Structure-Mapping Theory is a cognitive scientific theory of analogy proposed by Dedre Gentner in 1983. It explains why, when analogies are established between different objects or situations, correspondences between relationships—and the systematic organization of those relationships—can be more important than similarities in the individual attributes or appearances of the objects.
For example, the human circulatory system and an industrial material supply network consist of different substances and serve different purposes. However, both can involve transporting substances from a supply source to multiple destinations, distributing them through branching pathways, and regulating supply according to demand.
The important correspondence is not simply that “blood vessels resemble pipes.”
It is that a supply source delivers substances through transportation pathways to locations where they are needed.
Furthermore, if a particular mechanism is known in the source domain but has not yet been examined in the target domain, the correspondence may suggest a new hypothesis.
Structure-Mapping Theory attempts to explain analogical reasoning not merely as an intuitive association or sudden inspiration, but as a process of identifying correspondences between relational knowledge structures.
However, establishing a structural correspondence does not demonstrate that a new hypothesis derived from the analogy is correct in reality.
The central concern of Structure-Mapping Theory is not simply what resembles what, but which relationships correspond across domains and what new inferences can be generated from those correspondences.
The Origins and Development of Structure-Mapping Theory
People often draw on familiar examples when trying to understand unfamiliar problems. Analogies are also used when scientists compare natural phenomena or engineers attempt to apply technologies developed in other fields.
But what exactly do people compare when they judge two situations to be similar?
Similarity in shape, color, or other attributes is not necessarily the same as similarity in function or relationships.
For example, historical descriptions comparing an atom to the solar system associate the atomic nucleus with the Sun and electrons with planets.
This does not imply that the nucleus and the Sun have the same color or size. Instead, the analogy attempts to establish a correspondence between a central entity and other entities associated with it.
This historical analogy does not accurately represent the modern quantum-mechanical description of atoms. It is used here only to illustrate relational correspondence.
In her 1983 paper, “Structure-Mapping: A Theoretical Framework for Analogy,” Gentner described analogical reasoning as a process of mapping a relational knowledge structure from a base domain onto a target domain.
A distinguishing feature of the theory is that it formulates principles of analogy in terms of how knowledge is relationally structured, rather than relying exclusively on the particular meanings of the objects being compared.
Subsequent research by Gentner, Kenneth D. Forbus, Brian Falkenhainer, and others investigated how these principles could be implemented computationally.
Two important developments were the Structure-Mapping Engine (SME) and MAC/FAC (Many Are Called, Few Are Chosen).
| Year | Research | Main Contribution |
|---|---|---|
| 1983 | Gentner: Structure-Mapping Theory | Explained analogy through correspondence between relational structures |
| 1989 | Falkenhainer, Forbus, and Gentner: SME | Developed a computational approach to structural mapping |
| 1995 | Forbus, Gentner, and Law: MAC/FAC | Proposed a two-stage model for efficiently retrieving candidate analogical sources |
| 1997 | Gentner and Markman: Structural Alignment | Further explained the role of structural correspondence in analogy and similarity judgments |
These studies provided an important foundation not only for understanding analogy as a form of human reasoning but also for considering how computers might support analogical reasoning.
Basic Concepts of Structure-Mapping Theory
Structure-Mapping Theory distinguishes between two domains being compared.
The Base Domain, also called the Source Domain, contains relatively well-understood objects, situations, or knowledge.
The Target Domain is the object, situation, or problem we want to understand, explain, or investigate.
For example, when using blood circulation as a reference for designing an industrial supply network, the circulatory system is the base domain and the industrial supply network is the target domain.
Structural mapping does not simply compare individual objects.
It attempts to establish correspondences between the relationships in which those objects participate.
| Base Domain | Target Domain | Corresponding Role |
|---|---|---|
| Heart and circulatory driving mechanisms | Pump and supply equipment | Drive transportation |
| Blood vessels | Pipes and transportation pathways | Transport substances |
| Blood | Supplied material | Substance being transported |
| Organs and tissues | Production equipment and work locations | Receive supplies |
| Regulation of blood flow | Flow control | Adjust supply |
However, constructing such a correspondence table is not sufficient to understand the distinctive contribution of Structure-Mapping Theory.
