Revisiting the Sound–Meaning Systematicity

Code and resources for:

Revisiting the Sound–Meaning Systematicity: A Fine-Grained Computational Analysis

Xiaoxi Luo, Abeer Kashar, Yuxuan Li, Xiaoning Wang, and Freda Shi.

This repository contains the lexicons, sampled usages, curation artifacts, preprocessing and experimental code, result files, and figure-generation code used in the paper.

Repository structure

data/
  lexicons/
    filtered/                  Filtered lexicons
    unfiltered/                Unfiltered lexicons
    french_stages/             Intermediate French curation artifacts
    hindi_stages/              Intermediate Hindi curation artifacts
  annotations/                 Lexicon filtering and curation records
  sampled_usages/              Sampled contextual usages
  derived_representations/     Historical derived artifacts where needed

src/
  common/                      Shared preprocessing utilities
  chinese/
    preprocessing/             Chinese preprocessing code
  english/
    preprocessing/             English preprocessing code
  french/
  hindi/                       Language-specific experiment code

results/
  filtered/                    Filtered-lexicon results
  permutation_baseline/        Shuffled permutation baselines
  unfiltered_downsampled/      Unfiltered downsampling results
  static_baseline/             English fastText baseline

figures/
  make_paper_figures.py        Figure 2 and Figure 3 generation
  paper_outputs/               Historical final figure files

environment/                   Historical environment information

Experimental setup

The reported correlations are computed over distinct unordered lexical-item pairs. Same-item pairs and reverse duplicates are excluded.

Contextual semantic representations are averaged within each lexical item before pairwise semantic distances are calculated. Semantic distance is cosine distance between lexical-item representations. Phonemic distance is normalized dynamic time warping over PanPhon feature vectors.

For Chinese, pronunciation is derived from each usage context to account for polyphonic characters before the phonemic representations are aggregated at the lexical-item level. For English, French, and Hindi, pronunciation is generally context-independent.

The main experimental settings are:

• sampled usages per retained lexical item: 100
• permutation baseline: 100 permutations
• random seed: 13
• unfiltered-lexicon downsampling: 50 trials
• locality confidence intervals: 2,000 bootstrap samples
• locality neighborhood sizes: 500, 1,000, 2,000, 5,000, 10,000, and 20,000 pairs

For the permutation baseline, the historical analysis uses the following semantic representation keys:

• English: bert mean_embedding
• French: camembert mean_embedding
• Hindi: hindibert mean_embedding
• Chinese: bert mean_embedding

All four languages use panphon segments as the phonetic representation key.

Models

The contextual encoders used in the paper are:

Language	Model
English	bert-base-uncased
French	CamemBERT
Hindi	google/muril-base-cased (MuRIL)
Chinese	bert-base-chinese

The English static-embedding baseline uses pretrained fastText embeddings.

Pretrained model weights and complete third-party corpora are not redistributed in this repository.

Lexicons and sampled usages

Filtered and unfiltered lexicons are stored under data/lexicons/.

Where available, intermediate filtering and annotation artifacts are retained under data/lexicons/*_stages/ and data/annotations/.

The contextual usages sampled for the experiments are stored under data/sampled_usages/ as gzip-compressed JSON files. Each retained lexical item has exactly 100 sampled usages.

To decompress them while retaining the compressed copies:

gzip -dk data/sampled_usages/*.json.gz

For English, the filtered lexicon contains 800 items, of which 616 had at least 100 eligible usages and therefore entered the final representation pipeline.

The historical English full-list sampling pipeline read the lexical items true and false as Boolean values, so they appear as True and False in the sampled-usage JSON.

The number of lexical items entering a particular correlation analysis can be smaller than the corresponding curated lexicon because items without enough eligible usages or without usable representations are excluded from that stage.

Preprocessing code

Shared preprocessing utilities are stored under src/common/.

English preprocessing code is stored under src/english/preprocessing/ and includes usage sampling, contextual embedding generation, and phonetic feature extraction.

Chinese preprocessing code is stored under src/chinese/preprocessing/ and includes usage sampling, contextual embedding generation, and context-sensitive pronunciation and phonetic feature extraction.

The final correlation, permutation, downsampling, and locality analyses are implemented by the language-specific scripts elsewhere under src/.

Hindi filename note

Some historical Hindi scripts and representation files contain the label hindibert. These names are retained where necessary to preserve the original analysis path.

The embedding-generation code and stored metadata identify the model used as google/muril-base-cased (MuRIL), which is therefore the Hindi encoder reported in the paper.

Intermediate Hindi lexicon files are also retained to document successive etymological and morphological filtering stages.

Results

The JSON files underlying the reported analyses are organized as follows:

results/filtered/
results/permutation_baseline/
results/unfiltered_downsampled/
results/static_baseline/

Only outputs corresponding to analyses reported in the paper are included as final result files.

Reproducing Figures 2 and 3

Run from the repository root:

python figures/make_paper_figures.py

Generated figures are written to:

figures/generated/

Figure 2 consists of four language-specific panels comparing the filtered lexicon, downsampled unfiltered lexicon, and shuffled baseline across neighborhood sizes.

Figure 3 compares global and local Spearman correlations at k = 1000.

The PDF and SVG files used during paper preparation are retained under figures/paper_outputs/.

Environments

The experiments were run on Alliance Canada / Compute Canada infrastructure. Historical package information is provided under environment/.

Different stages of the project used different environments, particularly for the Chinese phonetic and semantic preprocessing pipeline. The archived environment files are therefore provided as provenance rather than as a single portable installation specification.

Citation

If you use these resources, please cite:

Xiaoxi Luo, Abeer Kashar, Yuxuan Li, Xiaoning Wang, and Freda Shi.
“Revisiting the Sound–Meaning Systematicity: A Fine-Grained Computational Analysis.”
Computational Linguistics.

A complete BibTeX entry will be added once the final publication metadata is available.