Learning Chinese Word Representations From Glyphs Of Characters

Tzu-Ray Su, Hung-yi Lee

Abstract

In this paper, we propose new methods to learn Chinese word representations. Chinese characters are composed of graphical components, which carry rich semantics. It is common for a Chinese learner to comprehend the meaning of a word from these graphical components. As a result, we propose models that enhance word representations by character glyphs. The character glyph features are directly learned from the bitmaps of characters by convolutional auto-encoder(convAE), and the glyph features improve Chinese word representations which are already enhanced by character embeddings. Another contribution in this paper is that we created several evaluation datasets in traditional Chinese and made them public.

Introduction

Traditional Natural Language Processing (NLP) models often treat words as discrete IDs, ignoring the internal structure of logographic languages like Chinese. However, Chinese characters are built from graphical components—such as radicals—that provide significant semantic and phonetic cues. For instance, the "water" radical (氵) visually links characters to aquatic concepts. While previous attempts to leverage this information relied on human-curated radical dictionaries, this research explores whether machines can learn these cues directly from raw character pixels.

Methodology

The proposed approach integrates visual information into word representations through a two-stage process:

  1. Feature Extraction: Traditional Chinese characters are rendered into 60x60 grayscale bitmaps. A Convolutional Auto-Encoder (convAE) is then trained to compress these images into 512-dimensional "glyph features." This unsupervised method allows the model to identify visual patterns and shared components without manual labeling.
Comparison between character glyphs and the 60x60 rendered bitmaps used for training the convolutional auto-encoder.
  1. Integration: These features are incorporated into word embedding frameworks via two primary methods:
  2. Glyph-Enhanced Word Embedding (GWE): Standard models (CBOW and Skip-gram) are modified to include glyph features of context or target words in the hidden layer.
  3. Direct Learning: A Gated Recurrent Unit (GRU) attempts to predict word meanings solely from sequences of glyph features, testing the viability of a purely visual representation.

Evaluation Datasets

The models were trained on the Chinese Gigaword corpus, containing over 316 million segmented words in traditional Chinese. To evaluate performance, the authors developed and released several traditional Chinese datasets:

  • Word Similarity: Translated versions of WordSim-240, WordSim-296, and SimLex-999.
  • Word Analogy: Standard analogy sets (Capital, City, Family) and a new "job&place" (J&P) dataset designed to test associations between professions and locations.

Key Insights and Analysis

A primary finding is that pixel-based models capture semantic information that radical-based models might miss. Many characters contain informative components that are not their official radicals. For example, in the character 伐 (attack), the right-side component 戈 (weapon) provides crucial meaning that a simple radical lookup might overlook.

Semantic decomposition of characters into visual components like 'human', 'weapon', and 'speech'.

Visualizing the convAE features using t-SNE showed that characters with similar visual components naturally cluster together. This confirms the model's ability to recognize sub-components autonomously. Furthermore, glyph features were most effective for "compositional" words—where the meaning is a clear sum of the parts—while being less effective for transliterated names where characters are chosen for sound.

Results

The Glyph-Enhanced (GWE) models demonstrated clear improvements over standard baselines:

  • Similarity: On the SimLex-999 dataset, adding target glyph features to CBOW increased the Spearman correlation from 0.3330 to 0.3493.
  • Analogy: In the "Family" category, accuracy rose from 0.4779 (Skip-gram) to 0.5662 (Skip-gram + context glyphs).

However, the "Direct Learning" models (RNN-Skipgram/GloVe) performed poorly. This indicates that while visual features are a powerful supplement, they cannot yet replace character IDs, likely because the mapping from raw pixels to abstract semantics requires the structural "anchor" provided by discrete IDs.

Conclusion

This research demonstrates that visual glyph features can significantly enhance Chinese word embeddings. While end-to-end visual models face training challenges, the integration of pixel-level information provides a more nuanced understanding of logographic semantics. Future work will explore joint training of the auto-encoder and embedding models, as well as applying these features to downstream tasks like named entity recognition and handling out-of-vocabulary characters.

Summary updated

Abstract

In this paper, we propose new methods to learn Chinese word representations. Chinese characters are composed of graphical components, which carry rich semantics. It is common for a Chinese learner to comprehend the meaning of a word from these graphical components. As a result, we propose models that enhance word representations by character glyphs. The character glyph features are directly learned from the bitmaps of characters by convolutional auto-encoder(convAE), and the glyph features improve Chinese word representations which are already enhanced by character embeddings. Another contribution in this paper is that we created several evaluation datasets in traditional Chinese and made them public.
BibTeX
@inproceedings{su2017learning,
  author = {Tzu-Ray Su and Hung-yi Lee},
  title = {{Learning Chinese Word Representations From Glyphs Of Characters}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2017},
  url = {https://openreview.net/forum?id=H14CrzG_WS},
}
Figure