What matters is how the elements are connected through relationships.
For example, rather than simply concluding that “a heart resembles a pump” and “blood vessels resemble pipes,” we can examine a relational structure in which a driving mechanism creates flow and that flow transports substances to destinations.
If the mechanisms regulating supply in response to demand can also be aligned, a more complex structural analogy may become possible.
The Difference Between Attributes and Relations
A fundamental distinction in Structure-Mapping Theory is the difference between attributes and relations.
An attribute describes a characteristic of an individual object.
Examples include a machine being large, a material being red, or a container having a cylindrical shape.
A relation describes how two or more objects, events, or states are connected.
Examples include Machine A supplying Substance B, pressure differences affecting flow rates, or Component C supporting Component D.
| Type | Example | What It Represents |
|---|---|---|
| Attribute | Machine A is large | A characteristic of Machine A |
| Attribute | Container B is cylindrical | The shape of Container B |
| Relation | Machine A supplies a substance to Container B | An interaction between objects |
| Relation | A pressure difference affects the flow rate | A relationship between physical quantities |
| Relation | Controller C regulates the flow rate | A control relationship |
In the original formulation of Structure-Mapping Theory, analogy emphasizes relational correspondences more than the matching of individual object attributes.
For example, blood vessels and industrial pipes do not need to have the same color.
They do not even need to have identical shapes.
The important correspondence is that both can serve as pathways for transporting substances.
However, this does not mean attributes are always irrelevant.
Differences in material properties, geometry, or other attributes may determine whether a proposed application is technically feasible.
The emphasis on relations means that an analogy should not depend exclusively on similarity of appearance.
Representing Relations — Predicate-Based Knowledge Representation
Computational structure-mapping requires an explicit representation of objects and their relationships.
Classical computational models, including SME, use structured representations involving predicates.
For example:
SUPPLY(Pump, Substance)
This represents a relationship in which a pump supplies a substance.
Other examples might include:
FLOWS_THROUGH(Substance, Pipe)
DELIVERS_TO(Pipe, Production_Equipment)
Here, SUPPLY, FLOWS_THROUGH, and DELIVERS_TO represent different kinds of relationships.
The elements within parentheses are arguments participating in those relationships.
Unlike ordinary natural-language sentences, these representations explicitly identify the type of relationship and the objects involved.
If similar representations are constructed for both the base and target domains, structural correspondences become easier to examine.
| Base Domain: Blood Circulation | Target Domain: Industrial Supply |
|---|---|
SUPPLY(Circulatory_Driver, Blood) | SUPPLY(Pump, Supplied_Material) |
FLOWS_THROUGH(Blood, Blood_Vessel) | FLOWS_THROUGH(Supplied_Material, Pipe) |
DELIVERS_TO(Blood_Vessel, Tissue) | DELIVERS_TO(Pipe, Production_Equipment) |
In this example, corresponding predicates have been used, and the objects participating in those relationships can be aligned consistently.
However, representing two statements using the same predicate does not prove that the correspondence is correct.
The choice of predicates itself depends on the analyst’s interpretation and method of knowledge representation.
Three Important Principles of Structure Mapping
Structure-Mapping Theory and its computational implementations emphasize principles for evaluating relational correspondences.
Three particularly important ideas are relational focus, structural consistency, and systematicity.
| Principle | Meaning |
|---|---|
| Relational Focus | Emphasizes correspondences between relationships rather than individual attributes |
| Structural Consistency | Requires a mapping to avoid conflicting correspondences |
| Systematicity | Prefers interconnected systems of relationships over isolated matches |
Structural consistency includes two important constraints: one-to-one correspondence and parallel connectivity.
One-to-One Correspondence
Within a particular mapping, an element in the base domain should not be mapped indiscriminately onto multiple elements in the target domain.
