The main purpose of this study is to automatically categorize the Kadi records by subject. For this, previously it was conducted an experiment based on the Turkish BERT model. However, the model was not very successful in the subject categorization. The main problem here is due to the structure of the Ottoman language, although it is Turkish, it has some differences with modern Turkish in terms of both words and structure. For this reason, it was realized that in order to achieve the subject categorization goal, we first needed to train a model consisting of Ottoman Kadi records. In order to train a new Ottoman Turkish BERT (SerBERT) model, ten hundred volumes of kadi registers from Center for Islamic Studies was extracted. These volumes had been previously transcribed by various people. The project, which was initiated by İSAM in 2008, received support from the 2010 European Cultural Agency within the scope of the Istanbul 2010 European Capital of Culture events and was completed and published in forty volumes in 2012. The 60-volume collection, prepared in partnership with İBB Kültür A.Ş. and Medipol University Faculty of Law, was completed and published in 2019. These volumes were also added to the database and 100 volumes were made available for access (https://www.kadisicilleri.org).

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.

Synonymy appears to violate the principle of least effort, as there is no clear justification for maintaining multiple words with the same meaning. However, according to the principle of no synonymy, any difference in form should entail a difference in meaning.

This study applies information-theoretic measures to quantify contextual differences between near-synonyms. Jensen–Shannon divergence (JSD) is used to measure distributional overlap between synonym pairs, while information gain (IG) captures how specific each word's contextual distribution is. The analysis examines 134 synonym pairs in Hindi, comparing lexemes of Sanskrit origin with their Perso-Arabic counterparts across two corpora (768M and 350M tokens) and four window sizes.

Results show that synonym pairs share substantial contextual overlap (lower JSD than random pairs), yet are not fully interchangeable (JSD consistently above zero). Analysis of information gain revealed a striking asymmetry: in over 80% of pairs, the Perso-Arabic variant exhibited higher IG than its Sanskrit counterpart. This suggests that Perso-Arabic vocabulary tends to occur in more specific, contextually constrained environments, while Sanskrit variants serve as a more neutral, broadly distributed baseline.

These findings suggest that synonymy, rather than violating the principle of least effort, is a consequence of system optimization. Etymology appears to constrain lexical behavior, leaving lasting semantic footprints long after borrowing. This may point to the existence of loose semantic frames composed of etymologically related words forming coherent conceptual groupings. Further research across languages with layered lexical histories is needed to confirm these patterns.