Systematic Literature Review on Neuro-Symbolic AI in Knowledge Graph Construction for Manufacturing

This paper focuses on the topic of Knowledge Graph (KG) construction in the manufacturing domain. In particular, the focus is on neuro-symbolic AI techniques. The research method used is that of a systematic literature review. The research is quite well executed and leads to interesting results. For publication, the paper should be improved in the following main areas:

(1) The study methods description should be revised by clarifying the rationale behind choosing the research questions, as well as some of the design choices (see detailed comments).

(2) Additional background knowledge should be added for the readers to understand all aspects of the paper, especially related to the AI system patters.

(3) In the sections describing the results, several figures miss a description of what they depict and/or an interpretation of the information they show. There are also inconsistencies between the numbers in the text and their counterparts on the figures. Some parts remain abstract and could be extended with examples.

(4) The discussion/conclusion sections do not revisit the RQs and as such they feel rather disconnected from the main part of the paper. This part should be re-written to capture the main conclusions along the RQs that the study investigated.

Detailed comments:

* What was the rationale behind deriving the fours RQs in Sect 3.1.1? Why are these the most important? Why does it make sense to investigate these and not other RQs?

* For the survey methodology: please clarify whether both primary and secondary studies (i.e. surveys) were included.

* The letter "I" is used to identify both inclusion criteria (Table 3) and the extracted data items (Table 5). This can lead to confusion and therefore the data items could be referred to with another letter than "I" (e.g., "D").

* "Balancing the large number of parameters of BERT, ALBERT is employed in three publications, as shown in Figure 8a." => please include more clarifications, as it is not clear how this statement can be deduced from Figure 8a.

* "We show the number of combinations of common building blocks of NER methods in Figure 8b." => please expand this sentence to describe what exactly is shown in Figure 8b. Additionally, please also provide an interpretation of the shown data.

* Sect. 4.3.4, paragraph starting on L3 makes a number of strong claims about joint knowledge extraction, which are not supported by relevant related work. Please include such support from literature as support for these claims.

* Section 4.4. assumes quite some background knowledge from the reader in terms of what are the I, Y, T patterns (p21/L5) as well as what d-M-s and d-K-s stand for. It would be advisable to provide some succinct material to explain the meaning of these notations.

* Fig.12 - please explain what the figure depicts exactly and provide an interpretation thereof.

* Sect. 4.5 => the last paragraph remains very abstract, mentioning "Innovative models, methods, frameworks" but does not give any examples thereof. Including some examples would make the text much more interesting.

* "4.6. Fairly Provisioning" => does fairly refer for FAIR? Please clarify and if yes, use FAIR instead.

* It is commendable that the results of the SLR are published as a KG. Did the authors consider ORKG as a publication space for their results? This would give their work a much higher visibility.

Further/minor comments:

* abstract: The main research question "Which role play neuro-symbolic AI approaches in knowledge graph construction for Manufacturing?" should be rephrased to "Which role do neuro-symbolic AI approaches play in knowledge graph construction for manufacturing?"

* p2/l3: "To work with KGs, they first must be constructed." - consider removing, as this states the obvious.

* p3: Fig1 should be explained in more detail in the text of the paper.

* Section 2: there is no need to reproduce the title of the discussed papers, it is sufficient if the references are added

* p4/l40 - "show a clear overview" => "show an overview"

* p5/l12 - "which is reasonable for the reviewed time period" - it si not clear how this statement was derived. Please consider adding an explanation of this statement

* p5/table 1 - please clarify what do X, (X) and - mean. Are they similar in semantics with the full/half/empty circle used in the MD column? If not, how do they differ? If yes, can the notation be unified?

* p5/l42 design patterns" representing "statistical ..." => design patterns" representing the combined use of "statistical ..."

* p6/l16 DT acronym needs definition

* p7/l28 3.1.3 => Section 3.1.3

* p12/l20 conduction => execution?

* p13/l47 "further stages 5" => "further 5 stages"

* p15/l17-18; "79% of the cases applied on unstructured text" => actually, in Fig 6, this is 60%

* p17/l39 "binary relations are relevant in our final search corpus" => please clarify why this is the case.

Summary
This paper presents a systematic literature review on neuro-symbolic AI approaches in knowledge graph construction for smart manufacturing. It follows a structured PRISMA methodology and analyzes 49 papers from 2020 to 2024, covering entity recognition, relation extraction, ontology reuse, and tools. While the methodology is rigorous and the analysis is detailed, the definition of neuro-symbolic AI is overly broad, and the distinction between manufacturing-specific and general KG construction is unclear.