For example, if the heart is mapped onto a pump, mapping the same heart simultaneously onto both the pump and a warehouse in the same structural alignment could create inconsistencies.
This constraint helps maintain a coherent correspondence.
However, it does not mean that one-to-many relationships cannot exist in real systems.
Nor does it prevent constructing alternative mappings from different perspectives.
One-to-one correspondence is a constraint governing the internal consistency of a particular structural mapping.
Parallel Connectivity
When two relations are mapped onto one another, their corresponding arguments must also be aligned appropriately.
For example, consider:
SUPPLY(Heart, Blood)
and
SUPPLY(Pump, Supplied_Material)
If the two SUPPLY relations correspond, the heart must correspond to the pump, and blood must correspond to the supplied material.
A mapping is not structurally coherent if the relational predicates correspond but their arguments are mapped inconsistently.
This principle is important because it distinguishes structural correspondence from simple keyword matching.
The Systematicity Principle
The Systematicity Principle is one of the central ideas of Structure-Mapping Theory.
It states that interconnected systems of relationships are preferred over collections of isolated relational matches.
Suppose two systems share the following characteristics:
“Both systems contain containers.”
“Both systems contain pathways.”
“Both systems contain controllers.”
These are similarities, but they may represent independent observations.
Now consider a different set of correspondences:
“A driving mechanism creates a pressure difference.”
“The pressure difference produces a flow.”
“The flow transports a substance to a destination.”
These relationships are connected within a common mechanism.
Structure-Mapping Theory emphasizes the importance of such systematic relational structures.
Higher-order relations are particularly important because they connect other relations, states, or events.
What Are Higher-Order Relations?
An ordinary relation can describe a connection between objects or quantities.
For example:
PRESSURE_DIFFERENCE(Source, Destination)
FLOW(Substance, Source, Destination)
A higher-order relation may represent how one state or event influences another.
For example:
CAUSE(Pressure_Difference, Flow)
Here, a causal relationship connects the states of pressure difference and fluid flow.
This notation simplifies the actual physical mechanism. In reality, flow also depends on factors such as hydraulic resistance, fluid properties, and boundary conditions.
The important distinction is that the system contains not merely individual objects or isolated relations, but relations that are themselves connected through other relationships.
Such higher-order relational structures can support richer analogies.
Candidate Inferences Generated Through Structure Mapping
Structure mapping does more than identify correspondences that are already known.
When a relationship is known in the base domain but has not been established in the target domain, the mapping may suggest that a corresponding relationship should be investigated in the target.
This is called a Candidate Inference.
For example, suppose the basic transportation relationships in blood circulation correspond to those of an industrial supply network.
Biological circulation includes mechanisms that regulate blood flow according to local conditions and demands.
Now suppose the industrial supply network does not have sufficiently developed local supply regulation.
The structural analogy might suggest the following hypothesis:
“Introducing supply regulation based on local demand could improve the efficiency or stability of the industrial supply network.”
This is a candidate inference projected from the base domain to the target.
However, the hypothesis may not hold in practice.
For example, local flow regulation could destabilize pressure throughout the network.
The cost of additional sensors and controllers might also exceed the economic benefits.
Candidate inferences are possibilities worth investigating, not automatically established facts.
A candidate inference generated through structure mapping is a prediction or hypothesis derived from an analogy. It is not knowledge that has already been verified in the target domain.
This distinction is particularly important when applying structure mapping to research and development.
A Practical Example — Comparing Supply Networks
Consider structural mapping between biological circulation and an industrial material supply network.
First, identify elements and relationships that may correspond between the two systems.
| Relationship | Biological Circulation | Industrial Supply Network |
|---|---|---|
| Driving transportation | Circulatory mechanisms generate flow | Pumps generate flow |
| Transportation pathways | Substances move through blood vessels | Substances move through pipes |
| Supply destinations | Substances are delivered to tissues | Materials are delivered to production equipment |
| Response to demand changes | Local blood flow is regulated through physiological processes | Supply may or may not be regulated according to local demand |
| Continuous supply | Circulation is maintained | Materials are replenished and transported from supply sources |
Relationships that are already known to exist in the target can form part of the structural alignment.
