10. Knowledge-Based, Graph, and Neuro-Symbolic AI

Introduction and Context

Artificial intelligence has developed along two influential traditions. The first is symbolic AI, which represents knowledge explicitly through concepts, relations, rules, logic, and procedures for reasoning. The second is machine learning, which infers patterns from data and can operate effectively even when explicit rules are unavailable. Each tradition addresses weaknesses in the other. Symbolic systems are interpretable and capable of exact reasoning, but they often require expensive knowledge engineering and struggle with ambiguity. Statistical learning is adaptable and effective with high-dimensional data, but it may be opaque, brittle outside its training distribution, and weak at systematic reasoning.

Knowledge-based, graph, and neuro-symbolic AI examine how these traditions can be used together. Knowledge-based AI asks how an intelligent system can store and manipulate structured knowledge. Graph-based AI represents entities and relationships as networks and applies algorithms ranging from symbolic graph search to graph neural networks. Neuro-symbolic AI combines learned representations with explicit concepts, constraints, logic, or reasoning procedures.

These approaches are important in doctoral research because they raise foundational questions about representation, generalisation, causality, explainability, and the relationship between perception and reasoning. They also support practical systems in medicine, finance, cybersecurity, scientific discovery, enterprise search, recommendation, robotics, and public administration. The central challenge is not simply to choose between symbolic and neural methods, but to determine which forms of knowledge should be learned, which should be encoded, how uncertainty should be represented, and how reasoning should interact with data-driven inference.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Distinguish knowledge bases, ontologies, knowledge graphs, rules, and formal logics.

2. Explain how automated reasoning, planning, and constraint satisfaction operate.

3. Compare symbolic graph algorithms with graph neural networks.

4. Design a conceptual architecture for a knowledge-intensive AI system.

5. Evaluate neuro-symbolic approaches in terms of accuracy, generalisation, interpretability, and computational cost.

6. Analyse risks involving incompleteness, inconsistency, bias, provenance, privacy, and governance.

7. Select appropriate methods for professional and research applications.

From Data to Explicit Knowledge

A knowledge-based system contains representations of a domain and mechanisms for deriving conclusions from them. A simple representation may consist of facts such as `WorksFor(Ada, LabX)` and rules such as:

`WorksFor(x, LabX) -> HasAccess(x, LabXResources)`.

The system can infer that Ada has access to the laboratory resources. This separation between knowledge and inference is a defining characteristic of classical symbolic AI. Knowledge can be inspected, revised, and reused across tasks.

A knowledge base is a collection of assertions, rules, definitions, and sometimes procedures. A knowledge graph is a graph-structured knowledge base in which nodes represent entities or concepts and edges represent relations. An ontology is a formal account of the concepts, categories, properties, and constraints within a domain. An ontology therefore provides more than a list of terms: it specifies intended meaning and permissible relationships.

For example, a clinical ontology may distinguish `Disease`, `Symptom`, `Medication`, and `Patient`; define that a medication treats a disease; and state that a particular class of diseases is a subclass of a broader class. Ontologies support semantic interoperability because different systems can refer to the same concepts using shared definitions. They can also encode constraints, such as domain and range restrictions, disjoint classes, and cardinality requirements.

Knowledge representation is not neutral. Decisions about categories determine what can be expressed and what remains invisible. A rigid ontology may fail to represent culturally diverse classifications, emerging scientific concepts, or contested entities. Consequently, ontology engineering is both a technical and epistemological activity involving scope definition, competency questions, expert consultation, versioning, and evaluation.

Logic and Automated Reasoning

Logic supplies formal languages and inference rules. Propositional logic represents statements that are either true or false. First-order logic extends this with variables, predicates, functions, and quantifiers. Description logics provide a computationally controlled fragment of first-order logic and underpin semantic-web languages such as OWL.

A reasoning task may involve:

• Deduction, deriving conclusions guaranteed by premises.
• Abduction, finding plausible explanations for observations.
• Induction, generalising from examples, usually associated with machine learning.
• Consistency checking, determining whether assertions can all be true together.
• Entailment, testing whether a conclusion follows from a knowledge base.

Suppose a knowledge base states that every registered device has an identifier, that Device17 is registered, and that Device17 is deployed in a hospital. A reasoner can infer that Device17 has an identifier. If another assertion says that Device17 is both authorised and unauthorised under mutually exclusive policy rules, a consistency checker can flag a contradiction.

