Abstract
Morphological awareness in Chinese linguistics involves the ability to distinguish between different senses of a shared character within compound words. This study presents a computational ensemble model designed to cluster morphologically related words by integrating syntactic, semantic, and contextual features. Using Hierarchical Agglomerative Clustering (HAC), the model achieves an F-NMI score of 50.84%, performing at a level statistically similar to human children (55.52%) but significantly below adult benchmarks (76.80%).
Introduction
In the Chinese language, individual characters often function as "chameleons," shifting their meaning based on the compound word they inhabit. For instance, the character 商 (shāng) may denote a historical dynasty (商代) or commerce (商店). For humans, recognizing these sub-word structures—known as morphological awareness—is fundamental to literacy. For natural language processing (NLP), this ambiguity complicates tasks such as word sense disambiguation and text segmentation. This research investigates the extent to which computational models can replicate human-like clustering of these morphological families.
Methodology
The researchers approached the problem as a clustering task, utilizing four distinct linguistic features to represent target words:
- Latent Semantic Analysis (LSA): Captures semantic similarity based on document co-occurrence.
- Document Context: Measures word frequency across 90 diverse genres and topics.
- Relation (Dependency Parsing): Analyzes grammatical roles using the Stanford Parser.
- Part-of-Speech (POS): Tracks the distribution of a word's grammatical categories.
The Ensemble Approach
To maximize performance, an ensemble model was developed by weighting these individual features and aggregating their similarity matrices. The final clusters were generated using Hierarchical Agglomerative Clustering (HAC), which iteratively merges the most similar word pairs.
Experimental Setup
The study utilized the Academia Sinica Balanced Corpus (5 million words) as a reference. The test set comprised 11 "morphological families" containing 285 target words. Ground truth was established by three psycholinguistic experts. To provide human benchmarks, 14 adults and 9 children performed the same clustering task.
Results and Analysis
Performance was evaluated using the F-NMI metric, which balances clustering accuracy with normalized mutual information.
As shown in the results, the POS (Part-of-Speech) feature was the most effective individual predictor (47.45%), indicating that grammatical function is a strong indicator of character sense. The ensemble model (50.84%) outperformed all single-feature methods, demonstrating the value of multi-faceted linguistic analysis.
A key finding is the "Child-Level" performance of the model. While the ensemble model significantly outperformed the random baseline (25.44%), its performance (labeled "Aggregation" in Figure 1) was statistically comparable to the child group (55.52%). However, a substantial 26% gap remains between the model and the adult group (76.80%).
Conclusion
The study demonstrates that current NLP techniques can capture basic morphological awareness similar to that of a young reader. However, the performance gap compared to adults suggests that models lack the sophisticated "world knowledge" and deep semantic understanding that humans acquire over time. Future work will focus on incorporating knowledge-based approaches, such as WordNet integration, to bridge this gap.
Summary updated