Mechanisms connected to this shared relational structure but not yet implemented or understood in the target may become sources of candidate inferences.
For example, if demand-responsive supply control is absent, the analogy might suggest introducing it.
However, this does not yet provide a specific control system design or establish that the proposed approach will be effective.
Structure mapping organizes correspondences and suggests what might be inferred from them.
Evaluating implementation requires knowledge of control engineering, fluid mechanics, equipment design, economics, and other relevant fields.
Surface Similarity, Analogy, and Abstraction
Gentner’s original theory distinguishes different forms of comparison according to which types of information are shared between objects.
| Type of Comparison | Main Characteristic |
|---|---|
| Surface Similarity | Objects share attributes such as color, shape, or material |
| Literal Similarity | Objects share both attributes and relational structure |
| Analogy | Correspondence between relationships is central |
| Abstraction | Particular examples are aligned with a more general relational structure |
For example, two similar kinds of industrial pumps may share many attributes and relationships, including shape, materials, functions, and operating principles.
By contrast, comparing a pump with the driving mechanisms of biological circulation emphasizes functional and relational correspondences over visual or material similarity.
At a more abstract level, one might describe both systems as:
“A supply source delivers resources to destinations through transportation pathways.”
This is a general relational description.
However, abstraction also removes specific conditions.
Discovering an abstract structure shared by two systems does not establish that the same practical solution can be implemented in both.
SME — A Computational Model of Structure Mapping
The Structure-Mapping Engine (SME) is a representative computational implementation of Structure-Mapping Theory.
It was developed by Brian Falkenhainer, Kenneth D. Forbus, Dedre Gentner, and colleagues. Its algorithm and applications were described in detail in a 1989 paper.
SME receives structured representations of a base domain and a target domain and searches for mappings between their relational structures.
Important outputs include:
| Output | Description |
|---|---|
| Correspondences | Which elements and relationships in the base correspond to those in the target |
| Candidate Inferences | Relationships that may be projected from the base to the target |
| Structural Evaluation Score | An evaluation of the structural quality and systematicity of the mapping |
A distinguishing feature of SME is that it does not merely search for similar objects.
It identifies combinations of relational correspondences that can be maintained consistently.
An Overview of SME Processing
Conceptually, SME processing can be explained through the following stages.
| Stage | Main Process |
|---|---|
| 1. Input | Receive structured knowledge representations of the base and target |
| 2. Generate Local Match Hypotheses | Identify relations and entities that may correspond |
| 3. Combine Correspondences | Construct combinations satisfying structural consistency |
| 4. Evaluate Global Mappings | Compare mappings according to systematicity and other structural criteria |
| 5. Generate Candidate Inferences | Identify relationships that might be projected onto the target |
This table is a simplified explanation rather than a complete description of SME’s internal algorithm.
SME combines local match hypotheses to construct globally consistent structural mappings.
These are commonly referred to as global mappings.
Structural evaluation scores allow alternative mappings to be compared.
What Does a Structural Evaluation Score Mean?
SME’s structural evaluation score measures aspects of the quality of a relational correspondence.
Mappings that preserve coherent and systematic structures generally receive stronger evaluations.
However, the score is not the probability that a proposed technology will work.
For example, a strong structural correspondence between biological transportation and an industrial supply mechanism does not establish that applying the biological principle will improve actual industrial performance.
Evaluating structural correspondence and evaluating scientific or technological feasibility are separate tasks.
MAC/FAC — Retrieving Analogical Sources from Large Knowledge Collections
SME addresses how to compare two structured knowledge representations once they are available.
However, analogical reasoning in practice involves an additional challenge.
How can potentially useful source examples be identified from a large collection of knowledge or memories?
A representative computational approach to this problem is MAC/FAC, presented by Kenneth D. Forbus, Dedre Gentner, and Keith Law in 1995.