The computational trade-off is fundamental. More expressive logics can describe richer domains but may make inference computationally expensive or undecidable. Practical systems therefore often use tractable fragments, rule engines, database-style query processing, or approximate reasoning. Open-world semantics, common in ontology languages, assumes that the absence of a statement does not imply its negation. This differs from the closed-world assumption used in many databases, where missing information is often treated as false. Confusing these assumptions produces serious errors in knowledge integration and compliance applications.

Reasoning may be monotonic or non-monotonic. In monotonic reasoning, adding facts cannot invalidate prior conclusions. In non-monotonic reasoning, new evidence may retract conclusions. Default reasoning is essential when systems must operate with incomplete information—for example, assuming a vehicle is available unless a maintenance record indicates otherwise. Belief revision and truth-maintenance systems provide mechanisms for managing such changes.

Planning and Constraint Satisfaction

Planning concerns the selection of actions that transform an initial state into a desired goal state. A classical planning problem specifies states, actions, preconditions, effects, and goals. A delivery robot, for example, may have an action `Move(RoomA, RoomB)` whose precondition is that the robot is in RoomA and whose effect is that it is in RoomB. A planner searches for an action sequence that satisfies the goal while respecting resource and safety constraints.

Planning can be deterministic or probabilistic, fully observable or partially observable, single-agent or multi-agent. Modern systems may combine symbolic planning with learned world models. A neural model can estimate likely outcomes or perceptual states, while a symbolic planner enforces action preconditions and safety rules.

A constraint satisfaction problem consists of variables, domains of possible values, and constraints restricting combinations of values. Scheduling examinations, assigning hospital staff, configuring a telecommunications network, and routing vehicles are typical examples. Constraint propagation removes values that cannot participate in any valid solution; search then explores the remaining alternatives. Satisfiability modulo theories extends Boolean satisfiability with theories such as arithmetic, arrays, or equality.

Constraints are valuable in AI because they express requirements that should not be violated merely because a statistical model finds a high-probability alternative. They can encode fairness rules, legal restrictions, physical limitations, and organisational policies. Nevertheless, constraints may conflict, be incomplete, or impose unacceptable computational costs. A practical solver therefore needs mechanisms for prioritisation, relaxation, explanation, and negotiation among constraints.

Knowledge Graphs and Semantic Technologies

A knowledge graph commonly represents information as triples: subject, predicate, and object. For example, `(DrugA, treats, DiseaseB)` and `(DiseaseB, subclassOf, ChronicDisease)`. Graph databases store and query such structures efficiently, while semantic technologies provide formal vocabularies and inference mechanisms.

Important technologies include RDF for graph-based statements, SPARQL for querying RDF graphs, OWL for ontology-based representation, and rule languages for deriving additional facts. Property graphs, used in many graph database systems, permit labelled nodes and edges with properties and often prioritise operational querying over formal semantics. The distinction matters: a property graph may store a relationship such as “employedSince=2020,” whereas an ontology additionally specifies the meaning, constraints, and inferential consequences of employment.

Knowledge graphs may be constructed from databases, documents, sensor streams, APIs, and human annotation. Entity resolution links different references to the same entity; relation extraction identifies connections in text; schema matching aligns vocabularies; and provenance records the source, time, confidence, and transformation history of each assertion. A graph without provenance may be difficult to trust because users cannot determine why a fact exists or whether it remains current.

Graph quality should be assessed along several dimensions: correctness, completeness, consistency, timeliness, coverage, and accessibility. Link prediction can suggest missing relationships, but predicted links should not automatically be treated as facts. In high-stakes applications, they require validation, confidence estimates, and traceable evidence.

Graph Neural Networks

Graph neural networks (GNNs) learn representations for nodes, edges, or entire graphs by aggregating information from neighbouring structures. In a message-passing layer, a node receives messages from adjacent nodes, combines them with its current representation, and applies a learned transformation. After several layers, the node representation contains information from a larger neighbourhood.

GNNs support node classification, link prediction, graph classification, recommendation, molecular property prediction, fraud detection, and traffic forecasting. Their strength lies in exploiting relational inductive bias: entities connected in meaningful ways are often statistically dependent. However, repeated aggregation can cause over-smoothing, in which node representations become indistinguishable. Deep graph models may also suffer from over-squashing, where information from exponentially large neighbourhoods is compressed into a limited representation.

