Reviewer has chosen to be Anonymous

Summary of paper in a few sentences:

This paper describes the interrelationship between the areas of data science and symbolic AI, by describing several areas where such an intersection has been explored.

Reasons to accept:

The paper deals with a very important topic, which is the intersection of these two areas of research that will surely be brought together (as they have been done in the past). The paper is obviously different from a common research paper, since it can be considered as a vision paper, which is what was sought for in this initial issue for this journal. However, as a vision/summary paper, it lacks, in my opinion, more clear discussions on how these two areas can actually intersect, and on the actual challenges and opportunities that this combination can bring in.

Reasons to reject:

The paper can be considered as an essay, and I have nothing against that, since it is the purpose for this inaugural issue.
However, my main concern is that the paper is not too clear about the challenges and opportunities that the combination of these two areas may bring in. I firmly believe that the strong focus that data science has placed on statistics will slowly move, in the coming years, into the need for understanding and explaining better the knowledge that is acquired using the techniques that are inherent to Data Science, and that we are not there yet. I also agree with the fact that knowledge as it is made available in today's Web (e.g., large knowledge graphs) is also a good source for the methods and tools that Data Science is being developing.
However, reading this paper does not bring in clearly a very clear landscape of where we are (just to name an example, I have not seen clear references to ontology learning techniques i section 3), but mostly a sequence of descriptions of existing efforts that are in between both areas, which do not flow very clearly and do not tell a clear story. Which are your main conclusions from the analysis that you have done? Which are the actual challenges and opportunities that you refer to in the abstract and title? You refer in section 4 to the limits of data science and the relationship to some theories in Science, but these are just examples, and it is not clear how symbolic AI can help in those specific cases, where it seems that mathematical formulations are potentially more useful.

Nanopublication comments:

Further comments:

In summary, I think that evaluating/reviewing a paper like this is always difficult, since it cannot be reviewed under the usual considerations of novelty, relevance, soundness, etc. I really agree with the authors that both worlds need to come together, and I am sure that this will happen. I also like the fact that the authors have identified some areas where such intersection is happening or has happened in the past. But I am missing a clear picture, a clear roadmap, for such a vision paper, where the low hanging fruits are identified, where the longer-term opportunities are discussed, and where the main challenges and potential limitations are also identified and discusssed, so that the paper can provide a good analysis of how such intersection of areas may happen, even if we may be totally wrong when looking back at this paper in a couple of years' time.

By Rinke Hoekstra https://orcid.org/0000-0001-7076-9083

Review Details

---

Reviewer has chosen not to be Anonymous

Summary of paper in a few sentences:

The paper discusses the interaction between symbolic and statistics inspired AI, and highlights what the role for symbolic AI is (or could be) for Data Science. The paper starts out with a discussion of what Data Science *is* as a field, then briefly introduces the two approaches within AI, and proposes ways in which knowledge can be seen as the 'data' for statistical approaches, or how statistical approaches can be used to extract knowledge (patterns) from data.

Reasons to accept:

I think the paper's discussion is highly topical. With the increasing precision of statistical approaches, and availability of larger datasets, we do have a shot at overcoming (or at least alleviating) what is traditionally known as the 'knowledge acquisition bottleneck': how to reliable elicit and acquire expert knowledge. Also, the high quality of symbolic knowledge representation can significantly contribute to the reliability and scope of statistical approaches (that currently have to deal with a 60-80% time investment for making data suitable for analysis). The authors are right at pointing out these benefits, and discuss several cases where these benefits have been shown.

Reasons to reject:

The 'position' of the paper, though topical, is also not a very surprising one. As the authors themselves already note, there exist several efforts in this direction already.

The paper is overly repetitive and detailed in explaining and introducing symbolic approaches, while statistical approaches are only mentioned by name. Technologies such as RDF, OWL/DL and Knowledge Graphs are introduced in a way that they seem to be very distinct, whereas they are strongly interconnected. I am sure this is merely a matter of presentation, but it makes the paper a rather dense read, and distracts from the message, the position, that the authors want to get across.

The discussion of the "grand challenge" for AI to solve is interesting (discovering the principle of inertia), but not detailed enough to be convincing. What concrete evidence is there that "an addition creative step" (sic) needs to be taken to meet that challenge?

Nanopublication comments:

Further comments:

Overall I think the paper does raise an interesting topic, but the paper suffers a bit from verbosity.

By Agnieszka Ławrynowicz https://orcid.org/0000-0002-2442-345X

Review Details

---

Reviewer has chosen not to be Anonymous

Summary of paper in a few sentences:

The paper tackles important issues, and I support such a position paper.
The general concepts and structure of the paper are good.
However, the paper is written often uncarefully, with typos, unclear symbols, unclear focus of specific paragraphs.
Especially, such things as typos should not occur in a journal submission.
The paper seems to be written in a hurry.

Below I list the detailed comments:

Page 1: world in which we „are living” or in which „we live”?

Page 2, second paragraph: the topic of this paragraph is unclear. Is this mostly about scientific experiments, or about data life cycle? Please choose one main topic per paragraph.
It also reads sketchy at some parts. It would be better to provide a complete list of steps of a *typical* data life cycle instead of giving only examples (maybe the sentence could be rewritten to avoid „mainly consists” and „among others”).
What is „archival”?

Page 3: something wrong with this sentence:
„In a physical symbol systems [35], physical entities (tokens, symbols) stand for, or denote, entities, are combined with other symbols to form complex symbol structures, and are manipulated by processes”

Page 4:
Fig.1: It is unclear to what refer the processes of „creating”, „processing” etc. To „data” coming from social, technology and science disciplines?
Can A.I. be classified under both „science” and „technology”? Why e.g. „Computing” is not under „Science” and „Meteorology” under „technology”. Why „Energy” is under Social? What is exactly meant by „energy”? There is a typo in the word „philosophy”.
Overall, the figure makes an impression of a bit too arbitrary. It could probably be more tidied up.

