Zero-Shot Topic Labeling for Hazard Classification

Round 1

Reviewer 1 Report

Here are my comments on the paper:

- Issues related to the paper technicalities should be included outside related work.

-The methodology section is not clear to me and here are some questions:

-  when you wrote earlier; Where is earlier? line 216

- Where is the cross validation experiment?  line 221

- What is this shallow auto-encoder? line 223

- Consequently, I could not follow the results section as I could not follow your explanations of the different methods you included in table-3

- I think it is better to rewrite the paper to be more structured. For example each method used must be described clearly as a subsection in the methodology section. 

-Separate anything related to the paper from related work and put it into the methodology section if appropriate.

Author Response

We would like to thank the reviewers for their helpful suggestions. Following the comments, we revised the structure and the presentation of our manuscript. In particular:

We restructured the related work section, moving away irrelevant passages.
We integrated a description of all the methods employed as baselines for comparison in the Methodology section.
We fixed unclear expressions and missing information.
We added more extensive background on the principles of mapping lexical items and labels to the semantic space.

Thank you very much and best regards.

Reviewer 2 Report

This paper presents their work in applying a Zero-shot method to text classification. The main contributions lie in proposing a framework of mapping both the label and the text-to-classify into the same semantic vector space and measuring the similarity of the two after dimensionality reduction. Three dimensionality reduction methods are applied and PCA shows its effectiveness. Another contribution of this work is building a hazard detection dataset and using the framework on this real dataset and showing its effectiveness.  

The work is interesting despite its simplicity. 

As mapping the label and the text to the same semantic space is a pivot part of the framework, the description of this part is lacking. 

Author Response

We would like to thank the reviewers for their helpful suggestions. Following the comments, we revised the structure and the presentation of our manuscript. In particular:

We restructured the related work section, moving away irrelevant passages.
We integrated a description of all the methods employed as baselines for comparison in the Methodology section.
We fixed unclear expressions and missing information.
We added more extensive background on the principles of mapping lexical items and labels to the semantic space.

Thank you very much and best regards.

Round 2

Reviewer 1 Report

The authors have revised the paper to meet my previous comments