Topical Key Concept Extraction from Folksonomy

Han Xue, Bing Qin, Ting Liu, Chao XiangIJCNLP 2013

Introduction: The Challenge of Concept Extraction

Building ontologies—structured maps of concepts and their relations—is a cornerstone of the Semantic Web. Traditionally, these are extracted from formal text corpora. However, formal text often lacks coverage for niche or fast-changing topics and can lead to "topic confusion" when words have different meanings across contexts.

This paper proposes leveraging folksonomies (user-generated tagging systems) as a solution. Platforms like Douban or Flickr provide real-time, rich collective wisdom that reflects how people actually categorize information. The challenge lies in the fact that these tags are inherently noisy, ambiguous, and lack hierarchy.

A movie page on Douban.com illustrating the 'Resource' (movie) and the associated 'Tag set' provided by users.

Methodology: From Noisy Tags to Structured Concepts

The authors introduce a multi-stage pipeline to transform messy tag clouds into structured key concepts:

  1. Topic Identification: Using Latent Dirichlet Allocation (LDA), the system discovers underlying topics within a resource collection, assigning each tag a probability distribution across these topics.
  2. Graph Construction: For every identified topic, a "tag graph" is built where nodes represent tags and edges represent their relationships.
  3. Hybrid Link Weighting: This is the paper's primary contribution. Instead of relying solely on co-occurrence, the authors use a weight function that combines Local Similarity (how often tags appear together on the same resource) and Global Similarity (the cosine similarity of tags' distributions across the entire dataset). This dual approach helps the model ignore noise while strengthening semantically valid connections.
  4. Topic-Specific Ranking: A variant of PageRank is applied to rank tags within each topic, with the highest-ranking tags extracted as the "Key Concepts."

Evaluation and Results

The researchers evaluated their framework using a dataset crawled from Douban.com Movie, focusing on the top 250 movies and approximately 1,700 unique tags. They compared their approach against two established baselines: Topical PageRank (TPR) and context-sensitive Topical PageRank (cTPR).

Mean Reciprocal Rank (MRR) comparison showing the proposed method consistently outperforming baseline TPR and cTPR models across various parameters.

Key findings include:

  • Superior Precision: The proposed method achieved a precision of 0.700, representing a 7.5% improvement over cTPR.
  • Enhanced Ranking Quality: As illustrated in the Mean Reciprocal Rank (MRR) metrics, the framework consistently pushed the most relevant concepts to the top of the results compared to baselines.
  • Robustness: The method remained stable even when aggressively pruning weak links in the graph, a scenario where baseline methods typically fail.

Conclusion and Future Work

By integrating global semantic similarity with local co-occurrence, the authors successfully demonstrated a more robust way to extract concepts from noisy folksonomies. While the system is currently limited by tag sparsity (it cannot identify concepts users haven't tagged), future work will focus on integrating multiple folksonomy sources and automatically generating full hierarchical Chinese ontologies.

Summary updated

BibTeX
@inproceedings{xue2013topical,
  author = {Han Xue and Bing Qin and Ting Liu and Chao Xiang},
  title = {{Topical Key Concept Extraction from Folksonomy}},
  booktitle = {IJCNLP},
  year = {2013},
}
Figure