MAC/FAC uses a two-stage retrieval architecture that narrows a large collection of candidates to a smaller number of promising examples.
The name stands for Many Are Called, Few Are Chosen.
The First Stage — MAC
The MAC stage uses a relatively inexpensive retrieval process to select candidates from a large knowledge collection.
In the original model, structured knowledge representations are converted into Content Vectors.
These vectors summarize features such as the predicates appearing in the knowledge representations.
This allows many candidate representations to be compared without performing detailed structural mappings for every candidate.
The Second Stage — FAC
The FAC stage subjects the smaller collection of selected candidates to more detailed structural evaluation using SME.
This enables the system to consider correspondences between relational structures rather than relying exclusively on coarse features.
| Stage | Purpose | Main Operation |
|---|---|---|
| MAC | Efficiently reduce a large candidate collection | Retrieval using computationally inexpensive feature representations |
| FAC | Evaluate promising candidates in greater detail | Structural comparison using SME |
MAC/FAC was developed as a computational model of how analogically useful examples might be retrieved from memory.
It was not originally designed simply as a commercial search engine.
Nevertheless, the general architecture of retrieving candidates broadly and then performing detailed relational comparisons on a smaller selection is relevant to modern knowledge exploration systems.
The Relationship to Modern Embedding-Based Retrieval
Modern AI systems frequently convert documents and other information into embedding vectors and retrieve items based on vector similarity or distance.
This approach is useful for efficiently identifying potentially related information in large collections.
MAC/FAC also uses vectors in its first stage, but these are not equivalent to modern neural embeddings.
The Content Vectors of MAC/FAC are derived from structured knowledge representations.
Modern embedding models typically use representations learned from data through machine learning.
The two approaches share the general purpose of narrowing candidate sets before expensive comparison, but they differ in representation and computation.
Furthermore, documents that are close together in an embedding space do not necessarily have corresponding relational structures.
For example, two academic papers may use similar technical terms while describing different mechanisms.
Conversely, papers from unrelated industries may use different terminology while describing systems with comparable functions or relational structures.
From a structure-mapping perspective, a possible architecture is therefore to use embedding retrieval to identify candidates and then extract and compare their functions and relationships.
Structure Mapping, Concept Maps, and Knowledge Graphs
Structure mapping requires representations of the relationships within the domains being compared.
Concept Maps and Knowledge Graphs offer possible ways to represent such information.
For example, a system might be described using the following propositions:
“A pump generates flow.”
“Flow transports a substance.”
“The transported substance is supplied to production equipment.”
These propositions form a relational knowledge structure describing a particular system.
If another system is represented in a comparable way, structural correspondences may be examined.
However, merely having Concept Maps or Knowledge Graphs does not mean structure mapping can be performed automatically.
The meanings and types of relations must be sufficiently consistent.
For example, consider:
“A transports B.”
“A carries B.”
“B is conveyed by A.”
These statements can express similar relationships, but their linguistic forms differ.
To support structure mapping, it may be necessary to normalize relationships and consistently represent argument roles and relational direction.
Moreover, causal relationships, correlations, part-whole relationships, and functional relationships should be distinguished.
Otherwise, mappings may be constructed between relations that are superficially similar but semantically different.
Challenges in Computational Structure Mapping
Although the principles of Structure-Mapping Theory are comparatively clear, applying them to large-scale real-world information presents several difficulties.
One of the most significant challenges is creating appropriate input representations.
SME receives structured knowledge representations.
However, corporate technical documents and scientific papers are generally written in natural language.
Accurately extracting entities, attributes, relations, and higher-order relations from such documents is not straightforward.
| Challenge | Description |
|---|---|
| Relation Extraction | Identify relationships between entities in natural-language documents |
| Semantic Normalization | Represent equivalent relationships consistently |
| Role Identification | Determine the roles played by elements within a system |
| Higher-Order Relation Extraction | Represent causation, control, dependence, and related structures |
| Context Preservation | Retain conditions, exceptions, and limitations |
| Scalability | Compare large collections of candidate structures efficiently |
| Inference Verification | Investigate whether hypotheses generated through analogy are valid |
Understanding the mechanisms described in technical documents may require substantial domain expertise.
