Concise definitions of the key terms used across these docs and the code.
Aggregator / combiner — the function that turns several base-measure scores into one ensemble score (mean, regression, evolved weights, …).
Algebraic stacking — combining base scores with a fixed statistical operator (mean, median, weighted sum); unsupervised.
Base measure — an individual similarity function f(a, b) → [0, 1] (lexical,
knowledge-based, embedding, transformer).
Blending — supervised stacking with a regression meta-learner fit to gold scores.
Bootstrap confidence interval — an uncertainty range for a metric, obtained by resampling the data with replacement and recomputing the metric many times.
Calibration — agreement between predicted scores and the actual proportion of similarity they imply.
Cross-validation (k-fold) — splitting data into k folds, training on k−1 and testing on the held-out fold, so reported scores are on unseen pairs.
Diversity — the degree to which ensemble members make uncorrelated errors; the main driver of ensemble gains.
Distributional measure — similarity from co-occurrence-based vectors (word2vec, GloVe, fastText) via cosine.
Gold score — the human reference similarity for a pair, usually the mean of several annotators' ratings.
Grammatical evolution — an evolutionary technique that evolves programs/ expressions defined by a grammar; used to design ensembles automatically.
Information content (IC) — −log p(concept); rarer (more specific) concepts
carry more information. Used by Resnik/Lin/JCN WordNet measures.
Kendall τ — rank correlation based on concordant vs. discordant pairs (τ-b corrects for ties).
Knowledge-based measure — similarity from a structured resource (e.g., WordNet) via taxonomy distance.
Least common subsumer (LCS) — in a taxonomy, the deepest concept that is an ancestor of both items; central to Wu–Palmer and IC measures.
Lexical / string measure — similarity from surface form (n-gram overlap, edit distance); no external knowledge.
Meta-learner — the model that learns to combine base scores (regression, neural net, …).
MSE / RMSE — (root) mean squared error against gold; here computed after min-max scaling so measures are comparable.
Pearson r — linear correlation between predicted and gold scores.
Relatedness — items are associated but not interchangeable (car / road); distinct from similarity (car / automobile).
Spearman ρ — Pearson correlation of the ranks; the headline metric for similarity/STS because it is invariant to monotonic rescaling.
Stacking (stacked generalization) — training a model on the outputs of other models; here, on the outputs of base similarity measures.
STS (Semantic Textual Similarity) — the task of scoring sentence-pair similarity, evaluated mainly by Spearman/Pearson.
Steiger–Williams test — a significance test for comparing two dependent correlations that share a variable (used to compare two methods on the same data).
Tournament selection — a genetic-algorithm operator that picks the fittest of a small random subset as a parent.
WordNet — a large lexical database of English organized into synsets linked by semantic relations; the basis of knowledge-based measures.