The comparative study of probabilistic knowledge structure between Japanese and Chinese

Yujie Zhang, Dong Yuan, Lan Tai, Asuka Terai, Masanori NakagawaInternational Journal of Human Culture Studies 2014 · DOI

Abstract

This study constructs probabilistic linguistic knowledge structures using large-scale Japanese and Chinese datasets to investigate the cultural and social systems underlying each language. Utilizing a rigorous probabilistic algorithm that addresses the limitations of Latent Semantic Analysis (LSA), we analyze latent semantic classes derived from noun-verb dependency relationships. Our findings reveal significant cross-cultural divergences in emotional regulation, economic outlook, and the conceptualization of power versus individual rights.

Introduction

While translation allows for the exchange of words, the underlying "mental maps" and cultural contexts often remain distinct. Traditional computational methods for semantic mapping, such as Latent Semantic Analysis (LSA), frequently suffer from data sparsity and noise because they rely on simple word proximity. This research addresses these challenges by employing a dependency-based probabilistic model to map the collective consciousness of Japanese and Chinese societies, moving beyond surface-level linguistic differences to uncover deep-seated structural variations in social systems.

Methodology

The study utilizes a probabilistic algorithm developed by Kameya & Sato (2005) to estimate latent semantic structures. This approach offers several advantages over standard LSA:

  1. Dependency Analysis: Instead of simple proximity, the model focuses on the co-occurrence of nouns and verbs (e.g., subject-verb or object-verb relationships), capturing more precise semantic links.
  2. Latent Class Modeling: The model assumes that word pairs are generated by hidden thematic "latent classes."
  3. Robust Estimation: The Expectation-Maximization (EM) algorithm is used for parameter estimation, incorporating a smoothing parameter ($\alpha$) to mitigate the "sparsity problem" where specific word combinations are absent from the training data.

The workflow involved morphological analysis and dependency parsing, followed by clustering these links into 100 latent classes for Chinese and 200 for Japanese.

Data Sources

To ensure a comprehensive "knowledge structure," the researchers utilized massive, diverse corpora:

  • Chinese Corpus (651.44 MB): Comprising the Chinese Treebank 4.0, People's Daily (1998), Beijing News (2010), and various electronic literary works.
  • Japanese Corpus (2.6 GB): An extensive dataset including 17 years of Mainichi Shimbun (1991–2008), major dictionaries, school textbooks, Aozora Bunko literary classics, and National Diet minutes.

Comparative Analysis of Knowledge Structures

By analyzing semantically equivalent latent classes across both languages, several key cultural insights emerged:

Emotional Regulation and Expression

In the "Emotion" cluster, Japanese nouns frequently related to internalized or suppressed stress (e.g., anxiety, insecurity), with associated verbs focusing on "wiping away" or "clearing" these feelings. Conversely, Chinese nouns were more direct (e.g., anger, desire), and associated verbs emphasized "expressing" or "releasing" emotions. This reflects the Japanese cultural ideal of kokki (self-control) compared to a more direct Chinese expressive style.

Economic Outlook

The "Economy" clusters reflected the historical context of the datasets. The Japanese data, covering the "Lost Decades," was dominated by downward-trend verbs such as "plunge" and "crash." In contrast, the Chinese data reflected a period of rapid expansion, characterized by verbs like "increase," "growth," and "surpassing targets."

Power vs. Individual Rights

Both languages utilize the character 権 (ken in Japanese, quán in Chinese). However, their structural associations differ sharply. In the Japanese structure, this character clustered with individual "rights" (human rights, privacy, basic rights). In the Chinese structure, it clustered with state "power" (military power, command, major power), highlighting different societal priorities regarding authority and the individual.

Cultural Hybridity

The "Ancient Culture" cluster in Chinese was strictly focused on domestic imperial history. The Japanese cluster, however, exhibited a "hybrid" nature, containing terms from Chinese history (Sui, Wei dynasties) alongside Western concepts (Celts, Puritans) and indigenous elements, illustrating Japan's historical tendency to absorb and integrate foreign cultural elements.

Conclusion and Future Directions

The study successfully demonstrated that probabilistic methods can extract meaningful cultural knowledge independent of specific grammatical rules. While the mathematical clustering is objective, the naming of these classes currently relies on human interpretation, suggesting a need for more automated matching techniques. Future research will aim to expand the Chinese corpus to balance the datasets and apply this framework to other languages, such as English, to develop a global map of probabilistic knowledge structures.

Summary updated

Abstract

本研究の目的は,日本語,中国語の大規模データに言語統計解析を用いて確率的言語知識構造を構築し,個々の言語の表面的な違いを超えて両言語の背景にある文化や社会システムを比較考察することである.本研究では,従来用いられている潜在意味分析(Latent Semantic Analysis)等の問題点を解消した,より厳密な確率的方法として,Kameya & Sato[6] のアルゴリズムを用いる.具体的には,同手法を用いて推定された両言語における多数の潜在クラスから,特に分析結果の解釈に基づき,中国語と日本語で同じネーミングになったクラスを選んで,比較考察した.本研究の方法を用いて,日本語,中国語に限らず,より多くの他言語,たとえば英語の大規模データベースを用いて,同様の確率的言語知識構造を構成し,多角的な比較研究を進めていくことも今後の重要な課題の一つである.
BibTeX
@article{zhang2014comparative,
  author = {Yujie Zhang and Dong Yuan and Lan Tai and Asuka Terai and Masanori Nakagawa},
  title = {{The comparative study of probabilistic knowledge structure between Japanese and Chinese}},
  journal = {International Journal of Human Culture Studies},
  year = {2014},
  doi = {10.9748/hcs.2014.222},
}
Figure