For example, extracting only superficial descriptions of a material’s properties or behavior may omit conditions essential to understanding its mechanism.
Furthermore, the relationships relevant to a document may depend on the purpose of the analysis.
When exploring material technologies, relationships between materials and functions may be central.
When improving manufacturing processes, relationships between process conditions and product quality may be more important.
Consequently, converting every piece of information into a single comprehensive Knowledge Graph is not necessarily the most effective approach.
An alternative is to extract the relationships required for a particular analytical purpose.
The Difference Between GNG-Based Concept Structuring and Structure Mapping
Unsupervised learning methods such as Growing Neural Gas (GNG) can organize similar data and represent aspects of the distribution or topology of a feature space.
For example, large collections of technical documents can be converted into embedding vectors and organized into conceptual groups using GNG.
However, the structures generated through this process differ fundamentally from the relational structures addressed by Structure-Mapping Theory.
The neighborhood relationships in GNG represent similarity in a feature space.
Structure mapping, by contrast, examines correspondences between semantically meaningful relationships among objects.
| Perspective | GNG-Based Structuring | Structure Mapping |
|---|---|---|
| Input | Numerical vectors and related representations | Structured representations of entities and relations |
| Main Criterion | Similarity in feature space | Correspondence between relational structures |
| Output | Data distribution, clusters, neighborhood relationships | Correspondences, structural evaluations, candidate inferences |
| Main Function | Organizing and exploring large information collections | Comparing knowledge structures across domains |
| Main Limitation | Feature similarity does not establish semantic relationships | Suitable relational representations must be constructed |
Therefore, edges connecting nodes in a GNG network should not automatically be interpreted as semantic relationships in a Knowledge Graph.
However, combining the two approaches may be worthwhile.
For example, GNG could organize a large collection of technical documents, after which the relevant conceptual groups could be selected for further analysis.
Functions, mechanisms, and constraints could then be extracted from the selected documents and compared using structure-mapping principles.
This would combine large-scale information organization with more detailed semantic comparison.
Whether the approach is effective must be evaluated using real exploration tasks and suitable datasets.
Combining LLMs with Structure Mapping
Large Language Models can extract entities and relationships from natural-language documents and compare examples from different domains.
They may therefore provide a way to construct the relational representations required for computational structure mapping.
For example, the following information could be extracted from technical documents:
- Functions provided by the technology
- Materials and equipment used
- Mechanisms responsible for the functions
- Inputs and outputs
- Causal, control, and dependency relationships
- Conditions and constraints under which the technology operates
If these elements are represented consistently, relationships between documents from different fields may become easier to compare.
However, LLM-based relation extraction is not always accurate.
An LLM may describe equivalent relationships using different predicates.
Conversely, excessive abstraction may create correspondences between technologies that do not share meaningful mechanisms.
For example, almost every technology can be represented through the general structure:
“Receive an input, perform a process, and produce an output.”
But such a broad description is usually insufficient for identifying technologically useful analogies.
It is therefore important to control the level of abstraction and preserve relevant evidence, mechanisms, and conditions.
Two possible computational approaches should also be distinguished.
In the first, an LLM performs the structural comparison and generates mappings directly.
In the second, an LLM constructs structured relational representations, which are then compared using a separate algorithm such as SME.
The first approach may provide greater flexibility, but consistency and reproducibility may be difficult to evaluate.
The second makes the mapping rules more explicit, but requires sufficiently accurate and compatible input representations.
The appropriate approach depends on the characteristics of the information, accuracy requirements, computational resources, and verification methods.
Applications to Cross-Domain R&D Exploration
When searching for new research and development themes, it can be valuable to investigate functions and mechanisms found outside established fields of expertise.
However, conventional searches based on familiar technical terminology or industrial classifications may fail to identify relevant examples from unfamiliar domains.
