Boolean vs semantic search for literature

A Boolean query is a set of literal string matches joined by operators. Its guarantee is narrow and valuable: every record containing your terms in the fields you searched will be returned, and the query can be written down, re-run and audited by someone else.

That reproducibility is why systematic reviews are built on Boolean queries and not on anything else. A review method section reports the exact string, and a reader can execute it. No semantic system currently offers an equivalent.

Its weakness follows from the same property. The query returns what you asked for, and you can only ask for words you already know. Every synonym you did not list, every regional spelling, every term the field used before it settled on the current one, is silently absent from the results.

• Controlled vocabulary (MeSH in PubMed, Emtree in Embase) partly fixes this by mapping synonyms onto one indexed concept — but only within databases that maintain a thesaurus, and only for concepts a human indexer has already assigned.
• Truncation and wildcards catch morphological variants, not conceptual ones. They find modelling from model*; they do not find simulation.
• Proximity operators tighten precision but do nothing for recall.

Semantic retrieval represents your query and each document as vectors in a space where proximity approximates similarity of meaning, then returns the nearest documents. Nothing is matched literally, so a query about degradation under thermal cycling can return a paper that only ever writes about fatigue from repeated heating.

In practice its most useful property is not the results themselves but the vocabulary they expose. Run a prose description of your problem, read the ten nearest papers, and you have the terms of art the field actually uses — which is the input a good Boolean query needed all along.

The limitation is that recall is unmeasurable. A ranked list of the nearest fifty documents says nothing about what sat at position fifty-one, and there is no query string a reader can re-run to check. If your method section has to defend completeness, semantic retrieval cannot carry that on its own.

The sequence that works is semantic first, Boolean second, semantic again at the end.

1. Describe the question in prose and run it semantically. Do not evaluate the results as findings yet — read them for vocabulary.
2. Harvest terms: the title and keyword phrases of the on-target hits, the controlled-vocabulary headings assigned to them, and the terms their abstracts use for your outcome.
3. Build the Boolean query from those harvested terms rather than from your own. This is the step that raises recall, and it is the whole reason for step one.
4. Run the Boolean query in the databases whose coverage you need to be able to claim. Log the exact strings.
5. After screening, run one more semantic pass using the two or three best papers you found as the query. Anything genuinely new that surfaces is a sign the Boolean vocabulary is still incomplete.

If step five keeps returning new relevant work, do not stop. Go back to step two with the new terms. That loop is the practical definition of a search that has not yet saturated.

Boolean alone is defensible when you are working inside a mature field with a maintained thesaurus, on a question whose vocabulary is settled, and the deliverable is an auditable search. Pharmacovigilance and clinical review work often fits.

Semantic alone is defensible when you are orienting in a field you do not know, when the question crosses disciplines, or when you need three good papers rather than all of them. It is not defensible in anything claiming exhaustive coverage.