GNN performance depends on graph construction. An inaccurate or biased graph can produce systematically misleading predictions. A model trained on a citation network, for example, may learn institutional or demographic patterns that correlate with outcomes without representing legitimate scholarly quality. Evaluation must therefore consider temporal leakage, distribution shift, graph sampling, and whether the target relation is available at prediction time.

Symbolic graph reasoning and GNNs are complementary. Symbolic algorithms provide exact traversal, path constraints, and interpretable proofs. GNNs provide soft generalisation over noisy and incomplete structures. Hybrid systems may use a GNN to rank candidate paths and a rule engine to verify them, or use logical constraints as regularisers during training.

Neuro-Symbolic Architectures

Neuro-symbolic AI is a family of approaches rather than a single algorithm. Its common purpose is to combine the perceptual and statistical flexibility of neural networks with the structured reasoning of symbolic systems.

In a symbolic-to-neural design, rules or logical structures guide representation learning. Logic can be translated into differentiable loss functions, constraints, or neural modules. In a neural-to-symbolic design, a neural model extracts symbolic objects, relations, programs, or facts from raw input. In a joint architecture, learned representations and symbolic reasoning interact iteratively.

A document-understanding system illustrates the pattern. A language model identifies entities and candidate relations in text. An ontology maps them to canonical concepts. A reasoner checks type compatibility and derives additional relations. If contradictions arise, the system requests evidence, lowers confidence, or sends the case for human review. The symbolic output can also provide structured context to the language model for a subsequent answer.

Neuro-symbolic methods address several weaknesses of purely neural systems. They can improve compositionality, data efficiency, interpretability, and adherence to domain rules. Yet integration is difficult. Neural outputs are probabilistic, while classical logic often assumes discrete truth values. Symbol grounding is imperfect; learned entities may not map cleanly to ontology classes. Differentiable approximations can weaken the guarantees of formal reasoning, while exact reasoning can make training slow or non-differentiable.

A useful design distinction is between hard and soft knowledge. Hard knowledge must be satisfied, such as a physical safety constraint or a legal prohibition. Soft knowledge expresses preferences, defaults, or uncertain relationships. Treating soft knowledge as hard can make a system unusable; treating hard requirements as soft can create unacceptable risk. Probabilistic logic, weighted rules, Markov logic networks, probabilistic programming, and constrained optimisation provide alternative ways to manage this distinction.

An Integrated Design Framework

A doctoral researcher designing a knowledge-intensive AI system should begin with the task rather than with a preferred technology. The following sequence is useful.

First, define the decision, users, consequences of error, and required explanations. Second, identify the knowledge sources and distinguish authoritative facts from hypotheses, predictions, and user-provided claims. Third, select a representation: a relational database may be sufficient for stable tabular data; a graph is appropriate when relationships and paths are central; an ontology is necessary when shared meaning and formal constraints matter.

Fourth, determine the reasoning regime. Does the system require deduction, search, planning, probabilistic inference, optimisation, or a combination? Fifth, decide where learning is required. Neural models are valuable for perception, language interpretation, similarity, forecasting, and uncertain pattern discovery. Symbolic procedures are valuable for validation, compliance, causal structure, planning, and explanation.

Sixth, establish provenance, uncertainty, access control, and version management. Seventh, evaluate not only predictive accuracy but also logical validity, calibration, robustness, explanation quality, latency, computational cost, and human usefulness. Finally, monitor the system after deployment. Knowledge changes, ontologies evolve, data distributions shift, and previously harmless errors may become consequential.

Applications Across Organisations and Industry

In healthcare, knowledge graphs connect patients, symptoms, diagnoses, treatments, genes, and clinical guidelines. Neural language models can extract information from clinical notes, while ontology-based reasoning checks terminology and identifies possible interactions. Human oversight remains essential because incomplete records and ambiguous symptoms make automatic conclusions unsafe.

In cybersecurity, graphs model users, devices, events, vulnerabilities, and network connections. GNNs can detect anomalous communication patterns, whereas symbolic rules can express known attack sequences and access-control policies. Combining both helps identify novel behaviour without discarding established security knowledge.

In finance, entity graphs link accounts, companies, transactions, ownership structures, and jurisdictions. Link prediction can identify suspicious relationships, but compliance decisions require provenance and an auditable rationale. Rules may encode reporting thresholds, while machine learning ranks cases for investigation.

In scientific discovery, knowledge graphs integrate publications, experimental results, compounds, organisms, and instruments. Graph learning can propose candidate associations, and symbolic constraints can eliminate chemically or biologically impossible hypotheses. The most promising systems treat predictions as research leads rather than established findings.

