Abstract
The measurement of semantic similarity between concepts is an important research topic in natural language processing. In the past, several approaches for measuring the semantic similarity between concepts have been proposed based on WordNet or Wikipedia. However, improvements in the measurement accuracy of most methods have led to a dramatic increase in time complexity, and the existing methods do not effectively integrate WordNet and Wikipedia. In this paper, we focus on designing an efficient semantic similarity method based on WordNet and Wikipedia. To improve the accuracy of WordNet edge-based measures, we propose an edge weight model for combining edge and density information, which assigns a weight to each edge adaptively based on the number of direct hyponyms of the subsumer. Second, to improve the computational efficiencies of the existing Wikipedia link vector-based measures, we propose a new Wikipedia link feature-based semantic similarity method that converts Wikipedia links into semantic knowledge and replaces the TF-IDF statistical weight model in the existing measures. In addition, we propose two new word disambiguation strategies to further improve the accuracy of Wikipedia link-based measures. Finally, to fully exploit the advantages of WordNet and Wikipedia, we propose two new aggregation schemas for combining WordNet “is-a” semantics and Wikipedia link semantics to replace the current aggregation schemas that combine WordNet “is-a” semantics with category semantics in Wikipedia. The experimental results show that our aggregation models are outstanding in terms of accuracy, efficiency and word coverage compared to state-of-the-art similarity measures.
Introduction
Measuring semantic similarity—determining the degree to which two concepts like "car" and "truck" are related—is a foundational task in Natural Language Processing (NLP). It is essential for applications ranging from search engines to machine translation. Traditionally, researchers have relied on two primary types of resources: manually curated taxonomies like WordNet and massive, evolving knowledge bases like Wikipedia.
However, existing methods face significant hurdles. WordNet-based approaches often suffer from "density non-uniformity," where the semantic distance represented by a path in the hierarchy varies depending on how detailed that specific branch is. Conversely, Wikipedia-based methods, while rich in information, are often computationally expensive, relying on complex statistical models like TF-IDF and high-dimensional vector spaces. This paper presents a framework that integrates both resources to achieve human-level accuracy with high computational efficiency.
Methodology
The proposed approach introduces three primary innovations designed to optimize both accuracy and speed.
1. Adaptive WordNet Edge Weighting
Instead of treating every link in the WordNet hierarchy as having a uniform length, the authors propose an adaptive edge-weighting model. This model assigns weights based on local density: if a parent node has many direct "children" (hyponyms), the edge weight is calculated using the logarithm of that count. This allows the model to "compress" or "stretch" semantic distances to better align with human intuition as the taxonomy grows more complex.
2. Wikipedia Link Feature-Based Ratio
To bypass the heavy computational requirements of traditional Wikipedia vector models, the authors treat internal "outlinks" within an article as semantic features. By analyzing the overlap of these links between two concepts, the model captures "has-part" relationships that WordNet’s "is-a" structure might miss. This link-feature ratio is significantly faster than TF-IDF-based calculations.
3. Volunteer-Aware Disambiguation
To address word sense disambiguation (e.g., distinguishing between "bank" as a financial institution vs. a riverbank), the authors leverage a unique heuristic: Wikipedia editors tend to list the most common meanings at the top of disambiguation pages. Two strategies—the Proportional and Number models—were developed to select the most relevant senses based on this crowdsourced ordering.
Aggregation Schemas
Finally, the paper proposes Linear and Maximal aggregation schemas to combine WordNet and Wikipedia data. This dual-source approach allows the system to maintain high coverage, falling back on Wikipedia when a term (such as a proper noun like "Arafat") is missing from WordNet.
Experimental Evaluation
The framework was evaluated against several gold-standard benchmarks:
- Classic Noun Sets: MC30 (30 pairs) and RG65 (65 pairs).
- Modern/Large Sets: AG203 and SimLex666.
- Domain-Specific Sets: Pedersen30, which focuses on clinical terminology from SNOMED-CT.
Key Results
The experimental results highlight the efficiency and accuracy of the proposed models:
- Accuracy: On the MC30 and RG65 datasets, the "Maximal Aggregation" model achieved a Pearson correlation of 0.91 to 0.92. This is comparable to the average correlation between human subjects (~0.90).
- Efficiency: The Wikipedia link-feature model proved significantly faster than traditional vector-based methods like Explicit Semantic Analysis (ESA) due to its streamlined processing and avoidance of complex statistical weighting.
- Coverage: The integration of Wikipedia successfully handled word pairs absent from WordNet, demonstrating the robustness of the hybrid approach.
Conclusion and Future Work
This research demonstrates that combining the structured hierarchy of WordNet with the link-based semantics of Wikipedia provides a powerful tool for NLP. By addressing density non-uniformity and computational overhead, the proposed method offers a scalable solution for real-time applications.
Future work will focus on automating parameter tuning (such as the smoothing factor $\alpha$ in linear aggregation) and exploring machine learning techniques like Support Vector Machines (SVM) to dynamically select aggregation strategies. Additionally, the authors plan to integrate further structured resources like DBpedia to enhance the model's semantic evidence.
Summary updated