Introduction
Wikipedia serves as a foundational resource for Natural Language Processing (NLP), yet it is constrained by "link sparseness." Manual editorial guidelines typically suggest linking a concept only once per page, resulting in a dataset where only approximately 19% of nouns are hyperlinked. This sparsity creates a bottleneck for training machine reading models. While existing datasets like SemCor provide human-annotated gold standards, they are too small for large-scale training. Conversely, massive automated datasets like Google’s Wikilinks focus predominantly on Named Entities, neglecting general concepts such as verbs and adjectives. To address this, researchers have developed Semantically Enriched Wikipedia (SEW), an automatically generated corpus that triples Wikipedia's semantic links to over 200 million annotations.
Methodology: A Heuristic-Based Pipeline
The construction of SEW relies on a fully automatic, heuristic-based pipeline that leverages Wikipedia’s internal structure and the BabelNet semantic network. The approach is built upon two primary pillars: the "one sense per page" assumption—the principle that a word's meaning remains consistent within a single article—and a cascade of propagation heuristics.
As detailed in Figure 1, the pipeline utilizes several levels of propagation:
- Intra-page Propagation (SP, LP, PP): If a term is hyperlinked once, the system identifies all other occurrences (including lemmatized forms) within the same article and applies the same link.
- Inter-page Propagation (WIL, BIL): The system analyzes "ingoing" links. If multiple pages link a specific term to a target article, that link is propagated to unlinked mentions of the term across the entire corpus.
- Category Propagation (CP): By using entropy-based ranking, the system identifies specific, informative Wikipedia categories to disambiguate mentions while filtering out noisy, broad categories.
- Monosemous Content Words (MP): Verbs, adverbs, and adjectives with only one possible meaning in BabelNet are automatically annotated.
Results and Scale
SEW represents a significant leap in the scale of available sense-annotated data. It contains over 206 million annotations covering 4 million distinct concepts, effectively doubling the original noun coverage of Wikipedia to nearly 40%.
As shown in Figure 6, SEW significantly expands the number of annotations and senses compared to the original Wikipedia dump and other specialized corpora like Wikilinks or MASC. Beyond scale, the dataset maintains high quality; evaluation against a hand-labeled gold standard revealed a precision of 93.4%.
Furthermore, SEW demonstrates substantial utility in downstream tasks. When a standard Entity Linking system (IMS) was trained on SEW rather than standard Wikipedia links, it achieved consistent performance gains across major benchmarks, such as improving from 0.758 to 0.882 on the SemEval-2015 task.
Conclusion and Future Work
By moving beyond nouns to include verbs, adjectives, and adverbs, SEW provides a comprehensive training resource for semantic tasks. While the "one sense per page" rule proved robust (holding true for ~98% of pages), future iterations may explore paragraph-level coherence to handle exceptionally long or diverse articles. Additionally, the authors plan to utilize BabelNet’s multilingual capabilities to expand SEW beyond English, creating a global resource for cross-lingual NLP research.
Summary updated