The RDF2vec Family of Knowledge Graph Embedding Methods

Tracking #: 3462-4676

Jan Portisch
Heiko Paulheim

Responsible editor: 
Claudia d'Amato

Submission type: 
Full Paper
Knowledge graph embeddings represent a group of machine learning techniques which project entities and relations of a knowledge graph to continuous vector spaces. RDF2vec is a scalable embedding approach rooted in the combination of random walks with a language model. It has been successfully used in various applications. Recently, multiple variants to the RDF2vec approach have been proposed, introducing variations both on the walk generation and on the language modeling side. The combination of those different approaches has lead to an increasing family of RDF2vec variants. In this paper, we evaluate a total of twelve RDF2vec variants on a comprehensive set of benchmark models, and compare them to seven existing knowledge graph embedding methods from the family of link prediction approaches. Besides the established GEval benchmark introducing various downstream machine learning tasks on the DBpedia knowledge graph, we also use the new DLCC (Description Logic Class Constructors) benchmark consisting of two gold standards, one based on DBpedia, and one based on synthetically generated graphs. The latter allows for analyzing which ontological patterns in a knowledge graph can actually be learned by different embedding. With this evaluation, we observe that certain tailored RDF2vec variants can lead to improved performance on different downstream tasks, given the nature of the underlying problem, and that they, in particular, have a different behavior in modeling similarity and relatedness. The findings can be used to provide guidance in selecting a particular RDF2vec method for a given task.
Full PDF Version: 

Minor Revision

Solicited Reviews:
Click to Expand/Collapse
Review #1
By Jedrzej Potoniec submitted on 24/May/2023
Review Comment:

All my remarks regarding the previous version of the paper were addressed by the authors to my satisfaction. I believe the paper is now suitable for publication, as I am no longer confused by it. While reading it, I noticed a few details that perhaps should be corrected in the camera ready:
* Paragraph right below Figure 2: "(see Fig. 2:" -> "(see Fig. 2):"
* e-RDFvec is typeset differently in different parts of the paper
* In the sentence right above Eq. (14), in the second DL expression, the dot is placed before the relation instead of after it.

Review #2
Anonymous submitted on 20/Jun/2023
Review Comment:

This manuscript was submitted as 'full paper' and should be reviewed along the usual dimensions for research contributions which include (1) originality, (2) significance of the results, and (3) quality of writing. Please also assess the data file provided by the authors under “Long-term stable URL for resources”. In particular, assess (A) whether the data file is well organized and in particular contains a README file which makes it easy for you to assess the data, (B) whether the provided resources appear to be complete for replication of experiments, and if not, why, (C) whether the chosen repository, if it is not GitHub, Figshare or Zenodo, is appropriate for long-term repository discoverability, and (4) whether the provided data artifacts are complete. Please refer to the reviewer instructions and the FAQ for further information.

The paper described the evaluation work of RDF2Vec methods on datasets, with regard to standard tasks, such as classification, regression, clustering, ....

Datasets are publicly accessible.

This improved version still contains errors, some writings (including those marked in blue) are clumsy.
In the middle picture of Figure 1, 'C p D' should be 'C r D'.

The contribution is weak, recent processes in neural embedding for KG embedding is ignored -- the vector embedding used in the paper is not preferred embedding for classification, query, etc. The Word-embedding method is quite out-of dated, and has little impact to the top-level research community of representation learning.

All three criteria in originality, significance of the results, and quality of writing are not good. This work should not appear in class A level publications, such as SWJ.

Review #3
Anonymous submitted on 20/Jun/2023
Minor Revision
Review Comment:

Summary of the paper:
In this paper, authors evaluate the quality of twelve RDF2vec variants on GEval benchmark and DLCC benchmark, and compare them to seven knowledge graph embedding methods. They analyze the results from the perspective that which ontological patterns in a knowledge graph can actually be learned by different embeddings.

I would like to order my review following the response(R) from authors by adding comments(C) on each point:

Q: Considering that some RDF2vec variants mentioned in the related work are not included in the evaluation, and DLCC is already proposed in previous work, the originality and the novelty of this paper are limited.
R: We agree that this paper builds on quite a few already published parts. However, it provides a comprehensive and compact account of a topic which can otherwise only be obtained by reading a bunch of different papers. Moreover, it still contains quite a few novel aspects. In particular, the discussion of the representational power of the different RDF2vec variants has not been published before.
As far as the variants from the related work section, we have added an explanation why certain extensions were not considered in the experiments.
C: I appreciate that authors add reasons for choosing extensions and agree the main contribution of this paper comes from the comprehensive experimental study and discussion on the representational power of the different variants.

Q: Quality of writing: The paper is easy to understand. Some paragraphs are exactly the same as paragraphs in [17], which is not expected to happen.
R: As this paper builds on top of existing works, we would assume that overlaps, both in terms of content as well as in terms of wording, are acceptable. The submission guidelines of SWJ state that “results previously published at conferences or workshops may be submitted as extended versions.” In our opinion, this does not require a complete rephrasing of the original publications’ contents.
C: I agree.

Q: Some of the statement is improper and should be rewritten.
R: We have tried to rephrase them such that it becomes clear which parts are genuine contributions in this submission and which are contributions in the papers that this submission is an extended version of.
C: I appreciate the updates.

Q: In the related work section, it is said [18] distinguish five different techniques for KGE, while such category is proposed for graph embedding which is a broader topic. Is it proper to use this category for KGE?
R: The authors of [18] list different kinds of graphs, including knowledge graphs, and make no restrictions in the sense that certain categories are only valid for particular kinds of graphs. Exceptions for methods which are not commonly used for KGE are mentioned in the paper. Therefore, we deem the categorization valid. We have added a corresponding remark in the paper.
C: I agree.

Q: For evaluation of KGE, MRR is also a common metric which is missing in related work of knowledge graph embedding evaluation.
R: MRR is mostly used in evaluations for link prediction, however, link prediction is not among the tasks used in this paper. For an evaluation of RDF2vec based approaches on link prediction tasks, we refer to our paper “Knowledge Graph Embedding for Data Mining vs. Knowledge Graph Embedding for Link Prediction - Two Sides of the same Coin?”, also published in SWJ.
C: I agree that MRR is not chosen to be used in this paper. But considering that this is a common metric, it could be added to the subsection "knowledge graph embedding" in Related work and put together with rank/HITT@10.

Q: SG and CBOW are first mentioned in the first line of section 3.2 without explanations.
R:We have added a general introduction into RDF2vec in that section, including SG and CBOW.
C: I appreciate the updates.

Q: It is not explained what is the difference between tick and (tick) in table 2.
R:This is obsolete, as the corresponding parts of the table were removed by suggestion of reviewer 1. The intended semantics were: tick means: “expected to be learnable”, (tick) means: “expected to be learnable to a certain extent”.

Q: Figures 2 and 3 are the same as in [17]
R: That is true. We have added the corresponding source to the figure caption.
C: Fine.

Q: As a work target at evaluating RDF2vec variants, it needs more in deep analysis of RDF2vec variants based on experiment results.
R: We have evaluated the RDF2vec variants on various datasets (GEval, and the two datasets of DLCC).
C: I agree.

Q: In summary, my main concern about this paper is the limited contribution and novelty.
R: We see that point, but we consider the paper rather an extended version of previously published works, rather than a new paper from scratch.
C: I agree.

Following are some minor points:
1. Page 3: "machine learning applications section 4" -> "machine learning applications in section 4"
2. Page 5: "by word embedding algorithm RDF2vec" -> "by word embedding algorithm word2vec"
3. Suggest putting Table 6 and 7 after Table 4.