Abstract
Polysemy is a very common phenomenon in modern languages. Under many circumstances, there exists a primal meaning for the expression. We define the primal meaning of an expression to be a frequently used sense of that expression from which its other frequent senses can be deduced. Many of the new appearing meanings of the expressions are either originated from a primal meaning, or are merely literal references to the original expression, e.g., apple (fruit), Apple (Inc), and Apple (movie). When constructing a knowledge base from on-line encyclopedia data, it would be more efficient to be aware of the information about the importance of the senses. In this paper, we would like to explore a way to automatically recommend the primal meaning of an expression based on the textual descriptions of the multiple senses of an expression from on-line encyclopedia websites. We propose a hybrid model that captures both the pattern of the description and the relationship between different descriptions with both weakly supervised and unsupervised models. The experiment results show that our method yields a good result with a P@1 (precision) score of 83.3 per cent, and a MAP (mean average precision) of 90.5 per cent, surpassing the UMFS-WE baseline by a big margin (P@1 is 61.1 per cent and MAP is 76.3 per cent).
Introduction
Polysemy—the existence of multiple meanings for a single word—presents a significant challenge for natural language processing. While Word Sense Disambiguation (WSD) focuses on selecting the correct meaning within a specific context, this research addresses Primal Meaning Recommendation. A "primal meaning" is defined as the core, frequently used sense from which other senses are derived or referenced. For instance, the fruit "apple" serves as the primal meaning for "Apple Inc." or the movie "Apple."
Online encyclopedias like Baidu Baike often list senses chronologically or based on user edits, which may not reflect semantic importance. Automatically identifying the primal meaning is essential for improving knowledge base construction and enhancing user experience in digital information retrieval.
Methodology
The proposed system utilizes a hybrid model integrating three sub-models to capture various semantic and structural signals.
1. Skip-Thought Model (Unsupervised)
This component measures semantic similarity between the expression and its textual description. It operates on the hypothesis that the primal meaning's definition will be semantically closest to the word itself in a high-dimensional vector space.
2. Pattern Detection Model (Weakly Supervised)
Primal meanings typically employ a formal, dictionary-like linguistic style. To capture these stylistic nuances, the researchers implemented a Long Short-Term Memory (LSTM) network enhanced with an attention mechanism.
3. Relation Graph Model (Weakly Supervised)
This model represents different senses as nodes in a directed graph. It calculates "deduction scores" between senses, testing the hypothesis that secondary meanings are derived from the primal hub.
Weak Supervision Strategy
To overcome the lack of labeled data, the authors used a "noisy" heuristic: assuming the first-listed sense in Baidu Baike is the primal meaning. Although this assumption is only accurate in approximately 44% of cases, it provided a sufficient signal for the models to learn underlying patterns.
Experimental Setup
The study used data from Baidu Encyclopedia (Baidu Baike):
- Training Set: 118,956 expressions (1.4 million senses) with noisy labels.
- Evaluation Sets: 547 validation and 1,200 test expressions, manually annotated by experts to establish ground truth. The dataset focused on general linguistic expressions, excluding specific names of people and places.
Results and Discussion
The hybrid model significantly outperformed the UMFS-WE (Unsupervised Most Frequent Sense with Word Embeddings) baseline across key metrics:
- Precision at 1 (P@1): 83.3% (Hybrid) vs. 61.1% (Baseline).
- Mean Average Precision (MAP): 90.5% (Hybrid) vs. 76.3% (Baseline).
Ablation studies revealed that the Pattern Detection Model was the most influential individual component, suggesting that the linguistic "style" of a definition is a powerful indicator of its status as a primal meaning.
Conclusion
This research demonstrates that combining unsupervised semantic analysis with weakly supervised pattern and relationship modeling can effectively identify the core meanings of polysemous words. While the model faces challenges with highly overlapping contexts (e.g., multiple adaptations of the same story), its success in Chinese suggests potential for cross-linguistic application. Future work will focus on integrating these insights into downstream tasks like Entity Linking and automated Knowledge Base Construction.
Summary updated