Mandarin classifier systems optimize to accommodate communicative pressures

Yamei Wang, Géraldine WaltherEMNLP 2023 · DOI

Abstract

This paper investigates the structural logic of the Mandarin Chinese classifier system, demonstrating that it functions as a mathematically optimized system similar to grammatical gender in European languages. By analyzing approximately 1 million noun phrases using information-theoretic measures and Generalized Additive Models (GAMs), we show that the system balances the need for learnability (regularity) with communicative efficiency (discriminability). Specifically, the system prioritizes regularity for low-frequency nouns to aid acquisition while emphasizing discriminability for high-frequency nouns to reduce processing ambiguity. Our findings suggest that Mandarin classifiers are not merely arbitrary measure words but a sophisticated mechanism for managing information entropy.

Introduction

In linguistics, "competing communicative pressures" describe the evolutionary tug-of-war between making a language easy to learn and making it efficient to use. A regular system allows learners to guess the category of a rare word by analogy, while a discriminative system helps fluent speakers distinguish between common words quickly. While grammatical gender systems (e.g., masculine/feminine in Spanish) have been shown to optimize this balance, "genderless" languages like Mandarin Chinese have been less studied in this context. This research explores whether the Mandarin classifier system—which is semantically transparent and far more numerous than typical gender markers—performs a similar functional role.

Methods

The researchers employed a multi-faceted computational approach to analyze a massive dataset from the Leipzig Corpora Collection, focusing on ~760,000 distinct nominal phrases.

  1. Semantic Mapping: Cosine similarity between nouns was calculated using pre-trained fastText embeddings to quantify semantic relatedness (e.g., "bench" vs. "table").
  2. Class Membership Score (SCM): A numeric score was developed to represent the overlap in classifier usage between noun pairs.
  3. Contextual Interaction: Pointwise Mutual Information (PMI) was used to measure how often two nouns appear in the same context, signaling a higher need for discriminability.
  4. Statistical Modeling: Generalized Additive Models (GAMs) were utilized to capture the non-linear interactions between frequency, similarity, and classifier usage.
Figure 1: SCM scores across noun frequencies, showing the broad distribution of classifier sharing as a function of word commonality.

Key Insights and Results

The "Frequency Flip"

The most significant finding is the divergent treatment of common versus rare words. For low-frequency nouns, the system prioritizes regularity: semantically similar nouns are highly likely to share the same classifiers, making the system easier to learn by analogy. Conversely, for high-frequency nouns, the system prioritizes discriminability. Even when two common nouns are semantically similar, they are more likely to use distinct classifiers, helping listeners distinguish them rapidly during real-time processing.

Functional Equivalence to Gender

Mandarin classifiers effectively reduce the "search space" for the listener. Much like hearing "la" in Spanish prepares the listener for a feminine noun, hearing a specific Mandarin classifier (like zhāng for flat objects) immediately narrows the set of possible upcoming nouns. The study found that classifiers immediately preceding nouns share significantly more information (~7.90 bits) than those separated by other modifiers (~6.79 bits), confirming their role in entropy reduction.

Contextual Dampening

The interaction between semantic similarity and contextual co-occurrence (PMI) is crucial. As shown in the modeling, high co-occurrence dampens the likelihood of sharing a classifier. If two words often appear in the same context, the language evolves to give them different classifiers to avoid ambiguity.

Figure 2: Interaction between semantic similarity and Pointwise Mutual Information (PMI) on classifier sharing (SCM), illustrating how contextual co-occurrence reduces classifier overlap.

Conclusion and Future Work

This research confirms that Mandarin Chinese optimizes its classifier system to accommodate the same communicative pressures found in gendered languages. The system employs a "dual strategy": relying on grammatical classification when nouns appear alone, and shifting toward probabilistic classification when adjectives are present to help reduce uncertainty.

Future work should aim to distinguish between "sortal" classifiers (categorizers) and "mensural" classifiers (measures), as preliminary data suggests sortal classifiers are the primary drivers of entropy reduction. Additionally, testing these information-theoretic predictions against child language acquisition data would provide further evidence for the "learnability" hypothesis.

Summary updated

BibTeX
@inproceedings{wang2023mandarin,
  author = {Yamei Wang and Géraldine Walther},
  title = {{Mandarin classifier systems optimize to accommodate communicative pressures}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  doi = {10.18653/v1/2023.findings-emnlp.843},
}
Figure