Dear Prof. Alessandra Mileo: * We sincerely appreciate your valuable comments and suggestions! Here are our concrete responses to your raised issues and how we address your concern in the revised version. * Issue 1: Regarding the presentation. * Response 1: Thanks for mentioning this! We made thorough proofreading to improve our presentation. We also add reference papers to consolidate our concept. * Issue 2: The relevance and importance of the proposal in this paper should be more strongly emphasized. * Response 2: Thanks for suggesting us to further highlight the motivation of our proposed architecture to establish the position of this paper. * In fact, the motivation of this paper comes from multiple aspects. For example, the neuro-symbolic framework resembles the human cognitive process, where the neural part corresponds to System 1 and the symbolic part behaves like System 2. Indeed, you have also provided a new angle to view neuro-symbolic framework—neural networks are based on very complex statistical and mathematical models but are still not able to capture the complexity of the human brain. * To address this concern, we revised Paragraph 4-5 to make clearer statement regarding our motivations. * Issue 3: Some claims are introduced without enough context * Response 3: Thanks for your suggestion! We provide clear reference for dual-process theory when these terms (e.g., System 2) are first mentioned. * Issue 4: The context of complex reasoning and question answering should come better defined and earlier * Response 4: Thanks for your suggestion! We make clear explanations for complex reasoning, knowledge intensive tasks, and their relationship with question answering in (Line 41, Page 2) - (Line 47, Page 2). We also supplement the advantages and disadvantages of the two parts in the beginning of Section 2, where we aim to propose our solution architecture. * Issue 5: 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? / Issues regarding Section 3 and Section 4. * Response 5: Thanks for raising the question! In fact, the proposed architecture is a general blueprint and its implementation is flexible. We do not specify how to implement each individual component and encourage further exploration. The role of Section 4 is to show how existing SOTA techniques can be potentially used to construct the architecture. However, implementation should not be limited to these methods only. Moreover, these methods are not tailored for our proposed architecture, which means they are less likely to be the optimal solution. The position of this paper is to call for more researchers to develop methods under such architecture. * Issue 6: Human-in-the-loop role seems minor… maybe it should not be when it comes to reasoning, and should be part of the architecture. * Response 6: Indeed, we would like to stress that human-in-the-loop means that humans interact with all the components of the architecture, as we show in the revised Figure 2. * Issue 7: I do not think LLM understands complex semantics. This is the reason why they hallucinate, there is no understanding of the semantics and constraints or knowledge guidance. * Response 7: We provide more evidence on LLMs understanding complex semantics in (Page 3, Line 39) - (Page 3, Line 41). * Issue 8: 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. * Response 8: Thanks for raising the question! Indeed, each component can be implemented with neural methods or symbolic methods. We would like to highlight this point in the caption directly. * Issue 9: Section 2.2 should also provide examples of manipulators that are neural or hybrid. * Response 9: Thanks for your suggestion! We address this issue with a running example through Section 2. * Issue 10: 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 * Response 10: We highlight that humans interact with all the four components in the architecture and we thus make revision to our framework illustration (Figure 2). However, due to page limitation, we are not able to further introduce the human-in-the-loop process and would like to address this issue in a technical paper. ------------------------------------------------- Dear Prof. Filip Ilievski: * Thank you for your constructive suggestions! We would like to address your issues as following: * Issue 1: A critical missing piece of the introduction is: why this paper and why now? * Response 1: See Response 2 to Reviewer Prof. Alessandra Mileo. * Issue 2: The paper makes a lot of claims that really need support. For instance, a repeated claim is that grounding the data to some symbolic formalism will reduce hallucinations - but this is a nontrivial claim. * Response 2: Thanks for raising the concern! We add more reference paper for clarity * Issue 3: he benefits of the different architectural aspects are presented without considering their nuances * Response 3: After revision, we introduce their nuances at the beginning of Section 2 * Issue 4: I am wondering who the "human" is in this manuscript. * Response 4: We specify the user profile of ``human’’ at the end of Section 2.5. * Issue 5: I understand that the paper is meant to be a vision paper on a high level, but I do think that it would have been very useful if the paper provided some discussion on some of the key challenges that someone considering this architecture would face / a certain roadmap towards a practical realization of this architecture would have been very insightful * Response 5: See Response 5 to reviewer Prof. Alessandra Mileo. * Issues regarding each sections * Response 6: Thanks for the suggestion! We make revisions to the paper accordingly to address these issues. ------------------------------------------------- For Reviewer #3 * Thanks for your review. * In this paper, we present an architecture for knowledge-intensive complex reasoning tasks, such as answering questions like “Which one came out first, the movie Harry Potter or the first related game?”, which requires neural-symbolic reasoning to combine various knowledge sources and complex planning. * We have added the motivation of why the four modules. Actually, the knowledge manipulator can influence the reasoning conductor, because the reasoning conductor needs the support of reasoning planner, and the reasoning planner needs the support of knowledge manipulator. * This is a position paper with limited length, so we mainly concentrate on clarifying the rationale behind the overall architecture, and providing feasible paths for detailed implementation. In the future, we will expand it with more “how” details.