Clarifying orthography: Orthographic transparency as compressibility

TL;DR

Orthographic transparency lacks a universal, script-agnostic metric. The authors introduce mutual compressibility, defined via $I_K(y; x) = K(x) - K(x|y)$ and $C_K(y; x) = I_K(y; x)/K(x)$, to quantify shared structure between orthography and phonology. They approximate these quantities with prequential coding using a neural sequence model, enabling a unified measure that captures both irregular spellings and rule complexity across 22 languages and diverse scripts. Findings show the measure aligns with intuitive script transparency (e.g., logographic being less transparent, syllabic like Katakana more so) and reveals directionality effects depending on transcription granularity. The approach offers a principled, general yardstick for evaluating orthographic transparency with potential educational and typological implications.

Abstract

Orthographic transparency -- how directly spelling is related to sound -- lacks a unified, script-agnostic metric. Using ideas from algorithmic information theory, we quantify orthographic transparency in terms of the mutual compressibility between orthographic and phonological strings. Our measure provides a principled way to combine two factors that decrease orthographic transparency, capturing both irregular spellings and rule complexity in one quantity. We estimate our transparency measure using prequential code-lengths derived from neural sequence models. Evaluating 22 languages across a broad range of script types (alphabetic, abjad, abugida, syllabic, logographic) confirms common intuitions about relative transparency of scripts. Mutual compressibility offers a simple, principled, and general yardstick for orthographic transparency.

Figures (6)

• Figure 1: The hypothetical orthographic transparency continuum imagines languages and their writing systems organized on a one-dimensional line.
• Figure 2: Phonological datasets are compressed both with Top: orthographic side-information and Bottom: no additional information. The difference between the lengths in the two compressed datasets yields the mutual algorithmic information between the dataset and the side information. We show this to be a good approximation of orthographic transparency.
• Figure 3: A depiction of prequential coding. Labels and inputs from previous samples at $t = 1, 2, ..., n-1$ are used to train the current model and yield parameterization $\theta_n$. The current model is used to make a prediction given input $x_n$. The resulting distribution is used to encode label $y_n$.
• Figure 5: Confidence intervals for mutual compressibility of orthography given phonology, and vice versa, in 22 languages. Error bars represent 95% confidence intervals as determined over all $40$ runs.
• Figure 6: Within languages the measure also can provide distinctions between script types. Here we compare katakana, kanji, and the full language including all three scripts. Hiragana was excluded because there were not enough stand-alone words made exclusively of hiragana characters to analyze.