TL;DR
This paper proposes a novel method to quantify how "logographic" a writing system is by measuring the mutual information between graphic forms and both phonological and morphological information. By incorporating morphology, their approach offers a more complete picture than previous methods, showing promise in distinguishing between writing systems.
What Problem Does This Paper Address?
Writing systems are traditionally categorized as either "phonographic" (representing sounds) or "logographic" (representing words or morphemes). However, this binary classification is often too simplistic, as many systems fall somewhere in between. The challenge lies in quantifying how logographic or phonographic a system is, moving beyond a simple label to a degree of measurement. Previous work has focused on the relationship between written forms and sounds (orthographic transparency), but this often overlooks the crucial role of morphology. This paper aims to fill this gap by proposing a metric that considers the interplay between graphic forms, sounds, and morphological units.
Methods
The core of the proposed method is the use of mutual information (MI), a concept from information theory that measures the statistical dependence between two random variables. The authors define a new metric, $I$, as the difference between the mutual information of graphic forms and morphemes ($I(G; M)$) and the mutual information of graphic forms and phonological forms ($I(G; P)$):
$I = I(G; M) – I(G; P)$
The intuition here is that in a highly logographic system, graphic forms ($G$) will carry more information about morphological identity ($M$) than about phonological forms ($P$). Conversely, in a phonographic system, $G$ will be more informative about $P$ than about $M$. A key advantage of this approach is its symmetry, allowing analysis from either a reader's or writer's perspective without needing separate measures.
To apply this, the authors also developed "phonographized" versions of the writing systems. This involves rewriting texts such that they exclusively use phonographic elements (like Japanese kana) or, for Sumerian, rewriting morphemes with their most common spelling to create a minimally logographic system.
Datasets
The study uses data from two unrelated languages:
- Modern Japanese: Data is drawn from the Balanced Contemporary Corpus of Written Japanese (BCCWJ). This corpus provides tokens annotated with graphic form and phonological information.
- Ur III Sumerian: Data comes from the Open Richly Annotated Cuneiform Corpus (ORACC), specifically from administrative documents of the Ur III period. This corpus is rich in lexical and morphological annotations.
For both languages, complex tokens (compounds, inflected words) were processed into morpheme-length units. The Japanese data includes approximately 63 million morpheme tokens from the 5,000 most frequent lemmas. The Sumerian data comprises over 1.8 million morpheme-sized tokens.
Key Insights and Ideas
- Morphology is Key: The paper strongly advocates for the inclusion of morphological information in quantifying logography, arguing that it completes a "triangle" of relationships (graphic form, sound, meaning/morphology) that is essential for a comprehensive understanding.
- Symmetry of Mutual Information: The use of mutual information offers a unified measure that is agnostic to the direction of analysis (reader vs. writer), a significant improvement over previous methods that often required separate considerations.
- Phonographized Systems as Baselines: Creating hypothetical "phonographized" versions of writing systems provides a valuable baseline for comparison, allowing the authors to isolate the effect of logographic elements.
- Distinguishing Minimally Logographic Systems: The method can differentiate between phonographized systems that might be considered equally "minimally logographic" by other metrics, suggesting a finer-grained analytical capability.
Key Results
The study reports preliminary results for Japanese and Sumerian, along with their phonographized versions.
- Japanese: Shows a clear distinction with $I(G; M) > I(G; P)$, indicating it is highly logographic, as expected. The phonographized Japanese resulted in $I(G; M) < I(G; P)$, as expected for a phonographic system.
- Sumerian: Also shows $I(G; M) > I(G; P)$, confirming its logographic nature, though the magnitude of the difference is smaller than for Japanese.
- Phonographized Sumerian: The results for phonographized Sumerian (both full and "MonoChar" versions) were close to zero ($I \approx 0$), meaning graphic forms provided roughly equal information about morphemes and phonological forms. This is interpreted as a system that is minimally logographic, aligning with the construction of these versions. Notably, the phonographized MonoChar Sumerian was found to be more logographic than the phonographized Japanese, a distinction the authors highlight as an advantage of their method.
Limitations and Future Work
- Limited Scope: The study is limited to only two languages, which is a significant constraint for a typology study. The authors intend to expand their analysis to a wider range of writing systems.
- Morphological Parsing Imperfections: The morphological parsing, especially for Sumerian, may have limitations due to the complexity of the language and current understanding. The handling of Japanese "jukugo" words as bimorphemic might also not always perfectly reflect linguistic reality.
- Future Work: The authors plan to incorporate semantics into their model and expand the range of writing systems analyzed. They also aim to make their code and datasets publicly available.
Summary updated