An Information Retrieval-Based Joint System for Complex Chinese Knowledge Graph Question Answering

Round 1

Reviewer 1 Report

This article suggests a current and attractive topic for the academy. The research is timely and worthwhile. The research problem is clearly defined. The authors provide fresh insight into the field.

Add limitations section to Discussion.

The first paragraph of the conclusion is too general. Add another that encompasses practical implications and what could be possible outcomes of these practical implications.

The structure of the paper has to be improved as the discussion wanders off the main topic in several places.

Articles should be reformatted according to standard structure, which is set out in the instructions for authors of the journal (sections are Introduction, Materials and Methods, Results, and Discussions, Conclusion). I suggest adding to the Results section to describe practical research results (using data visualization in diagrams and charts).

Authors should discuss the results and how they can be interpreted from the perspective of previously published studies. Possibly you will need to update your reference with a published paper:

- Kaushal, P., Bharadwaj N., Pranav M. S., Koushik S., Koundinya, A. "Myers-briggs Personality Prediction and Sentiment Analysis of Twitter using Machine Learning Classifiers and BERT", International Journal of Information Technology and Computer Science (IJITCS) 2021, Vol.13, No.6, pp.48-60. DOI: 10.5815/ijitcs.2021.06.04

- Zhao, Y., Li, H., Yin, S." A Multi-channel Character Relationship Classification Model Based on Attention Mechanism ", International Journal of Mathematical Sciences and Computing (IJMSC) 2022, Vol.8, No.1, pp. 28-36. DOI: 10.5815/ijmsc.2022.01.03

The paper approaches an interesting topic, bringing a valuable contribution, especially to practice.

The authors should check the new format given by the journal in the section "Instructions for authors." Authors need to use the template of the journal.

Author Response

Thank you very much for your time in reviewing our paper. Your comments and suggestions are very important for us to improve the quality of our paper. According to the suggestions, we carefully improved the writing, answered your concerns in this cover letter and revised the manuscript accordingly. 

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper claims in the introduction, that its contribution is a joint system for knowledge graph question answering  based on information retrieval methods that can avoid candidate path explosions by  three subsystems and a new text matching method 

The experiments are conducted on the CCKS2019 CKBQA dataset.

 

In the first step the system recognizes a name entity in the question. The name entity is linked with the entity in the knowledge graph. Paths from the knowledge graph are ranked by three subsystems. A path-fusing component selects the resulting path.

 

Chapter 2 briefly describes the other QA systems for knowledge graphs.  Chapter 3 presents the proposed system. Chapter 4 analyzes experiments.

 

The topic of the paper is actual and the approach seems to bring a scientific contribution. However, open issues prevent publishing it in the present form.

 

Major Issue 1:

 

It would be interesting to show how the results would be applicable to other languages. What changes are necessary to use the approach for the Latin alphabet? Does the approach need a word-boundary identification?

 

Major Issue 2:

 

The description of the proposed algorithm contains unclear parts and should be improved. It should be explained why a combination of CRF-BiLSTM-BERT was selected and how it is better than other options. The inputs and outputs of the training database for named entity recognition should be described.  To be more clear, section 3.3 should explain how the knowledge base looks and how it is connected to the templates in the search subsystems. 

 

Algorithm 1 should be better explained. 

Algorithm 1, Line 5 - “highest score” does not make sense, because the score is calculated only once. Line 5 in not in a loop. Malye lines 4 and 5 should be switched? How is the sliding window related to the question? There is no loop in the algorithm how the window “slides” in the question. 

 

Algorithm 2: How is embedding of the candidate path calculated?

 

Detailed list of unclear parts in chapter 3:

 

L 162: It is not explained how the NER is trained to link the entities in DBpedia. Is there such a training database?

L 164: Why is LSTM as a separate sequence model. Why is BERT not finetuned, but used as a word embedding? What kind of BERT model is used. The text references an monolingual English model which cannot be used with Chinese. 

L 176 - CRF is unexplained acronym. Is it Conditional Random Field? How CRF layer works with BiLSTM?  Is there any tokenization applied?

L191 - What is “related grade”?

L202  - What is “<?>” ?

L258 - “interactive information” - is unclear term, there is no interactivity in information. 

L266 - is “our semantic similarity” the same as “interactive information”?

L280 - “our semantic model” is not defined. How it was trained?

Equation 3: ‘x’ sign usually means vector product. Is it vector product?

Major issue 3

 

The experiments do not contain a proper comparison with a previous baseline approach. It is not clear if the proposed architecture is completely new, or it is just a slight modification of an existing one. Ifit is  a modification, which component contributed to the improvement?

Although the authors reference results from the previous competition CCKS2019, it would be more convincing if they repeat experiments with some other baseline approach and compare it. 

Minor issues:

• Non breakable space before reference are missing

• Figure 8  has  distorted caption

• Title - acronym IR in title - is it Information Retrieval?/ There should be no acronyms in the title.

• Abstract -  There should be no acronyms in the abstract

• L22 - why smart cities?

 

Author Response

Thank you very much for your time in reviewing our paper. Your comments and suggestions are very important for us to improve the quality of our paper. According to the suggestions, we carefully improved the writing, answered your concerns in this cover letter and revised the manuscript accordingly.

Please see the attachment where all our answers to reviewers are listed.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Accept in present form

Reviewer 2 Report

My objections were resolved, the paper can be accepted.