- Rating
- 1
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper introduces ConceptTensor, a method to represent a DAG as a tensor, weighting edges by it's parent's Dirichelt posterior over levels and a Gaussian span weight. It defines some querying operations, such as parent/child retrieval. The method is evaluated on 2 small (synthetic?) graphs.
Critical strengths
None.
Critical weaknesses
In my honest opinion, this paper is not at all close to being in a state where it is ready to be reviewed. Below, I have included some examples of this.
- The motivation in the introduction is very vague, and does not state the problem that's being studied. It has some mention about embedding entities and relations, and says that “White-box, generality, and scalability are difficult to achieve simultaneously.” But what this has to do with each other is not mentioned.
- The paper does not position its contributions adequately in relation to prior work. As an indication, the related work section only cites 4 papers.
- The presentation of the paper is very problematic. Basic information, such as having captions for tables, is missing. Many sections (e.g. 4.5, 4.6, 4.7) only contain definitions and no actual explanation what is going on. Moreover, many of these defenitions seems trivial or unnecessary (such as Proposition 4 which says that the if c is a child of p, p is a parent of c).
- The evaluation does not really explain its setup. I do not even know what the dataset is. Given that only 2 small graphs (with N = 25 and 50) are included, the experiments are also extremely limited in scope.
Questions for authors
I do not see anything that could materially affect my assessment, see weaknesses.