The definition of „ontology” is given two times, once in Section 2 (Page 4) and then again (very similar definition) in Section 3. Should I suspect that specific paragraphs were written by different authors but without consolidating and proof-reading the final result?

Last paragraph of Page 4: tenses are mixed (once there is „have been developed” and the other times „were developed”).
The citied references dealing with knowledge graphs seem to be picked a bit ad-hoc. Could there be more structure in referencing them?

Page 5:
Fig. 2: What kind of a relation/process is meant by two big black arrows? Maybe this can be explained in the caption.
„if”->”whether”?

I do not agree that only some OWL axioms may give rise to a graph structure. In principle, all OWL axioms can be serialized as RDF, i.e. as labelled graphs.

Fourth sentence from the end: the symbol of logical „models” or „entails” is corrupted.

Section 2 would benefit from more structured discussion on what are the particular problems of treating knowledge as data. The current one is a bit too sketchy.

Page 7:
What is „almost infinite”? Is it finite or not or just huge?

References:
Reference [1] lacks the name of the author.
There are missing letters in the names of some authors, e.g. [9, 20]. Some authors are referenced using their first name and some not.

Some typos:
-grpah
-Page 7: an addition -> in addition?

Reasons to accept:

Important and nicely, centrally placed topic for as a position paper to bootstrap a data science journal.

Reasons to reject:

Uncarefull writing, with typos, unclear notations and sometimes sketchy structure.

Nanopublication comments:

Further comments:

Overall, I think that the paper, as a position paper, tackles important and timely topics, and I would be happy to see this kind of a paper published.
It needs though careful proof-reading and polishing.

By Honghan Wu https://orcid.org/0000-0002-0213-5668

Review Details

---

Reviewer has chosen not to be Anonymous

Summary of paper in a few sentences:

This position paper presents an overview about the data science as a discipline with a focus on discussing the differences and associations of the two main methodologies in Artificial Intelligence - knowledge representation (symbolic representations and their computations) and statistical approaches (e.g., neural networks, deep learning and etc.). Specifically, the authors discussed 1) how explicitly presented knowledge can help data science / statistical methods? 2) how statistical approaches can help identify patterns / regularities from data, which can further be distilled as domain knowledge? 3) the future (grand) challenges in data science - challenges that might require the invention of disruptive methodologies that are able to abstract the unobservable, but fundamental laws of the world like Galileo’s principle of inertia.

Reasons to accept:

I think this is a very timely discussion on how the two different methodologies (symbolic and statistical AIs) can and should converge to go a step further towards human-like intelligence. More generally, the community definitely needs this kind of high-level thinking about what data science can do, why it has been successful in certain domains and what are fundamentally missing. The authors gave a clear overview on the current developments in both symbolic and statistical AIs, and presented a good review on the efforts so far of integrating the two. The grand challenge raised in the last section is a valuable opinion that will stimulate discussions and lines of thought about what are the most exciting directions of data science in the community.

Reasons to reject:

I found it is a bit unfortunate that the authors totally missed the consideration of human factors (such as human guided intelligence or human-machine interactions [1]) in this discussion. From the data science's point of view, it is very important to understand what humans are good at (putting in the era of big data) and how such human-factors can be utilised in or used to guide AI algorithms, such as [2]. From the societal point of view, with the increasing concerns from the public about Ai taking over our jobs, this topic also seems unavoidable. Just as Stephen Hawking pointed out "AI will be 'either best or worst thing' for humanity".

Following this direction, I think it is also worth touching the topic of cognitive computing [3], which was proposed as a interdiscipline between computing science, neuroscience and nanotechnology and is now more about technologies to mimic the functioning of the human brain.

[1] Sacha, Dominik and Sedlmair, Michael and Zhang, Leishi and Lee, John A. and Peltonen, Jaakko and Weiskopf, Daniel and North, Stephen C. and Keim, Daniel A. (2017) What you see is what you can change: human-centred machine learning by interactive visualization. Neurocomputing . ISSN 0925-2312
[2] Ofli, Ferda, et al. "Combining human computing and machine learning to make sense of big (aerial) data for disaster response." Big data 4.1 (2016): 47-59.
[3] Modha, Dharmendra S., et al. "Cognitive computing." Communications of the ACM 54.8 (2011): 62-71.

Nanopublication comments:

Further comments:

In addition to the related work introduced in the paper, there is another line of work of integrating reasoning with machine learning at the workflow level that is "an algebraic operation in a space of (machine learning) models"[4]. Furthermore, the paper mentioned "knowledge graph" in various sections without a definition. It would be helpful to provide a brief introduction or references, such as [5].

[4] Bottou, Léon. "From machine learning to machine reasoning." Machine learning 94.2 (2014): 133-149.
[5] Pan, Jeff Z., et al., eds. Exploiting Linked Data and Knowledge Graphs in Large Organisations. Springer, 2017.

Some minor suggestions:
- p2: "(meta)data description, interpretation, archival, publishing," -> (meta)data description, interpretation, archiving, publishing,
- p2: bank transaction -> bank transactions
- p2: "and turning it into knowledge" -> and turn it into knowledge
- p4: figure 1 puts AI in science. I think AI is similar to data science which is a kind of interdiscipline. If it has to be put in the figure, it seems better to be in the technology bit.
- p4: knowledge as data section - it might be worth mentioning the line of work that uses ontologies in natural language processing.
- p5: "owl-structured ontologies" - OWL is a language rather than a data structure. I would say owl-specified or owl-described.