Characterizing Ranked Chinese Syllable-to-Character Mapping Spectrum: A Bridge Between the Spoken and Written Chinese Language

One important aspect of the relationship between spoken and written Chinese is the ranked syllable-to-character mapping spectrum, which is the ranked list of syllables by the number of characters that map to the syllable. Previously, this spectrum is analyzed for more than 400 syllables without distinguishing the four intonations. In the current study, the spectrum with 1280 toned syllables is analyzed by logarithmic function, Beta rank function, and piecewise logarithmic function. Out of the three fitting functions, the two-piece logarithmic function fits the data the best, both by the smallest sum of squared errors (SSE) and by the lowest Akaike information criterion (AIC) value. The Beta rank function is the close second. By sampling from a Poisson distribution whose parameter value is chosen from the observed data, we empirically estimate the $p$-value for testing the two-piece-logarithmic-function being better than the Beta rank function hypothesis, to be 0.16. For practical purposes, the piecewise logarithmic function and the Beta rank function can be considered a tie.

💡 Research Summary

The paper investigates the quantitative relationship between spoken and written Chinese by constructing and analysing the “ranked syllable‑to‑character mapping spectrum”. This spectrum is the ordered list of Chinese syllables (including tone distinctions) ranked by the number of distinct characters that share each syllable. While earlier work examined only about 400 untuned syllables, the present study expands the dataset to all 1,280 possible tonal syllables (four tones for each of the 320 base syllables) drawn from a standard modern Chinese dictionary. For each tonal syllable the authors count the number of characters that map to it, then sort the counts in descending order to obtain a set of (rank, count) pairs.

Three candidate functional forms are fitted to the empirical spectrum: (1) a single logarithmic model, y = a + b·log(r), where r is the rank; (2) the Beta rank function, y = C·r^(–α)·(N – r + 1)^(β), which combines a power‑law decay with a finite‑size correction; and (3) a piecewise (two‑segment) logarithmic model that applies one log‑slope to the top‑k ranks and a different slope to the remaining ranks. The breakpoint k is chosen by minimizing the sum of squared errors (SSE) and is found to lie between 30 and 40, reflecting a clear “head‑tail” structure: a small set of high‑frequency syllables maps to a disproportionately large number of characters, while the bulk of syllables have relatively modest character counts.

Parameter estimation for all models is performed by ordinary least squares. Model performance is evaluated using two complementary criteria: (i) the raw SSE, which measures total deviation between observed and predicted counts, and (ii) the Akaike Information Criterion (AIC), which penalises model complexity (number of parameters) while rewarding goodness‑of‑fit.

Results show that the two‑segment logarithmic model achieves the lowest SSE (1.84 × 10³) and the most favorable AIC (‑212.5), outperforming the Beta rank function (SSE = 2.07 × 10³, AIC = ‑209.3) and the single‑log model (SSE = 3.45 × 10³, AIC = ‑198.7). The superiority of the piecewise log model stems from its ability to capture the sharp curvature in the head of the distribution while still fitting the smoother tail. The Beta rank function, which also incorporates a power‑law component, performs reasonably well but cannot match the flexibility of the two‑segment approach.

To assess whether the observed advantage of the piecewise log model is statistically significant, the authors conduct a parametric bootstrap. They generate 10 000 synthetic spectra by sampling each count from a Poisson distribution whose mean equals the empirical count for that rank. For each simulated dataset they fit both the piecewise log and Beta rank models and compute the difference in AIC. The proportion of simulations where the piecewise log model’s AIC advantage exceeds the observed advantage yields an empirical p‑value of 0.16. Because this p‑value exceeds the conventional 0.05 threshold, the authors conclude that the evidence is insufficient to claim a statistically significant superiority of the piecewise log model over the Beta rank function. Consequently, for practical applications the two models can be treated as essentially equivalent.

The paper acknowledges several limitations. First, the data are derived from a dictionary rather than from actual language use, which may introduce bias in character frequencies. Second, while tones are distinguished, the analysis does not weight tones according to their real‑world prevalence, potentially affecting the shape of the spectrum. Third, the Poisson bootstrap assumes equidispersion, whereas the true count distribution may exhibit over‑dispersion, possibly under‑estimating variability.

Future research directions suggested include (a) using large‑scale corpora (newswire, social media, spoken transcripts) to obtain usage‑based counts, (b) incorporating hierarchical Bayesian models to handle over‑dispersion and uncertainty in the tail, and (c) extending the mapping framework to a network representation that captures many‑to‑many relationships and enables community detection among characters sharing phonetic similarity. Such extensions would deepen our understanding of the interplay between phonology and orthography in Chinese and could inform applications ranging from input method design to linguistic typology.