Decomposed Two-Stage Prompt Learning for Few-Shot Named Entity Recognition
Round 1
Reviewer 1 Report
This paper proposes a decomposed two-stage prompt learning for the NER task in the few-shot learning setting. The two stages are entity identification and entity classification. Decomposing the NER task into such two sub-tasks is to improve the time efficiency and consistency of predictions, and their experiments show the effectiveness of the proposed approach.
The paper is well-written and the experimental designs are solid. The experimental results are sufficiently presented and described.
One argument of not splitting the task into subtasks to avoid error propagation. It would be good for the authors to address this common concern in a bit more detail in both the Introduction Section and the Experiment Section.
Author Response
Please refer to the attached one-to-one response PDF.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Review: Decomposed Two-Stage Prompt Learning for Few-Shot Named Entity Recognition
This paper present a a novel two-stage prompt Learning framework for NER tackling NER tasks using prompt learning. It employs the use of a location models to predict NER location and then a type to reconize the type of NER privily located.
The document organisation is ok, follows a general organization, the language is good.
However there are some things not clear
It is non clear how the train behaves when no entity is present in the sentence. This most surely will lead to a invalid or null loss. Have you set some kind of objective in this to be predicted?
When staying the few shot setting, can you specify the percentage of samples regarding the all set used ? What are the main characteristics in the crated low resource training data regarding the rich resource?
Although there are omission of several entity types, is possible to recover then? Or this is straight tied to the low resource data used in training?
For the early stopping strategy, we set 500 steps for (epocs is more usual)
On Conll2003 the increase of perfjoamcne are marginal, where in OntoNote are in the range of 15% on k 5, (although this increase is not due to the fact that you merged/skipped several classes that were considered in the other works? I think this vast increase has something to do with this. Also the speed is affected, since there are less entities to determined.
Distant labels are often valuable for location, its a fact, and what about close labels, this will not introduce some genre of redundancy or mislead the NER localisation stage? Its more easy to be wrong about close her that far ones that are surrounding by non new tokens.
“Our analysis revealed that model performance improves with an increase in data, highlighting the efficacy of the fine-tuning strategy in enhancing model performance”, This is a general obvious conclusion.
What about stability, the proposed method is able to address different set os data?
While the proposed research is somehow evident, however my concerns manly include:
1. Skipping some enteties is not straggly related to the boost in performance regarding other comparative works? This should act almost as a direct proxy to the increase of F1 score. The other works have this in consideration? If not is not suitable to be compared, at least, a fair comparison is needed.
2. The approach requires the creation of a low resource data. What about the NER distribution? It encapsulates all the new location and type variance?
Other Issues:
- In some parts of the text, the sentences are not clear or missing some word connector, They are minor but can be fixed easily
- More details with diference long range ner should be introduced
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors propose a decomposed prompt learning NER framework for few-shot settings, decomposing the NER task into two stages: entity locating and entity typing. A well-trained entity locating model is used to predict entity spans for each input. The input is then transformed using prompt templates, and the well-trained entity typing model is used to predict their types in a single step.
The article is very interesting, but I can highlight the following remarks:
1. The authors describe the proposed approach in section 3 very abstractly. I recommend adding more explanation and details. Perhaps the authors should add illustrative examples to help better understand the models and methods in section 3.
2. After reading the article, I still did not understand how the proposed model combines separate words to define one entity (Los Angeles, Hong Kong Disneyland).
3. Authors should describe the dataset and the process of its formation in more detail.
4. Few-shot learning requires a special approach to generalize a pre-trained models over new categories of data. I recommend the authors to describe this in more detail.
5. I recommend that the authors describe in more detail the characteristics of the environment in which they performed the experiments: library versions, program execution environments, processor model, memory size, etc.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Authors have addressed my comments.
Reviewer 3 Report
Can be accepted in current form
