AI Glossary / Retrieval and knowledge systems / Intermediate
Literature-based discovery
Definition
Literature-based discovery is the use of published research to identify potentially new connections or hypotheses by bringing together findings that have not been explicitly linked. It combines information across papers or research areas to suggest questions worth investigating. A proposed connection remains a hypothesis until appropriate evidence supports it.
What it means
Relevant knowledge can be scattered across separate publications, disciplines and vocabularies. One set of papers may describe a relationship between A and B, while another describes a relationship between B and C. Considering them together may suggest a possible connection between A and C.
This is more than summarizing an individual article, but it is not automatic proof of a new scientific fact. The proposed bridge can fail because the studies concern different conditions, the relationships have different directions or an extracted statement was misunderstood. Discovery requires interpretation of the underlying evidence.
How it works
A typical workflow gathers a defined literature collection, identifies concepts and extracts or represents relationships. It searches for potentially useful links across that material and ranks candidates for expert examination. Techniques can include term associations, knowledge graphs and learned representations.
In open discovery, a starting concept is given and the system searches for promising connected targets. In closed discovery, both endpoints are specified and the system looks for intermediate concepts or evidence that could connect them. Searching for a bridge does not itself verify the hypothesis.
Reviewers examine source passages, study conditions, contradictory findings and whether the connection was already known. Evaluation must also prevent later publications from leaking into an experiment that claims to predict an earlier discovery.
Three practical examples
1. A materials-research team searches separate literatures on surface properties and manufacturing defects to identify candidate mechanisms for further experiments.
2. An environmental research group explores connections between a known process and outcomes studied under different terminology in another field.
3. A research-support tool helps experts inspect intermediate concepts linking two specified topics, preserving citations so that each proposed relationship can be checked.
Think of it this way
Think of two researchers each holding one piece of a possible explanation. One knows that a process changes a material property; another knows that the property can affect an outcome. Bringing their papers together may reveal a useful question. The pieces still have to fit under the same conditions, and a plausible story needs testing before it becomes an accepted explanation.
Worked example
Consider an invented literature graph about packaging materials. Let A be a treatment process, B1 be surface texture, B2 be moisture uptake, and C be seal durability. Suppose the reviewed graph contains A-B1 and A-B2 links. It also contains B1-C, while no B2-C link has been identified in the collection.
The graph has one two-step path from A to C: A-B1-C. If another candidate outcome D has links from both B1 and B2, then there are two two-step paths from A to D: A-B1-D and A-B2-D.
A simple exploratory ranking by the number of intermediate concepts would therefore assign C a score of one and D a score of two. This is a deliberately crude ranking rule, not a standard proof or a probability calculation.
The two paths to D may rely on weak or duplicated evidence, while the single path to C may have stronger support. Neither score establishes that A causes either outcome. Reviewers must check the meaning, direction and conditions of every relationship, then decide which hypothesis merits further investigation.
Illustrative case study
Fictional case: Ash Materials uses a literature-discovery prototype to explore ways of improving packaging durability. The system proposes a bridge between a treatment and an outcome through a shared material property.
Experts read the cited studies and discover that one link was measured under conditions unlike those used in the other study. They retain the proposal as a qualified research question and design a small experiment to examine whether the conditions can be reconciled. The tool helps prioritize reading and experimentation, but the team does not describe the suggestion as a validated improvement.
Common misconceptions and mistakes
Two related statements do not automatically compose into a valid causal conclusion. A shared term can have different meanings, and an association may not represent a mechanism. Extraction can also miss negation, uncertainty or the fact that a paper is discussing another study's claim.
A connection absent from one database is not necessarily new to science. Duplicate publications and repeated citations can exaggerate apparent support. Closed discovery is sometimes described as hypothesis testing, but finding literature bridges is not a substitute for experimental or other domain-appropriate validation.
Business use
Literature-based discovery can support research planning, technology scouting and the identification of questions that span disciplines. Its practical value lies in directing expert attention toward plausible, traceable connections rather than automatically declaring discoveries.
Define the source collection, date cutoff and concept-mapping rules. Retain links to supporting and conflicting passages, and assess how much expert review is required per useful hypothesis. Evaluate novelty separately from plausibility and validation. If a generative model proposes a connection, verify that its citations exist and that the cited content supports each stated link.
People also ask
Is it the same as a literature review?
A review organizes existing findings, while literature-based discovery specifically seeks potentially new connections or hypotheses from them.
What is the difference between open and closed discovery?
Open discovery searches for targets from a starting concept; closed discovery searches for bridges between specified endpoints.
Does a high-ranked hypothesis count as a discovery?
It is a candidate for investigation. Ranking alone does not establish novelty, correctness or practical usefulness.
Check your understanding
A system finds that A is associated with B and B is associated with C. Can it conclude that A causes C?
Answer: No. The relationships may not be causal or compatible, and a shared intermediate concept does not establish transitive causation. The sources and proposed mechanism require further examination.
Related terms
- Information extraction
- Entity linking
- hypothesis generation
- knowledge graph
- scientific text mining
- evidence provenance
- link prediction and research validation
- traced and assessed
Learn next
construct a small graph from papers you can read fully. Record the exact evidence and conditions for every edge, then identify a candidate bridge and search deliberately for contradictory findings or prior publication of the same connection.
More in Retrieval and knowledge systems
- Inverted index - An inverted index is a search data structure that maps terms to the documents or records in which they occur.
- Item-item collaborative filtering - Item-item collaborative filtering is a recommendation approach that uses relationships between items inferred from users' interactions or ratings.
- Learned sparse retrieval - Learned sparse retrieval is a family of search methods that train models to produce sparse representations for matching queries and documents.
- Location-based recommendation - Location-based recommendation is the selection or ranking of suggested items using geographical context as part of the decision.
- Long tail - The long tail is the large collection of relatively infrequent items or events beyond a small group of highly popular ones in a ranked distribution.
- Low-dimensional embedding - A low-dimensional embedding represents an item using relatively few numerical coordinates while aiming to retain relationships useful for a task.
Sources
Primary sources: Swanson, Undiscovered Public Knowledge, describing the connection of separate published findings: https://www.journals.uchicago.edu/doi/pdf/10.1086/601720
Neural networks for open and closed Literature-based Discovery, an original study of graph-based approaches and evaluation: https://pubmed.ncbi.nlm.nih.gov/32413059/
Using the Literature-Based Discovery Paradigm to Investigate Drug Mechanisms, an original application explaining candidate relationships and further research: https://pmc.ncbi.nlm.nih.gov/articles/PMC2655783/
Previous: Literal Next: Ljung-Box test