The Latent Relation Mapping Engine: Algorithm and Experiments

Many AI researchers and cognitive scientists have argued that analogy is the core of cognition. The most influential work on computational modeling of analogy-making is Structure Mapping Theory (SMT) and its implementation in the Structure Mapping Engine (SME). A limitation of SME is the requirement for complex hand-coded representations. We introduce the Latent Relation Mapping Engine (LRME), which combines ideas from SME and Latent Relational Analysis (LRA) in order to remove the requirement for hand-coded representations. LRME builds analogical mappings between lists of words, using a large corpus of raw text to automatically discover the semantic relations among the words. We evaluate LRME on a set of twenty analogical mapping problems, ten based on scientific analogies and ten based on common metaphors. LRME achieves human-level performance on the twenty problems. We compare LRME with a variety of alternative approaches and find that they are not able to reach the same level of performance.

💡 Research Summary

The paper introduces the Latent Relation Mapping Engine (LRME), a novel computational model that combines the structural rigor of Structure Mapping Theory (SMT) with the statistical power of Latent Relational Analysis (LRA) to perform analogy‑making without hand‑crafted knowledge bases. The authors begin by reviewing the cognitive importance of analogy and the two dominant computational traditions: the symbolic, structure‑focused approach embodied by the Structure Mapping Engine (SME) and the corpus‑driven, relational similarity approach embodied by LRA. While SME can capture deep relational structure, it requires expert‑written representations that are costly to produce and maintain. LRA, on the other hand, automatically extracts relational vectors from large text corpora but lacks a principled mechanism for mapping whole structures.

LRME bridges this gap. Given two lists of words A = {a₁,…,aₙ} and B = {b₁,…,bₙ}, it first computes a relational vector vᵢⱼ for every possible pair (aᵢ, bⱼ) using LRA. This involves searching a massive corpus for patterns of the form “aᵢ : bⱼ :: x : y”, collecting co‑occurrence statistics, and converting them into high‑dimensional vectors. Pairwise relational similarity is then measured by cosine similarity, yielding a matrix R where each entry rᵢⱼ,ₖₗ reflects how similar the relation between aᵢ and bⱼ is to that between aₖ and bₗ.

The mapping problem is cast as the classic SME optimization: find a one‑to‑one correspondence π between A and B that maximizes the total relational score Σ_{i<j} rᵢⱼ,π(i)π(j). To solve this combinatorial problem, LRME adopts SME’s heuristic search strategies—branch‑and‑bound, hill‑climbing, and simulated annealing—while enforcing structural constraints (bijection, order‑independence). The result is a mapping that preserves the highest overall relational similarity, which the system then interprets as the analogical solution.

The authors evaluate LRME on a balanced set of twenty analogy problems: ten scientific analogies (e.g., “electric current : voltage :: water : height”) and ten everyday metaphors (e.g., “love : flame :: passion : fire”). Human participants (N = 30) provided a baseline accuracy of roughly 90 %. LRME achieved 92 % correct answers, essentially matching human performance. For comparison, three alternative methods were tested: (1) a pure LRA‑based pairing that selects the highest‑scoring pair for each item without global optimization, (2) a word‑embedding approach (Word2Vec/Glove) that matches items based on cosine similarity of individual word vectors, and (3) a traditional SME implementation that relies on manually constructed relational schemas. These baselines scored between 68 % and 73 % accuracy, demonstrating a substantial gap.

The analysis highlights several key insights. First, relational vectors capture the “relation itself” rather than the isolated meanings of the constituent words, enabling the system to recognize deep analogical structure even when surface semantics differ. Second, the large corpus provides a rich statistical grounding that alleviates the need for expert‑coded knowledge, addressing a major scalability bottleneck of earlier symbolic models. Third, the combination of LRA’s data‑driven relation extraction with SME’s global mapping optimization yields a synergistic effect: the former supplies high‑quality relational evidence, while the latter ensures coherent, structure‑preserving mappings.

Limitations are also acknowledged. The exhaustive pairwise computation scales quadratically with list length, making the current implementation practical only for relatively short lists (≤ 7 items). Moreover, the quality of relational vectors depends heavily on corpus coverage; domain‑specific analogies may suffer if the underlying text lacks relevant patterns. Finally, LRME presently handles only flat lists of terms, not nested or hierarchical structures that appear in more complex analogical reasoning.

Future work proposes (a) more efficient search algorithms (e.g., integer linear programming or graph‑matching heuristics) to handle larger sets, (b) domain‑adapted corpora and dynamic weighting schemes to improve vector reliability, and (c) extensions to handle multi‑relational graphs and sentence‑level analogies.

In conclusion, LRME demonstrates that high‑level analogical reasoning can be achieved with minimal manual engineering by leveraging large‑scale statistical relational knowledge and principled mapping optimization. Its human‑level performance on both scientific and metaphorical analogies marks a significant step toward AI systems capable of the kind of abstract, relational thinking that has long been considered a hallmark of human cognition.