In enterprise operations, semantic layers allow employees to query distributed data using domain concepts rather than database-specific terminology. Planning and constraint solving support scheduling, supply-chain design, and resource allocation. Such systems can reduce duplication across departments, but only if governance establishes ownership, definitions, and procedures for correcting knowledge.

Illustrative Case Study: Clinical Treatment-Pathway Assistant

Background

Consider a fictional hospital network developing an assistant for multidisciplinary teams managing patients with complex chronic conditions. The assistant is not authorised to prescribe treatment. Its purpose is to organise evidence, identify possible guideline conflicts, and explain why a recommendation deserves review.

Situation and Problem

Relevant information is distributed across electronic health records, laboratory systems, medication lists, clinical guidelines, and specialist notes. Terminology differs among departments. A conventional keyword search misses indirect relationships, while a purely neural summariser may produce fluent but unsupported statements. The hospital also needs an auditable record of evidence used in each case.

Analysis

The design team identifies four distinct tasks. Language models are appropriate for extracting entities and events from unstructured notes. An ontology is needed to align diagnoses, medications, tests, and clinical concepts. A knowledge graph is suitable for representing relationships among the patient, observations, treatments, contraindications, and evidence sources. A rule and constraint layer is required to check dosage restrictions, incompatible conditions, and guideline prerequisites.

Uncertainty is unavoidable. A statement in a recent specialist note may be more reliable for a current condition than an older coded diagnosis. Therefore, each graph assertion receives provenance, timestamp, source type, and confidence. The system uses an open-world interpretation for missing clinical information: absence of a documented allergy is not treated as proof that no allergy exists.

Approach and Solution

A neural information-extraction model proposes structured facts from notes. Entity linking maps terms to ontology concepts, with uncertain mappings routed for review. The graph stores patient-specific facts separately from general medical knowledge and applies access controls to both. A rule engine checks formal contraindications and derives candidate pathways. A GNN ranks related evidence and identifies similar cases, but its output cannot override hard safety constraints.

When presenting a result, the assistant displays the proposed pathway, supporting sources, timestamps, assumptions, rule firings, and unresolved contradictions. A clinician can accept, reject, or correct an extracted fact. Corrections are versioned and used to improve later extraction, but they do not silently rewrite historical decisions.

Outcome

In this illustrative scenario, the integrated design improves retrieval of relevant evidence and makes recommendations easier to audit than an unstructured summarisation tool. It does not eliminate clinical uncertainty. Its main value is to reduce search burden, expose missing information, and make reasoning inspectable. The hospital therefore defines success through a balanced evaluation: retrieval recall, factual accuracy, contradiction detection, calibration, clinician workload, explanation usefulness, and safety incidents.

Learning Points

The case demonstrates that integration is architectural rather than cosmetic. A graph alone does not guarantee truth; a neural model alone does not guarantee reasoning; and an ontology alone does not guarantee operational usefulness. Reliability emerges from carefully defined responsibilities, provenance, uncertainty management, human review, and evaluation against the real decision context.

Limitations, Risks, and Common Errors

The first risk is knowledge incompleteness. A graph can appear comprehensive while omitting important entities or relations. The second is inconsistency caused by conflicting sources, version mismatches, or incorrect extraction. The third is semantic drift: a term or policy may change meaning over time.

A further problem is ontology bias. Categories may encode institutional assumptions and exclude minority perspectives. Graph-based models can amplify structural bias because highly connected entities receive disproportionate influence. Neural components may exploit spurious shortcuts, and explanations generated after prediction may not reflect the model’s actual decision process.

Common implementation mistakes include treating predicted links as verified facts, mixing open-world and closed-world assumptions, ignoring provenance, applying transitive rules without checking their domain validity, and measuring only accuracy. Another error is forcing all domain knowledge into a rigid ontology when a probabilistic or narrative representation is more appropriate.

Scalability is also significant. Reasoning over large graphs can be expensive, especially when expressive ontologies, temporal data, or uncertainty are involved. Approximate retrieval, graph partitioning, caching, incremental reasoning, and selective materialisation can reduce cost, but they introduce possible completeness and consistency trade-offs.

Ethics, Governance, and Professional Responsibility

Knowledge-intensive AI systems can affect rights, access to services, financial opportunity, and professional judgement. Governance must therefore address lawful data use, privacy, consent, security, accountability, and contestability. Sensitive graph structures can reveal social relationships even when individual attributes are removed. Access control must protect both graph data and inferred relationships.

