Introduction
Entity Linking (EL) is the critical task of mapping ambiguous text mentions to unique entries in a knowledge base, such as Wikipedia. A primary challenge in EL is polysemy, where a single name (e.g., "Pink Floyd") can refer to multiple distinct entities depending on the context. Traditional EL systems have often relied on "collective linking" or coarse heuristics like tf-idf to match document text with entity descriptions. However, these methods frequently fail to capture subtle semantic nuances. This paper proposes a neural approach using Convolutional Neural Networks (CNNs) to distill text into meaningful topic vectors for more precise semantic matching.
Methodology: Multi-Granularity Semantic Matching
The proposed model addresses the limitations of coarse heuristics by comparing the semantic "topic" of a document at multiple levels of detail against potential target entities.
The architecture employs CNNs to generate dense vector representations for three granularities of the source text:
- Mention: The specific text string being linked (e.g., "Pink Floyd").
- Context: The immediate local neighborhood (sentence or surrounding words) of the mention.
- Document: The global context of the entire source text.
Similarly, candidate entities from Wikipedia are represented by their Entity Title and Entity Article text. The model computes the cosine similarity between all pairs of these source and target vectors (e.g., Document vs. Article). These neural similarity scores are then integrated into a log-linear model alongside traditional "sparse" features, such as prior link probabilities from Wikipedia. This hybrid approach allows the system to leverage both statistical popularity and deep semantic understanding.
Experimental Results
The model was evaluated on four diverse entity linking benchmarks: ACE, CoNLL-YAGO (noted for rare entities), WP (Wikipedia snippets), and a random sample of 10,000 Wikipedia articles.
As shown in the results table, the "Full" model—which combines sparse features with multi-granularity CNN features—achieved significant improvements over previous state-of-the-art systems. For instance, on the ACE dataset, accuracy reached 89.9%, a substantial increase from the 79.6% reported in prior work. Ablation studies confirmed that while CNN features alone are powerful, the combination of dense neural features and sparse statistical features provides the most robust performance across all datasets. Additionally, the researchers found that word embeddings trained on Wikipedia outperformed those trained on Google News, likely due to better domain alignment with the target entities.
Discussion and Insights
A key insight from this work is that CNNs effectively act as "topic distillers," filtering text to identify relevant n-grams that define a specific subject area. Visualization of CNN filters revealed that some learn to activate for specific domains, such as "weapons" or "legal proceedings."
Furthermore, the research highlights the importance of granularity. Neither broad document context nor narrow local context is sufficient in isolation. The global document provides a "genre" estimate, while the immediate context offers specific disambiguation clues. By utilizing all levels of granularity, the model captures a more complete semantic picture. Despite these successes, some CNN filters remain difficult to interpret, and the computational overhead of running multiple CNNs per candidate entity remains a consideration for large-scale applications.
Conclusion
This paper demonstrates that deep learning, specifically multi-granularity CNNs, can significantly enhance entity linking by capturing semantic similarity more effectively than traditional heuristics. By stacking neural features with sparse statistical indicators, the model achieves state-of-the-art performance across multiple benchmarks. Future work may address the "NIL" entity problem for mentions not present in the knowledge base and further improve the interpretability of neural filters.
Summary updated