We operationalize semantic structure using Binder et al.’s (2016) multidimensional feature norms. To scale these ratings beyond the original 535 English seed words, we fine-tune transformer-based language models (BERT large for English; HerBERT for Polish) with a custom regression head to predict 65-dimensional semantic vectors from contextualized sentence embeddings. The models achieve near-human performance (median R² ≈ 0.97) and are applied to ~40,000 words per language, generating high-resolution semantic profiles.

Phonological representations are derived from standardized IPA transcriptions and encoded as sparse binary matrices capturing phoneme inventories, articulatory features, phoneme bigrams, morphological variables, and part-of-speech categories. We employ XGBoost regression to identify the strongest phonological predictors for each semantic dimension, followed by distribution-appropriate Generalized Linear Models with Bonferroni correction and effect size estimation via Cliff’s delta. Interaction filtering controls for potential morphological confounds.

Results reveal consistent phonosemantic effects across both languages. For example, high front vowels show negative associations with visual salience, nasals correlate positively with duration and weight, and sibilants align with audition and unpleasantness. These findings replicate classic phonesthetic patterns while extending them into a high-dimensional experiential semantic space.

The study demonstrates that distributed iconicity permeates everyday vocabulary and shows how transformer-based semantic modeling combined with gradient boosting enables systematic large-scale investigation of sound–meaning mappings across languages.