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
In natural language processing (NLP), words are typically represented as high-dimensional vectors (embeddings). For alphabetic languages like English, these are often learned based on context. However, Chinese is unique: its characters are logograms composed of graphical components (like radicals) that carry inherent semantic meaning.
While previous research attempted to use sub-character information by looking up radicals in a dictionary, this approach is limited. Dictionaries don't capture every nuance, and many non-radical components also provide semantic or phonetic hints. This research explores whether machines can learn Chinese semantics directly from the visual "glyphs" of the characters, capturing all graphical information without relying on human-annotated dictionaries.
Methodology
The researchers proposed a multi-stage pipeline to integrate visual information into word embeddings:
Character Bitmap Feature Extraction
The team rendered 8,780 unique Traditional Chinese characters into 60x60 grayscale bitmaps. They then trained a Convolutional Auto-encoder (convAE)—consisting of five convolutional layers—to compress these images into 512-dimensional "glyph feature" vectors.
Glyph-Enhanced Word Embedding (GWE)
They modified the standard Character-enhanced Word Embedding (CWE) model, introducing two variants:
- ctxG (Context Glyphs): Enhances the representation of context words using their visual features to help predict a target word.
- tarG (Target Glyphs): Uses the visual features of the target word itself to refine its own representation.
Additionally, they experimented with learning word representations solely from sequences of glyph features using Recurrent Neural Networks (GRUs), termed RNN-Skipgram and RNN-GloVe.
Experimental Setup
The models were trained on a massive corpus of Traditional Chinese news from the Central News Agency (1991–2002), totaling over 316 million segmented words. To evaluate performance, the authors translated standard datasets (WordSim-240, WordSim-296, and SimLex-999) into Traditional Chinese and introduced a new "job&place" analogy dataset.
Results and Analysis
The glyph-enhanced models showed significant improvements across various tasks:
- Word Similarity: On SimLex-999, the
CBOW + CWE + ctxGmodel achieved a Spearman’s correlation of 0.3538, outperforming the standard CBOW baseline (0.3330). This is likely because visual cues are strong indicators of "genuine" similarity (e.g., coast vs. shore). - Analogy Tasks: In the "Family" category, the
Skipgram + CWE + ctxGmodel reached an accuracy of 0.5662, a significant jump over the baseline Skipgram (0.4779). - Visual Clustering: t-SNE visualization revealed that characters with similar components naturally clustered together, confirming the convAE captured structural similarities without explicit linguistic supervision.
Conclusion and Limitations
While visual information is a powerful booster, it can be "noisy." For example, shared graphical components do not always imply semantic similarity (e.g., "mountain" and "honey" share a component but are unrelated). Furthermore, learning embeddings purely from glyphs performed poorly, suggesting that co-occurrence information in large corpora remains the essential foundation of word meaning. Future work will apply these visual features to downstream tasks like POS tagging and dependency parsing.
Summary updated