Strengths:
- The PRISMA-based approach ensures transparency and reproducibility.
- The paper provides a solid overview of ML-based KG construction in manufacturing.
- The discussion on KG construction methods, particularly in manufacturing, is valuable.
- The paper contains informative visualizations that support the analysis.

Weaknesses:
1. Inaccurate Definition of Neuro-Symbolic AI
The paper uses a too broad definition of neuro-symbolic AI, classifying any ML-based entity extraction followed by ontology mapping as neuro-symbolic. Standard definitions emphasize integration between neural and symbolic methods, such as reasoning over embeddings or logical constraints guiding learning. Most reviewed papers do not meet this criterion. I know that these approaches are covered by the Boxology of Neuro-Symbolic approaches, however, I don’t think that the discussion of neuro-symbolic methods adds anything to this work and is more confusing than actually insightful.
2. Weak Manufacturing-Specific Focus
The paper does not sufficiently explain how manufacturing KG construction differs from general KG construction. Many methods discussed are standard NLP techniques and could apply to any domain. Especially in RQ2 it is often unclear how the mentioned methods are different from what is currently going on in the general KG construction and information literature.
3. Unstructured Discussion in RQ2
The response to RQ2 is lengthy but lacks a clear structure. Subsections cover overlapping topics without an underlying red threat. It isn't easy to follow how the different elements connect.
4. Unclear Explanation for RQ3
The KG construction patterns (Atomic, Fusion, I-Type, Y-Type, T-Type) are not well explained. The terminology is technical and assumes prior knowledge, making it difficult for readers unfamiliar with neuro-symbolic architectures.
5. Readability and Figure Quality Issues
Some figures (e.g., Figure 8) are low resolution and difficult to read. The PRISMA methodology is overly detailed, while the discussion on neuro-symbolic reasoning is underdeveloped.
6. No LLM approaches
As mentioned in the survey itself, I find it a bit surprising that there are no LLM approaches in this survey at all. This has become one of the major technologies in information extraction and it is not really covered in this survey. I see this as a big shortcoming, however, I see that adding this without changing the complete literature study, is not feasible here. At least, it is discussed in detail in the discussion section.

Details:
- Why do you use Industrie 4.0 as a search term?
- Is there a reason you did not use Google Scholar or Semantic Scholar as a search engine?
- Fig 3. (b) I am a bit surprised that there are mostly journal publications in your survey since this usually is a very conference-heavy field. Do you have an explanation for this?
- Page 18, line 5: I don’t understand the citation “weak correlation and high entity density”. I think it would be nice to rephrase these citations and explain them if actually needed here.

Overall:
This paper provides a well-structured survey on ML-based KG construction in manufacturing but misrepresents neuro-symbolic AI as I see it and mostly focuses on standard information extraction and knowledge graph construction techniques.
These are major points that I think that should be improved.
1. Clarify the definition of neuro-symbolic AI or reframe the paper.
2. Strengthen the discussion on manufacturing-specific aspects.
3. Reorganize RQ2 for better structure and clarity.
4. Improve explanations of KG construction patterns in RQ3.
5. Enhance figure resolution and rebalance content.

This paper conducts a literature review on neuro-symbolic AI approaches in KG construction in Smart Manufacturing.

They are preforming a reviews of SRL on KC construction and conclude that there is not sufficiend and updated SRL on neuro-symbolic methods to build KG for maniguacturing domain. They argue that manifacturing domain has an importnat specificity and deserves a specific survey of nesy approches for this domain.

In the first part of the paper, the authors outline a general methodology for Systematic Literature Review in the general case. This methodology is taken from a paper that studies the problem of producing good SLRs. The author applies the method to select the papers for review, and after the selection phase, 49 papers on NeuroSymbolic integration for manifacturing have been chosen for the review.

The type of literature review produced is a 'meta-review,' meaning that the paper reports figures on topic containment (e.g., does the paper use a neural approach or not?), publication venue, data type used to build a knowledge graph (KG), and approaches to building a knowledge graph, etc. However, the analysis of the different approaches never touches on technical aspects, which would provide good insight for applications.

In such a review paper, I was expecting a description of a general architecture for neuroSymbolic tools used to solve tasks in manufacturing, including components with their possible features, as well as a description, comparison, and classification of the different approaches.
See for instance Table 1 of the paper [1]

A further constraint is that the contribution to Semantic Web technologies and AI topics feels limited. Given its specificity to industrial applications, the paper might be better suited for a journal specializing in manufacturing or industrial AI.

With that said, I suggest to reject the paper.

[1] Giuseppe Marra, Sebastijan Dumančić, Robin Manhaeve, Luc De Raedt, From statistical relational to neurosymbolic artificial intelligence: A survey, Artificial Intelligence, Volume 328, 2024.