Elastic CRFs for Open-Ontology Slot Filling
, Yichi Zhang 1,2,†, Hong Liu 1,2, Zhijian Ou 1,2,3,*, Yi Huang 2,4 and Junlan Feng 2,4Round 1
Reviewer 1 Report
- Abstract, line 11: "cross-doamin" -> "cross-domain"
- Please put number for equations.
- It is better to clearly explain "IOB tag", and how it is used in this paper. For instance, where "B", "I", and "O" should be placed respectively.
- Please put related work in the beginning instead of the end before conclusion. Instead, if exists, a discussion should be put after experiment, before conclusion, not related work. Put the related work, which should include the novel contributions of this work, before proposed method section.
- Dataset and tasks section should not be put before proposed method. Put it before or merge with experiments section.
- NER, first used in Section 3, paragraph 1, is not defined in the text.
- In Section 3.3, it is said that a pretraining is utilized by ignoring the edge potential term and using only node potential. Please briefly explain the details on for how long this is performed. For instance, this pretraining is done until convergence, and then it is retrained with both edge and node potential terms.
- Section 4.4: "licoln" -> "lincoln"
- The paper proposes a method called elastic CRF for open-ontology slot-filling. At first, it is expected that a new formulation of CRF with the addition of "elastic" characteristic is proposed. However, it turns out that the CRF formulation stays the same, where the difference is the use of "IOB tags" to encode the slot representation, which is connected to the final CRF layer. Hence, it is quite confusing to digest what is the meaning of "elastic" in the proposed CRF formulation. Do the authors mean to propose a whole new formulation of CRF? Or do the authors use the usual CRF formulation, but with an improved network architecture/data flow? Please clearly describe about this point, and clearly explain what does "elastic" mean.
Author Response
Thank you so much for your valuable review.
Q: Improving paper writing in terms of spell errors, organization, and unclear sentences.
A: Thanks. We have revised our paper according to your suggestions.
Q: What does elastic mean?
A: The key idea of eCRFs is to use slot descriptions to create semantically meaningful IOB tags, which are further used for a new calculation of potential functions in the CRF framework. Compared to traditional fixed IOB tags in original CRFs, our eCRFs are able to process new slots unseen during training without retraining the model. Such flexibility is the motivation we call it a ''elastic'' CRF model.
Reviewer 2 Report
This paper proposes the use of a so-called elastic conditional random field to open ontology slot filling in dialog systems.
The paper is technically sound and the proposed model outperforms two approaches: CT (concept tagging), which is very close to what the authors propose; and BT, which is a simplified form of CT.
The negative aspect of the paper is the literature review. Not much is mentioned about latest techniques for open ontology slot filling. I would suggest a more up-to-date bibliographical research, if possible. In fact, most of the references are up to year 2016. Aside from that I think the paper is well written, covers an interesting topic and shows promising results.
One minor issue is that Fig.3 is mentioned before Fig.2 in the manuscript.
Author Response
Thank you very much for your valuable reviews!
We have revised our related work and the figure issue according to your suggestions.