Structure-Mapping Theory provides one theoretical perspective for addressing this problem.
For example, suppose an R&D challenge requires a system capable of repeatedly delivering substances to locations where they are needed.
Rather than searching only for particular materials or equipment, the problem could be represented through relationships such as supply, transportation, demand variation, control, and replenishment.
Candidate systems sharing some of these relationships could then be investigated across different industries and biological domains.
When relevant examples are found, their relational structures can be mapped to the target problem.
This may suggest mechanisms that could be adapted or applied.
However, Structure-Mapping Theory does not itself provide a complete method that automatically discovers valuable new R&D themes.
Additional decisions are required concerning problem formulation, information retrieval, relation extraction, relevant constraints, and verification.
The evaluation of new R&D themes must also consider technical feasibility, applications, market demand, competing technologies, intellectual property, and cost.
Structure mapping is best understood as one mechanism for connecting knowledge from different domains and generating hypotheses.
Limitations of Structure-Mapping Theory
Structure-Mapping Theory is important for understanding analogical reasoning, but several limitations must be recognized.
First, the mapping depends on the knowledge representations provided as input. Different selections of attributes and relations may produce different mappings of the same objects.
Second, structural correspondence does not establish that a mechanism will operate under the conditions of the target domain. Differences in materials, environments, scale, time, and cost may prevent practical transfer.
Third, a high structural evaluation score does not guarantee technological usefulness. A correspondence may be structurally coherent while offering little practical value for research or development.
Fourth, computational models such as SME require structured representations. Constructing those representations from natural-language documents is a separate problem that must be addressed.
Fifth, if potentially useful sources are overlooked during retrieval, even an accurate mapping algorithm cannot evaluate those missing candidates.
Sixth, the relationships that matter in complex real-world systems depend on the analytical purpose. A single fixed representation may not be suitable for every analogical comparison.
Finally, Structure-Mapping Theory provides a cognitive scientific account and computational models of important aspects of analogy, but it is not the only theory relevant to analogical reasoning or human creativity.
Structure-Mapping Theory identifies important structural principles supporting analogy. It does not establish a universal law guaranteeing successful discovery or invention through analogical reasoning.
Structure-Mapping Theory in ThinkNavi
ThinkNavi treats Structure-Mapping Theory as a cognitive scientific foundation for understanding analogical thinking.
The separate page on Analogy introduces practical applications, including understanding, idea generation, problem-solving, and cross-domain exploration.
This page focuses on the theory and computational mechanisms underlying relational correspondence.
For ThinkNavi Free, a possible implementation would provide worksheets allowing users to describe entities and relations in two domains and organize their correspondences manually.
For example, users could record the elements, functions, relationships, and operating conditions of both the base and target domains.
They could then identify which relationships correspond, where important differences exist, and what new hypotheses might follow from the comparison.
Even a relatively simple approach could help transform analogy from an informal association or metaphor into a more explicit, examinable comparison of relationships.
Potential LLM-assisted capabilities include extracting relationships from technical documents, generating candidate correspondences, and comparing structural consistency and applicability conditions.
More specialized functionality could involve algorithms based on computational structure mapping.
However, such implementations would require evaluation of the quality of the input representations and their computational requirements.
For ConceptMiner R&D, a potential direction is to combine machine-learning methods that organize large collections of technical information with mechanisms for comparing semantically meaningful relational structures.
For example, GNG could organize many technical documents into conceptual groups.
Relationships relevant to a particular exploration task could then be extracted from selected documents and compared.
Rather than maintaining every relationship in a single massive Knowledge Graph, it may be possible to construct local relational structures according to the current exploration objective.
Candidate inferences produced through structural correspondence could also be stored separately from established facts and linked to their underlying technical sources for subsequent verification.
This combination represents one possible technological direction for retrieving candidate analogical sources and organizing new hypotheses.
However, GNG-based similarity analysis, relation mapping through structure-mapping principles, and LLM-based hypothesis generation are different computational processes.
Whether combining them produces an effective cross-domain exploration system must be demonstrated through evaluation using specific problems and datasets.
