When Conset Meets Synset: A Preliminary Survey of an Ontological Lexical Resource Based on Chinese Characters

Shu-Kai Hsieh, Chu-Ren HuangACL 2006

Abstract

Traditional Chinese Natural Language Processing (NLP) often struggles with the "wordhood" problem and the underutilization of the ideographic nature of Chinese characters (Hanzi). This paper introduces HanziNet, an ontological resource that shifts the focus from words to characters. By defining the "Conset" (Concept Set)—a character-level counterpart to WordNet’s Synset—the authors provide a framework for capturing the conceptual information inherent in the Chinese writing system.

Introduction: The Challenge of Chinese Wordhood

In Western linguistics, writing is frequently viewed as a secondary representation of speech. However, applying this "surrogational model" to Chinese presents significant hurdles. First, defining a "word" in Chinese is notoriously difficult, as native speakers often disagree on boundaries. Second, standard word-based resources like WordNet overlook the rich, evolutionary information encoded within individual ideographic characters. HanziNet addresses these issues by treating characters as primary conceptual units rather than mere building blocks for words.

The HanziNet Framework

HanziNet is built on the concept of the Conset, which groups characters sharing a primitive meaning derived from traditional philology. Unlike a Synset, which focuses on word-level synonymy, a Conset captures the underlying conceptual root shared by characters. The model recognizes characters as tripartite entities:

  1. Logographic: Representing specific syllables or morphemes.
  2. Phonological: Functioning as phonetic components within larger characters.
  3. Ideographic: Interpretable as direct instances of concepts.

The resource employs a tree-structured taxonomy where characters are instances of Consets, linked via hypernym/hyponym (IS-A) relations. This structure allows for a dual perspective: the "Atomic Globule," representing the rooted tree of concepts, and the "Cobweb," representing the complex network of character combinations in words.

The hierarchical structure of HanziNet, mapping ontological categories to specific Consets.

Network Modeling and Small-World Properties

A key insight of the research is the "small-world" nature of the Chinese character network. Statistical analysis reveals that while the network is sparsely connected, it possesses very short average paths and high local clustering. This suggests that the conceptual organization of Hanzi is highly efficient for information retrieval and cognitive processing.

The authors analyzed the network using four metrics: the number of nodes (N), the average degree (k), the clustering coefficient (C), and the average shortest path length (L). As shown in the table below, the actual configuration of the lexicon (N=6493) exhibits a clustering coefficient of 0.64, which is significantly higher than the 0.06 found in a random configuration. This high clustering, combined with a short path length of 2.0, confirms that the conceptual space of Chinese characters is structured for optimal connectivity.

Comparison of network statistics between the actual character configuration and a random model.

Key Results and Performance

The authors utilized the Academia Sinica Balanced Corpus of Modern Chinese to map over 5,000 characters and 30,000 derived words. At the time of this survey, 3,600 characters had been integrated into 309 distinct Consets.

HanziNet demonstrated strong interoperability by successfully aligning its Consets with the 164 Base Concepts of the Global Wordnet Grid (derived from EuroWordNet and SUMO). Preliminary applications in sense prediction for unknown two-character words showed high performance, validating the utility of character-grounded ontologies in practical NLP tasks.

Conclusion and Future Directions

HanziNet represents a significant step toward a multi-level knowledge resource that integrates radicals (Hantology), characters (HanziNet), and words (Chinese Wordnet). While the current survey is preliminary, future work aims to complete the character integration, refine the distinction between semantic heads and modifiers, and apply the resource to full-scale Word Sense Disambiguation (WSD) tasks. The ultimate goal is to provide a cognitively grounded framework that captures the deep conceptual information "wired" into the Chinese writing system over millennia.

Summary updated

BibTeX
@inproceedings{hsieh2006when,
  author = {Shu-Kai Hsieh and Chu-Ren Huang},
  title = {{When Conset Meets Synset: A Preliminary Survey of an Ontological Lexical Resource Based on Chinese Characters}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2006},
}
Figure