{"path":"research/corpus-graph/what-research-says-about-finding-far-analogies.md","content":"# What research says about finding far analogies (2026-10-10)\n\n*Looked up online on the night of 2026-10-10, after the founder asked how the matching of distant\nconnections between this project's notes could be improved (*\"Think deeply about how to improve this\nconnection matching\"*). Read depth: the abstract and the passages a search returned for each paper,\nfrom the publisher's page, the authors' copy or arXiv. No paper was read in full. Where a number is\ngiven it stood in the passage returned. What it means for the index of notes is in\n[README.md](README.md) ยง 18.*\n\n## 1. People fail at the same step our readers fail at: retrieval, not judgment\n\n- **Gentner, Rattermann & Forbus (1993)**, *The Roles of Similarity in Transfer*, Cognitive Psychology\n 25. People were reminded of an earlier story by shared characters and settings far more often than by\n shared causal structure, and then *rated* the structural matches as the sounder ones. A later summary\n by Gentner (2002) puts it at two to five times more likely for surface matches. Their model is named\n for the two stages: many are called by a cheap match, few are chosen by a structural one (MAC/FAC).\n \n- **Gick & Holyoak (1980, 1983)**, as reported in the same papers: about one in ten solved the\n radiation problem alone, about three in ten after reading an analogous story, and eight or nine in\n ten once told to think back to that story. The story was in memory. It was not retrieved.\n\n**For us.** A reader who has both passages open judges kinship easily. The loss is in getting the\nsecond passage in front of the reader at all. More reading with the same way of searching will not\nclose it.\n\n## 2. What makes a far case retrievable\n\n- **Comparing two cases makes a schema, and the schema is the better probe.** Gentner, Loewenstein,\n Thompson & Forbus (2009), *Reviving Inert Knowledge*, Cognitive Science. People who compared two\n analogous cases at recall time retrieved more analogous cases from their own past than people who\n read the same two cases without comparing them. Reading two cases one after the other gave no\n advantage over reading one. The authors' account: the schema has fewer features that fail to match\n and gives the relations more weight, so it matches a stored case better than either example does,\n and it works both forward and backward in time.\n \n- **A relational label does the same.** Jamrozik & Gentner (2020), *Relational labeling unlocks inert\n knowledge*. Giving a situation a label that names its relational pattern (their example\n is *positive feedback loop*), at the time it is stored or at the time of recall, raised retrieval of\n other cases of the pattern. Labels naming the domain did not help. Their explanation of expert\n retrieval: experts store cases in a uniform relational vocabulary, so a new case overlaps with a\n stored one.\n \n- **The wording has to be the same wording.** Gentner (2002) reports Clement and colleagues (1994):\n retrieval was more likely when the two stories used synonymous terms for the relation than when they\n used loosely similar ones, while mapping, with both stories present, was hardly affected.\n \n\n**For us.** Three things follow. A name for the shared pattern is the strongest known aid to finding\nits cases. A probe built from two known cases finds a third better than either case does. And a\nsubject-free sentence written freely for each passage will vary in wording and miss its partner, so\nthe sentences need a small fixed vocabulary of moves, which is what the list of named ideas is.\n\n## 3. Machines: compare a short statement of purpose and mechanism, never the whole text\n\n- **Hope, Chan, Kittur & Shahaf (2017)**, *Accelerating Innovation Through Analogy Mining*, KDD. Product\n descriptions were reduced to two parts, the purpose and the mechanism, and matched on those. The\n authors report higher precision and recall than matching on the text (TF-IDF, LSA, GloVe, LDA), and\n that analogies found this way raised the share of ideas rated good by at least a quarter.\n \n- **Chan, Chang, Hope, Shahaf & Kittur (2018)**, *SOLVENT*, CSCW. Research papers were annotated with\n background, purpose, mechanism and findings, at about a minute a paper. Matching on those parts found\n analogies between papers about 25 to 30 percent better than a standard text method, across fields as\n well, and annotators outside the field could do it. Leaving the *background* part out of the match\n helped, since the background is the field.\n \n- **Gilon, Chan, Ng, Lifshitz-Assaf, Kittur & Shahaf (2018)**, *Analogy Mining for Specific Design\n Needs*, CHI, citing Yu and colleagues: abstraction has to be aimed. A problem stated abstractly with\n its key constraint kept concrete found useful far cases; a problem made abstract in every part found\n cases that could not be used.\n \n- **Kang and colleagues (2022)**, *Augmenting Scientific Creativity with an Analogical Search Engine*:\n the same idea over about 1.7 million papers. A search for analogies has to keep a\n distance from the query, since the nearest match is the same thing again, and the useful matches sat\n at a middle distance.\n\n**For us.** The form to compare is short and has parts: what the passage is trying to achieve, the\nmove it makes, and the condition under which the move works, with the field left out. The condition\nstays concrete. And the candidates worth a reader's time are the ones near in that form and far in\ntheir words.\n\n## 4. Language models: weak at exactly the long version of this task, and a larger model does not fix it\n\n- **AnaloBench** (EMNLP 2024; its code is published by the language-technology group at Johns Hopkins). Models were asked to pick the analogous story among\n four, and to find analogous stories in a pool of about two hundred. With one-sentence stories the\n best model was within about seven points of people. With thirty-sentence stories the gap was about\n twenty-nine points, and larger models gained little. Finding analogues in the pool of two hundred\n was near the level of picking at random for long stories.\n \n- **StoryAnalogy** (Jiayang and colleagues, 2023). Sentence embeddings tracked likeness of entities\n and relations and did poorly on analogy; a model answered about three multiple-choice items in ten\n where people answered more than eight in ten.\n \n- **A 2026 preprint on narrative analogy** (arXiv 2603.29997). Mapping structure on raw event phrases\n did worse than chance. Mapping after each event had been replaced by an abstraction did better than\n the model alone on far analogies, and worse on near ones.\n \n\n**For us.** This is the likeliest reason our readers returned kinship in one finding of sixteen: they\nwere holding eleven long notes and looking for analogues among four hundred more, which is the\ncondition where models are near chance. The remedy the papers agree on is to shorten each side to its\nabstraction first and match those. A stronger model is not the lever.\n\n## 5. What was not looked up\n\nWork on structure mapping proper (the engine that aligns two relational descriptions), on case-based\nreasoning's indexing by goals, and on how patent and legal search handle the same problem. And none of\nthe above was checked against a replication.\n"}