A Study on Standardization of Security Evaluation Information for Chemical Processes Based on Deep Learning

Round 1

Reviewer 1 Report

This paper deals with a method of intelligent HAZOP auxiliary analysis system based on knowledge graph by establishing the ELMO-DCNN-BiLSTM-CRF deep learning framework. The proposed method is especially effective for special problems such as a large number of polysemous words in texts of chemical industry and lack of available annotated data. Experimental results show the effectiveness of the proposed method and the pertinence of the tool combination.

The paper can be improved in the following points:

- In the paper, the motivation for the choice of each tool was lacking, even if the choices are relevant, there are several methods that make it possible to do the work of extraction, selection or classification of attributes, the tools chosen must be better motivated.

- Convolutional neural network is used for feature selection, this method is based on a convolution of signals. compared to other techniques that can be implemented separately, like ReliefF which is a filtering technique, why the authors chose the convolution of the signals? See for example: A review of feature selection methods in medical applications. Computers in Biology and Medicine 112 (2019) 103375.

- The introduction can be enriched by review papers on attribute extraction techniques for diagnostics, based on data from physical sensors such as for example. Accurate Detection and Discrimination of Pollutant Gases Using a Temperature Modulated MOX Sensor Combined with Feature Extraction and Support Vector Classification.

Author Response

Dear reviewers,

We would like to express our sincere appreciation to the editors and reviewers for their insightful comments and constructive suggestions, which have helped us to improve the manuscript significantly. For clarity, we have uploaded a copy of the original manuscript with all the changes highlighted by using the track changes mode in MS Word. Appended to this letter is our point-by-point response to the comments raised by the reviewers. The comments are reproduced and our responses are given directly afterward in a different color (red).

We would like also to thank you for allowing us to resubmit a revised copy of the manuscript.

We hope that the revised manuscript is accepted for publication in the Journal of Process.

Sincerely,

Ms, Lanfei Peng

Author Response File: Author Response.docx

Reviewer 2 Report

The paper presents an interesting approach to extract and systemize the knowledge of HAZOP studies. However, the manuscript focuses only on the introduction of this new approach and nothing else. The literature survey contains references to those building blocks which are necessary to the proposed approach. It suggests that the authors are the first on this topic, who would like to automatize the HAZOP studies. There can be found some works in the literature in which the authors proposed other approaches (sometimes the proposed approach is based on using machine learning) to achieve the same. Just e.g.: DOI 10.1007/978-3-642-34419-0_6; 10.1205/psep.04055; B978-0-12-818634-3.50093-X; 10.1016/j.compchemeng.2012.06.007 and so on. It is necessary to show us how your approach can differentiate from the already existing ones, to see what are the pros and cons using your method.

In the results section only the numerical values of the statistical metrics can be seen, nothing about the developed model. Unfortunately, the Tables do not help me, because I cannot understand the headlines in Table 2. and 6.

At the present form of the manuscript, we get a description of the approach but nothing which can be show that it works, hence I recommend to reject this manuscript and asking the authors to make complete this work with a proper literature review and some results which proves that the proposed method is applicable.

Some minor drawbacks which should be modified in the text:

1. The number of references should be maximized in 2 at the end of sentences, or probably you should refer to only one work at a time to show the reader, why is important that reference and how can you use that information in your work.
2. In Section 2. you use a different reference format than in other sections of the manuscript.
3. What is the primary aim of this work? Please explain this carefully in the abstract and the introduction. Based on the title and the abstract I think I will see how this method can be used to support HAZOP studies.
4. In the abstract you mentioned: “Experiments were carried out on the HAZOP report of coal seam indirect liquefaction project, and the results were significantly improved compared with other models.” Please give a detail explanation what is the result what you would like to see and how can this improved.
5. In lines 325-327 you mentioned: “It shows that method proposed in this paper has a good entity in the hazard and operability analysis data.” What does this mean? How can we see this?

Author Response

Dear reviewers,

We would like to express our sincere appreciation to the editors and reviewers for their insightful comments and constructive suggestions, which have helped us to improve the manuscript significantly. For clarity, we have uploaded a copy of the original manuscript with all the changes highlighted by using the track changes mode in MS Word. Appended to this letter is our point-by-point response to the comments raised by the reviewers. The comments are reproduced and our responses are given directly afterward in a different color (red).