Organisations should assign responsibility for ontology ownership, fact correction, model monitoring, and incident response. Every consequential output should be traceable to sources, rules, model versions, and human interventions. Where uncertainty is material, the system should communicate it rather than presenting a single confident conclusion.

A technically correct inference may still be ethically inappropriate if the underlying category is discriminatory or if the system uses data beyond the user’s legitimate authority. Professional practice therefore requires impact assessment, stakeholder participation, bias testing, documentation, and mechanisms for appeal. Human oversight should be meaningful: reviewers need sufficient evidence, time, authority, and training to challenge system outputs.

Current Developments and Research Directions

Research is moving toward foundation models connected to external knowledge graphs, retrieval systems, tool-using agents, and verifiable program execution. Graph-enhanced language models aim to improve factual grounding and multi-hop reasoning. Temporal and event-based knowledge graphs represent how facts change rather than treating the world as static. Federated and privacy-preserving graph learning seek to learn across organisations without centralising sensitive data.

Other active directions include causal representation learning, differentiable theorem proving, neuro-symbolic program synthesis, probabilistic ontologies, continual ontology evolution, and benchmark designs that test systematic generalisation rather than memorisation. A major research question is evaluation: success requires measuring not only whether an answer is plausible, but whether it is entailed, sufficiently supported, robust under distribution shift, and appropriate for the decision context.

Best Practices for Design and Evaluation

A robust project should:

• begin with competency questions and explicit use cases;
• separate observations, authoritative facts, predictions, and inferred claims;
• record provenance, timestamps, confidence, and version history;
• choose the weakest formalism that adequately expresses the task;
• combine hard constraints with calibrated probabilistic models where appropriate;
• test graph quality independently from model quality;
• evaluate counterfactual, temporal, adversarial, and out-of-distribution behaviour;
• provide explanations tied to evidence and actual inference steps;
• include domain experts and affected stakeholders in ontology development;
• design correction, appeal, rollback, and monitoring procedures before deployment.

Chapter Summary

Knowledge-based AI represents concepts, facts, rules, and constraints explicitly so that systems can query, explain, and reason over structured information. Ontologies provide shared conceptual meaning, while knowledge graphs connect entities and relations across heterogeneous sources. Logic supports entailment, consistency checking, planning, and constraint satisfaction, but expressiveness must be balanced against computational cost and uncertainty.

Graph neural networks learn from relational structure and are effective for classification, prediction, and representation learning, yet depend strongly on graph quality and may be difficult to interpret. Neuro-symbolic AI combines neural perception and generalisation with symbolic structure, verification, and reasoning. Its success depends on careful handling of soft and hard knowledge, provenance, semantic alignment, uncertainty, and human oversight. Across healthcare, finance, cybersecurity, science, and enterprise operations, reliable systems require more than a graph or a model: they require an integrated architecture and governance regime.

Key Terms

Ontology: A formal specification of concepts, properties, relations, and constraints in a domain.

Knowledge base: A structured collection of facts, rules, definitions, and related inference resources.

Knowledge graph: A graph representation of entities and semantically meaningful relationships.

RDF: A framework representing information as subject–predicate–object triples.

SPARQL: A query language for RDF and semantic graphs.

Description logic: A decidable family of logics used to represent and reason about concepts and roles.

Entailment: A relation in which a conclusion follows logically from stated premises.

Open-world assumption: The principle that missing information is not automatically false.

Non-monotonic reasoning: Reasoning in which new information can retract earlier conclusions.

Automated planning: Computational selection of actions that transform an initial state into a goal state.

Constraint satisfaction problem: A problem defined by variables, domains, and restrictions on allowed assignments.

Knowledge graph embedding: A vector representation of graph entities and relations for computational learning.

Graph neural network: A neural architecture that learns from nodes, edges, and neighbourhood structure.

Message passing: The exchange and aggregation of learned information among connected graph elements.

Neuro-symbolic AI: An approach combining neural learning with symbolic representations or reasoning.

Provenance: Information recording the source, time, transformation, or authority of a fact.

Entity resolution: The process of determining when different records refer to the same entity.

Symbol grounding: Linking symbolic representations to objects, observations, or experiences.

Over-smoothing: The loss of distinction among node representations in deep graph networks.

Semantic interoperability: The ability of systems to exchange information with shared and preserved meaning.

