Reviewer has chosen not to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes
Adequacy of the bibliography: Yes

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

This paper presents a succinct and important overview of the state of research on OWL-based knowledge graphs within neursymbolic systems. As the authors state (and I agree), the OWL language and broader ontology community are under-represented in neurosymbolic research, which is a pity given that OWL is the largest accepted standard for knowledge representation across a multitude of application domains including biomedicine. It is fantastic and much needed to steer more neurosymbolic research in this direction.

I would like to see the paper say more about the limitations of OWL embeddings, for example how the neighbourhood-based or fixed-length random walk-based embeddings lose potential long-range logical relationships (which can be mitigated to some extent by using an explicitly materialised form of the ontology as the input).

I would also like to see the paper say more about the approaches for knowledge injection (mentioned very briefly in the penultimate paragraph of section 2.3 on page 4)

Finally, this point might be taken as a shameless plug for a citation, but I wonder if the authors might consider translations from OWL to FOL as a promising strategy to mention in section 3.4., as this was one of the reasons motivating our work on this topic. See, for example, Flügel et al., 2022 "When One Logic Is Not Enough: Integrating First-Order Annotations in OWL Ontologies" for a description of our implemented translation mechanism and some case studies (https://www.semantic-web-journal.net/content/when-one-logic-not-enough-i...). This point is of course only a suggestion for the authors to consider.

Overall, a very exciting contribution to the journal!

Reviewer has chosen to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes, but limited
Adequacy of the bibliography: Yes

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

This position paper makes a compelling case for the value of leveraging OWL-based knowledge graphs and their reasoning capabilities in neurosymbolic AI systems. The paper does a good job describing the benefits of OWL-based KGs, including the availability of open standards and reusable resources like reasoners and triplestores, the ability to create custom domain ontologies and knowledge graphs, and the reasoning and logical consistency enforcement capabilities. Concrete examples illustrate how OWL reasoning could be used to guide neural learning, such as by inferring new knowledge to strengthen weak supervision in training data or providing feedback when predictions violate ontological constraints.

Several promising research directions leveraging OWL KGs are discussed in depth, such as enabling neural networks to emulate KG reasoning, using KG reasoning to compensate for incomplete training data, and integrating KG reasoning into existing neurosymbolic frameworks like Logic Tensor Networks. The authors make a compelling argument that OWL KGs are well suited for emulating approaches like logical constraints on training, while avoiding issues like combinatorial explosion.

A particularly noteworthy contribution is the introduction of the new NeSy4VRD resource, which provides a dataset and tailored OWL ontology to enable more research on knowledge graph-based neurosymbolic AI for visual relationship detection. The authors correctly identify that a lack of suitable datasets with aligned ontologies is likely inhibiting research in this area currently.

In terms of weaknesses, there are a few areas where additional detail would strengthen the paper. The overview of neurosymbolic AI is very brief - expanding this to better contextualize the role ontology reasoning could play would be helpful. More concrete examples demonstrating OWL reasoning in action would also reinforce the arguments made, the main focus seems to be Visual Relationship Detection.

Overall though, this is an important contribution highlighting an underexplored area with significant promise in advancing neurosymbolic AI. The authors make a persuasive case that the reasoning capabilities of OWL knowledge graphs are extremely relevant and impactful capabilities for hybrid neural-symbolic systems that deserve more research attention. Developing tailored resources like NeSy4VRD will help accelerate progress in this exciting field.