We would like also to thank you for allowing us to resubmit a revised copy of the manuscript.

We hope that the revised manuscript is accepted for publication in the Journal of Process.

Sincerely,

Ms, Lanfei Peng

Author Response File: Author Response.docx

Reviewer 3 Report

This paper describes a Named Entity Recognition system for hazard and operability analysis in the field of chemical safety. They use a deep learning system to achieve good results - a F score of 91.72%. The techniques used are plausible and reasonably modern, but do not include the latest Transformers-based techniques, as used in for example BERT.

The description of relevant work in Section 2 is broad-ranging, but would be improved by mention of some of the fields similar to chemical safety, for example the chemical named-entity recognition word evaulated in the recent BioCreative V.5 challenge (https://biocreative.bioinformatics.udel.edu/media/store/files/2017/BioCreative_V.5_Proceedings.pdf)

I would have liked a lot more detail in section 4.1.1. The new dataset, opening up work in the HAZOP domain, is the most original part of the paper (and the part most relevant to the Processes journal) - the fact that the work is conducted on Chinese language texts adds to the originality. They should describe more what the source texts were (the abstract mentions "the HAZOP report of coal seam indirect liquefaction project" but I can't find any mention of this in the body of the paper), whether the domain experts were given any guidelines in their annotation, and whether inter-annotator agreement was studied. They should also give more examples of the types of entities to be annotated - they mention "materials" and "equipment" but it would be useful to know more. It would also be useful to know whether the annotated texts will be made available to other NLP researchers.

The disadvantage of working on a new NLP task is that it makes it harder to assess the technical contributions of the paper against the current state of the art. This is one reason why I think that more emphasis is needed on the task, and less on the technical aspects.

The authors could save some space by omitting some of the parts of the methods that are common knowledge in the field. For example, Figure 4 shows the internal workings of an LSTM - this could be replaced by a citation to a relevant paper. Likewise the P, R and F measures described in section 4.2 are common knowledge to any NLP researcher and do not need to be spelled out it the paper.

In the methods section (section 3) it is not always clear which things are happening at a word level and which things are happening at a character level.

One interesting aspect of the paper which the authors do not discuss in detail is the fact that the texts the system works on are in Chinese. This is not mentioned early on, it was only on page 8 where an example was given that I realised the texts the system was processing were in Chinese. I do not know any Chinese language but I understand that it provides unique challenges and opportunities for NLP systems; for example, Chinese contains much more information per character than English text, and so character-level techniques are likely to work differently. Also, word segmentation is more difficult in Chinese. It would be interesting to see some discussion of the aspects of the problem that relate specifically to the Chinese language. Also - most NLP systems and resources are targeted at the English language, and so it is good to see systems for other languages. The paper is written in English - it would be useful for readers who do not know Chinese to have translations of the pieces of example text that are given.

In general I think the paper has some potential, but I think that more emphasis is needed on the NLP task and dataset in order to highlight the most original aspects of this paper.

Author Response

Dear reviewers,

We would like to express our sincere appreciation to the editors and reviewers for their insightful comments and constructive suggestions, which have helped us to improve the manuscript significantly. For clarity, we have uploaded a copy of the original manuscript with all the changes highlighted by using the track changes mode in MS Word. Appended to this letter is our point-by-point response to the comments raised by the reviewers. The comments are reproduced and our responses are given directly afterward in a different color (red).

We would like also to thank you for allowing us to resubmit a revised copy of the manuscript.

We hope that the revised manuscript is accepted for publication in the Journal of Process.

Sincerely,

Ms, Lanfei Peng

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

I really appreciate the hard work what the authors have performed to revise the manuscript. My personal opinion is that the manuscript is become much more enjoyable after the review. My major problem with this work was that the authors only showed us some statistical evaluations at the end, but nothing about the obtained knowledge base, which is generalized with the proposed method. That was the reason, I rejected this work at the first place. However, after I read the revised version and I see that all of my other concerns are answered in the text, I should revise my first decision. I can accept that it is not possible to show any more interesting information about the resulted knowledgebase that these statistics, so I recommend to accept this work in this form.

Reviewer 3 Report

The changes to the paper are extensive and do a lot to bring out the more interesting and novel aspects of the work. The paper still needs much work to improve the quality of the English, but apart from that it is now acceptable.