Relational Structure Mapping

145 concepts106 papers1 proposal

If you ask a standard machine learning model to find something similar to a "computer virus," it will likely point you toward "malware," "trojan horse," or "cybersecurity." This happens because standard models rely heavily on surface-level features, literal vocabulary overlap, or close semantic attributes.

However, a human can easily map a computer virus to a biological disease. We recognize that both infect a host, replicate autonomously, cause systemic harm, and can be mitigated by specialized defenses (antivirus software versus an immune system).

This cognitive leap is made possible by Relational Structure Mapping. Instead of comparing the literal attributes of entities, structure mapping focuses entirely on the network of relationships, causal chains, and functional roles that connect them. It allows intelligent systems to perform abstract reasoning and transfer knowledge across completely disjoint domains.

The Failure of Surface Similarity

Why do traditional embedding models struggle with abstract analogies? As noted by the authors of Life is a Circus and We are the Clowns: Automatically Finding Analogies between Situations and Processes, standard sentence embedding models prioritize attribute similarity and word overlap. While this makes them exceptionally good at identifying similarities within the same topic, it renders them ineffective at discovering "far analogies" across unrelated domains where the vocabulary differs entirely.

To find deep analogies, a system must look past what an entity is and focus on how it functions within its environment.

Surface Matching (Fails across domains):
[Computer Virus] <--- No Semantic Overlap ---> [Biological Pathogen]

Structure Mapping (Succeeds across domains):
[Host PC] ------(infected by)-----> [Computer Virus]
   |                                      |
(Equivalent Role)                  (Equivalent Role)
   v                                      v
[Human Body] ----(infected by)-----> [Biological Pathogen]

By shifting the focus from entity attributes to relational roles, systems can identify functional counterparts in entirely separate fields. For instance, in domain adaptation for sentiment analysis, the authors of Unified Feature and Instance Based Domain Adaptation for Aspect-Based Sentiment Analysis note that robust cross-domain reasoning requires recognizing conceptual analogies—such as mapping the functional role of "service personnel" in a restaurant review to "technical support" in a laptop review.

The Mechanism of Structural Alignment

To formalize this process, relational structure mapping treats domains as networks of relational predicates. An analogy is established when a system finds a structural isomorphism—a mapping that preserves the internal relational web of the source domain within the target domain, even if zero low-level features are shared.

In Transfer Learning by Structural Analogy, Hua-Yan Wang et al. demonstrate that knowledge transfer between representationally disjoint domains is highly successful when the mapping preserves the internal relational structure of each system. Rather than looking for literal feature overlaps, the model optimizes for how features interact within their own respective systems.

This alignment operates on three core principles:

  1. Functional Role Equivalence: Elements are mapped if they occupy equivalent positions in their respective structures. As Aaron Stockdill et al. show in Cross-domain Correspondences for Explainable Recommendations, this mechanism allows systems to translate abstract concepts into alternative representations—such as mapping algebraic operations to visual diagrams—based purely on structural similarity.
  2. Preservation of Causal Logic: The mapping must prioritize higher-order relations like implications and causal chains over simple object descriptions. Joseph A. Blass et al. demonstrate in Moral Decision-Making by Analogy: Generalizations versus Exemplars that evaluating similarity based on causal chains and logical dependencies ensures that scenarios with similar surface elements but different logical implications are correctly distinguished.
  3. Systematicity: The mapping favors deeply interconnected webs of relations over isolated matches. In Multimodal Analogical Reasoning over Knowledge Graphs, Ningyu Zhang et al. show that prioritizing the modeling of relational structures between entities rather than focusing on individual entity attributes leads to superior performance in complex reasoning tasks.

Mathematical and Algorithmic Implementations

Modern architectures implement structure mapping through several distinct mathematical frameworks.

Vector Symbolic Architectures

One approach is to design high-dimensional vector spaces that explicitly encode structural relationships. Emre Neftci et al. demonstrate in A Grid Cell-Inspired Structured Vector Algebra for Cognitive Maps that structuring high-dimensional vectors into modules mimicking biological grid cells allows models to perform abstract relational inference, such as mapping hierarchical family trees, using the same algebraic framework used for spatial path integration.

Graph-Based Abstraction

Another common paradigm represents environments as graphs to compute systematic correspondences. Melanie Mitchell in Abstraction for Deep Reinforcement Learning argues that representing environments as relational graphs allows agents to identify systematic correspondences between objects and their relationships across entirely different scenarios, bypassing the limitations of standard deep reinforcement learning.

Decoupled Optimization Metrics

When training or evaluating models on analogical tasks, the objective function must explicitly separate structure from content. The authors of StoryAnalogy: Deriving Story-level Analogies from Large Language Models to Unlock Analogical Understanding formalize this by defining an analogical strength metric that explicitly maximizes relational similarity (the underlying plot or logical structure) while minimizing entity similarity (the specific characters or settings).

$$\text{Analogical Strength} \propto \text{Similarity}(\text{Relations}) - \text{Similarity}(\text{Entities})$$

This decoupling ensures that the system is rewarded for finding deep structural mappings rather than getting distracted by surface-level attribute overlaps.

Unlocking Cross-Domain Transfer

The primary reason relational structure mapping works so well is that it enables zero-shot generalization to novel environments.

In reinforcement learning, Peter Stone demonstrates in Value-Function-Based Transfer for Reinforcement Learning Using Structure Mapping that task alignment can be achieved by identifying correspondences between relational structures rather than matching specific variable names. State variables are treated as equivalent if they share the same causal relationships with other components in the environment, allowing an agent to instantly transfer strategies from a simple source task to a complex target task.

This same principle applies to how large language models handle complex workflows. Jinglong Gao et al. show in ExpeTrans: LLMs Are Experiential Transfer Learners that transferring task-solving experiences is significantly more effective when based on the alignment of execution steps and underlying goals rather than surface-level semantic similarity.

By forcing models to align the process and the logic rather than the literal text, relational structure mapping provides a robust blueprint for true, human-like abstract reasoning.

Key Foundational Work

Analogies in Spontaneous Discourse (1981) — contributed 2 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 106 papers
  2. Version 1 · 2026-10-04 · 145 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_relational_structure_mapping_43dcaf52,
  title = {relational structure mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/relational-structure-mapping/rdr_43dcaf52858e46f98a7e58885aee0854},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_relational_structure_mapping_43dcaf52_v2,
  title = {relational structure mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/relational-structure-mapping/rdr_43dcaf52858e46f98a7e58885aee0854/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake