-point responses to the reviewers’ comments are as follows.

The paper proposes a novel, transfer learning based solution for semantic address matching.

 

1.The paper focuses on chinese address model which is not commonly known and should be explained or referenced in the paper in order to be more comprehensive. for instance it is not clear weather the position of a character in a string affects the meaning or not and what is the goal of permutations

 

Response: Thank you for pointing out this problem in manuscript. In the second paragraph of our introduction, we compare the address patterns of different countries and also specifically analyses the characteristics of Chinese addresses (page 2, lines 51-62). In addition, we take a simplified Chinese address record for example in section 2.3.2 to explain whether the position of a character in a string affects its meaning, and what the purpose of the permutation is.

 

2.The authors do not cite any of their previous work in the similar field so the path for the research presented is not clear.

 

Response: We gratefully appreciate for your valuable comment. In the fifth paragraph of the Introduction, we focus on the past work of address matching based on language understanding. It also contains a study of address matching based on pre-trained language models. To enable a clearer path of research in this field, we have added eight important references common to address matching research in recent years, i.e., references 53 to 60, and divided them according to methodology.

 

3.Several equations miss the explanation of the symbors. i.e. eq 1 what is E and p ..

 

Response: Thank you for pointing out this problem in manuscript. We have added notes to all the symbols included in the formulas in this paper. In Equation 1, E denotes the maximum Expectation and p denotes the predicted probability (page 7, lines 261-262).

 

4.Evaluation - it is not clear if the evaluation was performed over the same dataset? If so, what is the implementation of the methods used?, if not, reference the results used.

 

Response: Thank for your comments. We have evaluated different methods on the same dataset, and the evaluation metrics include precision, recall and F1 score. We have elaborated on these three evaluation metrics in the last paragraph of section 2.4, and the various methods of comparison are shown in Figure 7.

 

5.Literature cites rather old solutions while dealing with the modern problem. Only a few references have been published in the last 2 years. More recent literature should be included in overview

 

Response: We gratefully appreciate for your valuable comment. We have added a total of eight important references from address matching studies in recent years, namely references 53 to 60. The additions are on page 2, lines 95 to 99; page 3, lines 102 to 105; page 3, lines 112 to 113.

 

Round 2

Reviewer 3 Report

The auhtors have incorporated the changes/suggestions properly. However, the figures (i.e., Fig: 2,3,4,5,6,7) need to be replaced/reconstruct as they are not cleared and blurred.

After these changes, I accept the paper for publication in the Applied Sciences journal.  

Author Response

Thanks for your helpful suggestions! We have replaced Figure 2 with a clear Visio format and also replaced Figures 3, 4, 5, 6 and 7 with the original in HD.

Reviewer 4 Report

It can be accepted.

Author Response

We thank you for reading our paper carefully and giving it a positive review and endorsing it for publication.

Reviewer 5 Report

The paper is significantly improved and ready for publishing

Author Response

We thank you for reading our paper carefully and giving it a positive review and endorsing it for publication.