News Recommendation Based on User Topic and Entity Preferences in Historical Behavior
Round 1
Reviewer 1 Report
The article presents an interesting method of incorporating information on topic and entities extracted from a title and user preferences for topics and entities extracted on the basis of user's past behavior. The effectiveness of the method is supported by good results.
The problem of the article is the level of presentation. The description of the algorithm is not detailed enough. Sections 3 and 4 require more details. If Section 3 - Preliminaries is given, it means the authors assume that a reader may not have enough background in the field, but section 3 would not help. There are a number of symbols not explained, e.g. those for head, relation and tail. Obviously their meaning should be explaided. Moreover, symbol E looks different in different lines, e.g. 163 and 164.
The preliminary could also contain some concepts or methods mentioned later, such as word2vec, doc2vec, TransE algorithm, attention mechanism etc. If all these common techniques are briefly decsribed in Section 3, then the complex description in Section 4 could be more readable
The decsription of authors' own method in Section 4 should be divided into more steps (or subsections) and a summary of it with references to proper subsections could be given at the begining. To enhance the readability, each step should be illustrated by an example (if possible).
In lines 256 and 320 the word "contact" appears. Shouldn't it be "concat"?
It is not clear if the authors used any pre-trained embeddings or trained on their own?
What if a new title is not close to any topic centroid. Is there a threshold set to find outliers or is it always assigned to one of the clusters, even if the topic is completely new (not presented in the training data).
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Leveraging knowledge graphs for news recommendations is an interesting topic. However, I was expected to see how the authors use actual knowledge graphs for recommendations. The overall approach makes sense to me, but the issue is the validation of the proposed system.
First of all, Wikipedia is not a knowledge graph. The semantic web-based of this knowledge source is DBpedia. Extracting the entities and information from a DBpedia is standard and is usually done using SPARQL queries. It is unclear how the authors validate/recommend the news using this knowledge source.
Secondly, I have not seen a recent article (after 2019) in the domain cited by the authors. The Semantic Web community has published many papers in this domain recently.
I recommend authors update their approach to truly use DBpedia or an actual knowledge graph to validate their proposed approach.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors addressed the issues that I mentioned and they provided a knowledge graph in their experimentation. I still think that the literature review should be improved by recent publications (2021-2022).
Author Response
Please see the attachment.
Author Response File: Author Response.pdf