Neural Article Pair Modeling for Wikipedia Sub-article Matching

Muhao Chen, Changping Meng, Gang Huang, Carlo Zaniolo

Abstract

Wikipedia editors frequently split long articles into sub-articles to improve readability, a process that leads to "knowledge fragmentation." This fragmentation violates the "article-as-concept" assumption critical for automated knowledge base (KB) construction and semantic search. This paper introduces a hierarchical neural framework designed to automatically link sub-articles back to their parent main articles. By integrating deep learning document encoders with explicit structural metadata, the proposed model achieves near-perfect matching accuracy on a large-scale benchmark dataset, WAP196k.

Introduction

The "article-as-concept" assumption is a cornerstone of modern AI systems that utilize Wikipedia as a source of structured knowledge. However, the manual curation of Wikipedia often involves splitting complex topics (e.g., Harry Potter) into multiple specialized pages (e.g., Fictional Universe of Harry Potter). While beneficial for human readers, this creates a fragmented landscape for machines, leading to incomplete facts and poor cross-lingual alignment.

Example of Wikipedia sub-articles for 'Bundeswehr', illustrating the fragmentation of a single concept into multiple pages.

The authors address this by automating the identification of sub-article relationships, effectively restoring the one-to-one mapping between articles and real-world concepts.

Methods

The proposed solution is a hierarchical neural architecture that evaluates pairs of articles $(A_i, A_j)$ to determine if one is a sub-article of the other.

The hierarchical neural architecture combines latent semantic embeddings from document encoders with explicit structural features for article pair classification.

The framework operates in two primary stages:

  1. Neural Document Encoders: The model employs encoders to extract latent semantic meanings from article titles ($t$) and summary paragraphs ($c$). The researchers evaluated three architectures: Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), and Attentive GRUs.
  2. Feature Fusion: Latent representations are augmented with "hand-crafted" explicit features, including symbolic similarity (token overlap), structural similarity (link-based metrics like the Milne-Witten Index), and term frequency of the main title within the sub-article text.

A final Multi-Layer Perceptron (MLP) processes these combined features to output a match probability.

The WAP196k Dataset

To train and evaluate the model, the authors developed WAP196k, a benchmark dataset containing approximately 196,000 Wikipedia article pairs. This dataset is significantly larger than previous efforts, which typically featured only a few thousand pairs. It includes ~17,000 positive matches identified via crowdsourcing and ~178,000 negative cases generated through strategies like pair inversion and sibling-article pairing.

Results and Discussion

The experimental results yielded several key insights:

  • CNN Superiority: CNN-based encoders outperformed sequential models (GRUs). This suggests that sub-article matching relies more on identifying specific "trigger" keywords and local semantic patterns than on long-range narrative flow.
  • Hybrid Modeling: The inclusion of explicit structural features provided a substantial performance boost, indicating that Wikipedia's metadata (links and hierarchy) is as informative as its textual content.
  • Title Significance: Ablation studies confirmed that article titles carry the strongest relational signal, more so than the body text.

The best-performing model variant, CNN+F, achieved a precision of 99.13% and a recall of 98.06%, vastly outperforming traditional baselines like Random Forests.

Conclusion and Future Work

The proposed neural framework demonstrates high scalability and accuracy, successfully identifying 200,000 new sub-article pairs across the entire English Wikipedia with 85.7% precision. Future research will focus on extending this model to cross-lingual contexts using bilingual word embeddings and refining the taxonomy to distinguish between abstract sub-concepts and specific refined entities.

Summary updated

BibTeX
@inproceedings{chen2018neural,
  author = {Muhao Chen and Changping Meng and Gang Huang and Carlo Zaniolo},
  title = {{Neural Article Pair Modeling for Wikipedia Sub-article Matching}},
  booktitle = {ECML/PKDD},
  year = {2018},
  url = {https://openreview.net/forum?id=ZVaDm-bKrL2},
}
Figure