An Improved BiLSTM Approach for User Stance Detection Based on External Commonsense Knowledge and Environment Information

Round 1

Reviewer 1 Report

An Improved BiLSTM Approach for User Stance Detection Based on External Commonsense Knowledge and Envinment Information

In this paper the authors have introduced a deep learning framework based on BiLSTM to detect user's stance on topics discussed in social media. The proposed model is experimented on a supervised setup and the core novelty of the proposed work is two-fold: (i) they used external common sense knowledge graph data obtain the details about topics and (ii) they introduce a new representation learning method for social media users to detect their stance. Extensive experiments on two real world data shows that the propsoed model outperforms the baseline methods. The ablation study supports the need for each model present in the proposed model.

This paper is neatly written in most of the parts and it is easy to understand. However I found the following issues in the current draft. I suggest authors to fix the following issues for publication:

1. It is a bit unclear on what the input data is and how it pre-processed for the model. I undestand the input data is a base feature for a user. I do not understand how this is combined with a topic of interest

2. The social network module of Figure 1 is misleading. How this social graph is constructed, given that each node is a tweet? Is it constructed based on a timeline? If yes, then it will not be graph. So, it is not clear to me.

3. I feel the results given in the draft to be redundant. Same results are given in table format and picture format. I suggest to use only one format in the final version. There is literally no use in giving both formats.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

·       The problem definition of this work is not clear. In Sect. 1, several paragraphs are too wordy, including the first and second paragraphs. Moreover, the citations used can be improved by using recent research. Several articles are discussed in the research survey. However, no comparison is shown with these techniques. The research survey and references are meaningless. The authors must improve the introduction. 

·       In Sect. 4, the authors not explain in detail how topic information is used as input for the proposed method. Moreover, the reason for using LDA as topic extraction is thin.

·       The proposed method is maze. The authors not explain the process of getting neighbors tweets. The authors should explain the details of the proposed techique by visulize the proposed method, not only the network architecture.

·       In Sect. 5, the figures look meaningless and redundant. The results study have been shown in the table so that the use of images is not required.

·       The effectiveness of this work is not clear. Through simulations/experiments, the authors must justify the effectiveness of the proposed method by comparing with the other latest methods. As a comparison, the authors can use the state-of-the-art methods of stance detection (not only based on SemEval 2016 results).

·       The authors explained that there is a difference between user stance and user viewpoint. Make a comparison table with examples of tweets for a proper comparison.

Author Response

Please write down attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

It seems previous review comments have been accommodated with green highlights in the revised paper.