By Alessandra Mileo
Review Details
Reviewer has chosen not to be Anonymous
Overall Impression: Average
Content:
Technical Quality of the paper: Average
Originality of the paper: Yes, but limited
Adequacy of the bibliography: Yes, but see detailed comments
Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Limited
Organization of the paper: Needs improvement
Level of English: Satisfactory
Overall presentation: Average
Detailed Comments:
This paper identifies the key components that should be part of a neural-symbolic architecture for complex reasoning. It also suggests, for each components, what functionalities they should have, how they should interact and what existing solutions from both neural and symbolic side could be used to implement such components. Future directions could be more indicative of the different opportunities in all components identified.
Some of the references for some of the components (e.g. entity extraction from textual corpora in 3.1) are not very recent.
Language could be improved (see comments below).
English language presentation could be improved: several incorrect or improper grammar forms, use of terms (e.g. initiative vs initial), repeated words and other language issues to be fixed throughout.
Detailed comments (per section) as follows:
INTRODUCTION
In general the relevance and importance of the proposal in this paper should be more strongly emphasised.
There is a very simplistic claim that connectionist AI relies on the idea of computerising neural networks that mimic neural networks in living beings. This is quite simplistic in that we know today that neural networks are based on very complex statistical and mathematical models but are still not able to capture the complexity of the human brain. This should also be one of the reasons why cognitive + neural is and should be gaining attention.
The success of modern approaches to connectionist AI is not only given by better models, but most importantly by more data (to train the models) and more computational resources (to do the training).
Some claims are introduced without enough context: what is System 2 deep learning and what are the other systems? References should be provided to contextualise this sentence.
The context of complex reasoning and question answering should come better defined and earlier. NeuroSymbolic combination for reasoning has specific characteristics (e.g. can we say that a neural network is not able to do REASONING but can instead to LEARNING? In QA, for example, LLM uses attention mechanism and statistical properties to generate the answer, but there is no causality or comparison or other reasoning tasks performed. Therefore, in this case the combination is more a fusion of two different capabilities. I think this should be better highlighted as it is a bit confusing atm in the introduction.
The structure of the paper is a bit confusing too: the components are described, then the state-of-the-art is used to discuss the implementations of these components: are you suggesting to reuse existing components in a new way in the architecture? What other approaches combining Learning and Reasoning are there that leverage Neuro-Symbolic combination?
Human-in-the-loop role seems minor… maybe it should not be when it comes to reasoning, and should be part of the architecture.
SECTION 2
- I do not think LLM understand complex semantics. This is the reason why they hallucinate, there is no understanding of the semantics and constraints or knowledge guidance.
- some sentences in Section 2 are not well formed, I suggest grammar checking
- Lack of clarity re. Concepts around system 1 and system 2 is also present here as well las in the introduction.
- in Fig 2. What is neural? What is symbolic? It all reads symbolic from the boxes but in the paragraph each component is described and possible implementations are suggested, it is clear that some functionalities of the components can be either neural or symbolic in nature. This duality in the components should be clearer. Also, the process is described backwards. It would be better to start from the block where everything starts and go from there.
- Section 2.2 should also provide example of manipulators that are neural or hybrid.
- The role of the human in the components is under defined and simplistically added in terms of the high level cycles the human could/should be involved in.
I think that if the human role is critical, it should be added in the framework conceptual diagram, and better defined in terms of what intervention the human can introduce, what components do the cycles involve, and how does each intervention affect the framework (and how this can be measured). If the human role is in addition to the main framework ability, then it should be seen as external and maybe discussed later on in section 4.
SECTION 3 and SECTION 4
In general I would have expected this section to be not considered stats-of-the-art, but rather a list of concrete suggestions (maybe using the most recent techniques) on how each component/functionality could be implemented.
A SOTA section to me should rather compare other approaches where neural and symbolic AI have been used to combine Learning abilities on the neural side and complex reasoning abilities on the symbolic side (even just looking at Question Answering or Language Understanding). Comparing those approaches with a framework like the one proposed should be focused on highlighting what specific advantages would this framework have (e.g. transparency at multiple level, or better flexibility or …).
An alternative way (or additional perspective) of clarifying related work versus SOTA and open challenges, is indicating for each component in the possible implementations (Sec 3), where are the gaps in existing solutions for certain functionalities. For example, in the interplay between the symbolic and the neural approaches to translating questions into sub-tasks in the reasoning planner, or in the way humans are involved in some of the tasks and where we can do better or what challenges are still to be addressed…). At the macroscopic level one example is provided in section 4 for the former, and at the macroscopic level a general consideration of the role of human in the loop is also provided. However, I think this can be done more systematically (e.g. for different functionalities discussed section 3 at the microscopic level, and for different components/ interactions at the macroscopic level. An account of more limitations would serve as the basis for developing future research ideas in many directions.
