MeSH Ranking Feature Signals: Fifteen Additional Separators

Additional features supply facts that frequency, hierarchy location, and one MiniLM score cannot express. Each feature describes one query-candidate relationship. The ranker learns when that fact changes the order.

Worked query: Similar candidates require different evidence

Query: Acute MI following PCI with a drug-eluting stent → Target: D009203 · Myocardial Infarction → Alternatives: Acute Coronary Syndrome; Myocardial Ischemia

Signal groups: Lexical, source, hierarchy, and candidate context

Group	Representative signals	Question answered
Lexical	Preferred term; synonym; acronym overlap	Does the query state the concept?
Source	Title; condition; intervention; source prior	Where does evidence occur?
Hierarchy	Depth; ancestors; branch density	Which specificity is correct?
Candidate context	Score margin; rank; hard-negative similarity	How ambiguous is the set?

Feature discipline: One value, one meaning, one shared producer

Compute the same feature during training and inference.
Keep deterministic features inspectable beside model scores.
Measure incremental value by feature group.
Remove duplicate signals without held-out value.

Feature register: Fifteen signals and their operational purpose

#	Signal	Purpose
1	Preferred-term overlap	Recognize literal official names.
2	Official synonym overlap	Recognize exact concepts expressed through entry terms.
3	Scope-note keyword match	Match distinctive definition words.
4	Acronym expansion	Resolve supported short forms without deleting originals.
5	Character similarity	Survive spelling, punctuation, and hyphen variants.
6	Query-field signals	Preserve title, condition, intervention, and description location.
7	Source-specific MiniLM	Keep PubMed and ClinicalTrials semantic opinions separate.
8	Hard-negative MiniLM	Separate correct terms from plausible wrong candidates.
9	Candidate prevalence	Measure identifier frequency within each source.
10	Query-conditioned hierarchy	Compare candidates with likely query branches.
11	Relative hierarchy	Summarize candidate-group branches without pairwise loops.
12	Cross-encoder score	Allow direct interaction between query and candidate words.
13	SapBERT score	Add a biomedical synonym model opinion.
14	MeSH-name ambiguity	Discount lexical matches in crowded name neighborhoods.
15	Original frequency rank	Preserve retriever order across variable score ranges.

Measured rollout: Cost increases across three rounds

Round	Features	Required work
1	1–6, 9, 14, 15	Deterministic calculations and stored counts
2	8, 10, 11	New training and hierarchy aggregation
3	7, 12, 13	Additional encoders and higher inference cost