These capabilities represent possibilities for future development and do not imply that all such functions are currently available.
The significance of Structure-Mapping Theory lies in providing a clear explanation of why meaningful analogy often depends on correspondences between relationships and their systematic organization, rather than merely on matching superficial attributes.
Its development into computational models such as SME and MAC/FAC is also important for considering how computers may support analogical reasoning.
However, similarity of structure is not the same as equivalence of underlying mechanisms.
Candidate inferences generated through analogy require examination of their evidence and conditions of validity.
Structure-Mapping Theory explains how correspondences between different knowledge structures can support new inferences. It provides a foundation for extending knowledge into unfamiliar domains, but the validity of the resulting hypotheses must be established through further investigation and verification.
Related Topics
- Analogy — Connecting Knowledge Across Domains to Generate New Ideas and Hypotheses
- Concept Formation | How Categories Are Made
- Induction, Deduction, and Abduction — Discovering Patterns, Drawing Conclusions, and Generating New Hypothese
- The Scientific Method and Falsifiability — Generating Hypotheses and Testing Them Critically
- Concept Maps — Representing Relationships Between Concepts and Understanding Knowledge Structures
- Synectics | Discovering Solutions Across Domains Through Analogy
- TRIZ | A Systematic Approach to Inventive Problem Solving Through Technical Contradictions
- The KJ Method | Building Meaningful Structures from Fragmented Information
- Problem Structuring — Breaking Down Problems, Understanding Relationships, and Finding Paths to Solutions
- Systems Thinking | Interdependence and Change Over Time
- Logical Thinking — Building Sound Arguments, Communicating Reasoning, and Achieving Mutual Understanding
- Critical Thinking | Examining Claims and Evidence
- Morphological Analysis | Explore Combinations
- Concept Exploration
References and Further Reading
Stanford Encyclopedia of Philosophy — Analogy and Analogical Reasoning
Dedre Gentner, “Structure-Mapping: A Theoretical Framework for Analogy,” Cognitive Science, 7(2), 1983, pp. 155–170.
Brian Falkenhainer, Kenneth D. Forbus, and Dedre Gentner, “The Structure-Mapping Engine: Algorithm and Examples,” Artificial Intelligence, 41(1), 1989, pp. 1–63.
Kenneth D. Forbus, Dedre Gentner, and Keith Law, “MAC/FAC: A Model of Similarity-Based Retrieval,” Cognitive Science, 19(2), 1995, pp. 141–205.
Dedre Gentner and Arthur B. Markman, “Structure Mapping in Analogy and Similarity,” American Psychologist, 52(1), 1997, pp. 45–56.
Arthur B. Markman and Dedre Gentner, “The Effects of Alignability on Memory,” Psychological Science, 8(5), 1997, pp. 363–367.
Arthur B. Markman and Dedre Gentner, “Structure Mapping in the Comparison Process,” The American Journal of Psychology, 113(4), 2000, pp. 501–538.
Dedre Gentner and José Medina, “Comparison and the Development of Cognition and Language,” Cognitive Studies: Bulletin of the Japanese Cognitive Science Society, 4(1), 1997, pp. 112–149.
Arthur B. Markman, “Structural Alignment in Similarity and its Influence on Category Structure,” Cognitive Studies: Bulletin of the Japanese Cognitive Science Society, 4(4), 1997, pp. 19–37.
Mary L. Gick and Keith J. Holyoak, “Analogical Problem Solving,” Cognitive Psychology, 12(3), 1980, pp. 306–355.
Mary B. Hesse, Models and Analogies in Science, University of Notre Dame Press, 1966.
Keith J. Holyoak and Paul Thagard, Mental Leaps: Analogy in Creative Thought, MIT Press, 1995.
Douglas Hofstadter and Emmanuel Sander, Surfaces and Essences: Analogy as the Fuel and Fire of Thinking, Basic Books, 2013.
Qualitative Reasoning Group, Northwestern University — Structure-Mapping Research