Multitask Learning with Knowledge Base for Joint Intent Detection and Slot Filling
Round 1
Reviewer 1 Report
Lexicons and word search is really not my expertise, however, I have some comments, at least towards the delivery of the content:
1) in the introduction I have not identified some works, i.e.
https://www.sciencedirect.com/science/article/abs/pii/S0950705119303211
https://arxiv.org/pdf/2101.08091.pdf
as well as the concept of knowledge should be elaborated
2) The method, including the maths, is kind of obscure to me. I have failed to follow it and hence I cannot understand it. maybe diagrams like in the above work could also be used.
3) The results among the various methods are not that different, at least in the first reading. Maybe some extra runs are needed, or some other kind of Error (absolute values?) can be defined to better showcase the difference (I am saying this based on the discussion and the relation between "America" and "United States").
4) Conclusions should not be a summary of the paper. this part is a bit weak.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
In this paper, the authors present a joint model of intent detection and slot filling based on multitask learning (MTL) with knowledge base, which makes full use of the external knowledge and the high-quality relationship information between intents and slots. Especially, as the latest method to pre-training language representation, Bidirectional Encoder Representations from Transformers (BERT) is used to convert text inputted into a vector in this model.
The paper has merit and could be accepted for publication, if the authors tackle the following issues:
- In the abstract and/or introduction, the motivation, contribution and novelty of this research should be clearly stated.
- The authors are advised to include the following references to support several assertions:
- F. Giannakas, C. Troussas, I. Voyiatzis, C. Sgouropoulou, A deep learning classification framework for early prediction of team-based academic performance, Applied Soft Computing, Volume 106, 2021, 107355, https://doi.org/10.1016/j.asoc.2021.107355.
- C. Liu, T. He, Y. Xiong, H. Wang and J. Chen, "A Novel Knowledge Base Question Answering Model Based on Knowledge Representation and Recurrent Convolutional Neural Network," 2020 IEEE International Conference on Service Science, 2020, pp. 149-154.
- Troussas C, Krouska A, Sgouropoulou C, Voyiatzis I. Ensemble Learning Using Fuzzy Weights to Improve Learning Style Identification for Adapted Instructional Routines. Entropy. 2020; 22(7):735. https://doi.org/10.3390/e22070735
- In Section 3, a vizialisation of the approach should be provided.
- In Section 4, the discussion should be enriched. This section should make it clear how the finding lead in new knowledge.
- Limitations of this study are missing at the moment.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The presented research tries and, I believe, achieves in contributing within the field of a task-oriented dialog system, i.e., their design and performance improvement. While reading the paper, I have not found any issues regarding the quality of the presentation, study design, or discussions. In my view, the authors invested considerable effort into explaining their model. The block scheme in Figure 1 is very informative and helps with understanding the rest of the paper. The model innovativeness is very high, especially for connecting the external knowledge and improving the model accuracy.
Only a few minor issues to fix:
- Line 50: explain SVM abbreviation,
- Line 257: „…to extend…“
- Line 428: „We…“
- In the concluding chapter, the first few sentences are maybe too repetitive. I would focus on highlighting the obtained results and outlining the future research paths, avoiding explaining the model architecture once again.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have dedicated time to address the comments. It has been reshaped and is much more comprehensive.
Just a clarification from my side: the comment on the error implied adding one more indicator and not substituting it. In any case, it is more than OK as it is.
Reviewer 2 Report
The paper is improved based on the reviewers' comments and it can now be accepted for publication.
There is a typo in the sentence:"Trousers constructed an intention detection model ... to determine learning style [20]."
Based on the Reference [20], Trousers must be replaced to Troussas.