Review Questions

Short Answer Questions

1. What distinguishes an ontology from a simple vocabulary?

2. Explain the difference between open-world and closed-world assumptions.

3. What are the principal components of a constraint satisfaction problem?

4. What role does provenance play in a knowledge graph?

5. Define message passing in a graph neural network.

Long Answer Questions

1. Explain how first-order logic, description logics, and rule engines support automated reasoning, including their computational trade-offs.

2. Compare symbolic knowledge graphs and graph neural networks in terms of representation, generalisation, interpretability, and failure modes.

3. Describe an integrated neuro-symbolic architecture for a high-stakes domain and justify the allocation of tasks between neural and symbolic components.

4. Discuss the importance of uncertainty, provenance, and ontology governance in knowledge-intensive AI.

5. Analyse how planning and constraint satisfaction can improve the safety and reliability of AI agents.

Analytical / Application Questions

1. A fraud-detection graph contains transactions inferred by a GNN but lacks source information. Design a validation and governance process before the links are used for investigation.

2. A hospital system treats the absence of an allergy record as proof that no allergy exists. Identify the semantic error and redesign the reasoning process.

3. An enterprise wants to combine a language model with a policy ontology for automated compliance decisions. Propose an evaluation framework covering accuracy, logical validity, explainability, robustness, and human oversight.

22. Multiple Choice Questions with Answers

Question 1

Which statement best describes an ontology?

A. A collection of unlabelled data points used to train a classifier

B. A formal specification of domain concepts, relations, properties, and constraints

C. A database index designed only to speed up numerical queries

D. A neural network that aggregates neighbouring nodes

Correct Answer: B. A formal specification of domain concepts, relations, properties, and constraints

Question 2

Under the open-world assumption, what does the absence of a fact normally imply?

A. The fact is false

B. The fact is true

C. The fact is unknown unless negated explicitly or otherwise established

D. The knowledge base is inconsistent

Correct Answer: C. The fact is unknown unless negated explicitly or otherwise established

Question 3

A planning system primarily seeks to do what?

A. Compress all observations into one vector

B. Select actions that transform an initial state into a goal state

C. Replace every rule with a probability

D. Remove all constraints from a problem

Correct Answer: B. Select actions that transform an initial state into a goal state

Question 4

Which item is essential to a constraint satisfaction problem?

A. Variables, domains, and constraints

B. Only labelled images

C. A graph neural network with many layers

D. A single deterministic action

Correct Answer: A. Variables, domains, and constraints

Question 5

What is the main purpose of provenance in a knowledge graph?

A. To hide the source of sensitive information

B. To record the origin, time, authority, or transformation of an assertion

C. To guarantee that every predicted link is true

D. To eliminate the need for ontology design

Correct Answer: B. To record the origin, time, authority, or transformation of an assertion

Question 6

A GNN repeatedly aggregates information from neighbouring nodes. Which problem may arise when the network becomes very deep?

A. Over-smoothing of node representations

B. Automatic conversion to first-order logic

C. Guaranteed causal identification

D. Elimination of graph bias

Correct Answer: A. Over-smoothing of node representations

Question 7

Which architecture most clearly represents a neuro-symbolic design?

A. A rule engine with no learned component

B. A neural model that extracts relations, followed by ontology-based validation and reasoning

C. A database containing only numerical columns

D. A random graph with no semantic interpretation

Correct Answer: B. A neural model that extracts relations, followed by ontology-based validation and reasoning

Question 8

Why should predicted links in a knowledge graph not automatically be treated as facts?

A. Graph algorithms cannot represent relations

B. Predictions may be uncertain, biased, or unsupported by authoritative evidence

C. Ontologies prohibit all newly created relations

D. Symbolic systems cannot store provenance

Correct Answer: B. Predictions may be uncertain, biased, or unsupported by authoritative evidence

Question 9

A hospital system infers that a patient has no allergy because no allergy is recorded. Which design principle is violated?

A. The distinction between open-world and closed-world reasoning

B. The definition of message passing

C. The need for graph partitioning

D. The use of entity resolution

Correct Answer: A. The distinction between open-world and closed-world reasoning

Question 10

Which evaluation approach is most appropriate for a high-stakes neuro-symbolic system?

A. Measure accuracy only on randomly sampled training-like data

B. Evaluate predictive performance, logical validity, calibration, robustness, provenance, and human oversight

C. Evaluate the fluency of generated explanations without checking evidence

D. Evaluate only inference speed

Correct Answer: B. Evaluate predictive performance, logical validity, calibration, robustness, provenance, and human oversight