[BUPT Research] Bridging the Semantic Gap: Normalizing Social Tags for Hierarchical User Interest Discovery

TL;DR

Researchers from Beijing University of Posts and Telecommunications have developed a Tag Normalization Algorithm (TNA) that transforms chaotic, "long-tail" social media tags into a structured two-layer interest hierarchy. By using Wikipedia as a "world knowledge" anchor, the system can understand niche interests (e.g., mapping "Jonny Buckland" to "Music/Guitarist") even when traditional content-mining methods like LDA fail due to data sparsity.

The Problem: The Chaos of Human Expression

On platforms like Sina Micro-blog, users are defined by their tags. However, these tags are notoriously difficult for machines to process for three reasons:

1. The Long Tail: ~75% of tags appear only once, making statistical clustering nearly impossible.
2. Ambiguity & Specificity: Users use "non-mainstream" terms that lack explicit semantic links.
3. Lack of Hierarchy: There is no native structure to distinguish between a broad interest (e.g., "Sports") and a specific one (e.g., "NBA").

While previous works relied on user posts (tweets), many users are "lurkers" who consume but don't create content, leaving their profiles blank.

Methodology: Wikipedia as the Universal Translator

The core insight of this paper is that every niche tag can be "explained" if you have a large enough encyclopedia.

1. Building the Word Similarity Net

For every tag $t_{ij}$, the system fetches the top relevant Wikipedia pages to create a pseudo-document. Instead of simple frequency counts, it uses Augmented Expected Mutual Information (AEMI) to calculate the relationship between words. Unlike standard Mutual Information, AEMI subtracts "counter-evidence" to find more meaningful phrases.

2. The Voting Mechanism

The system treats words as vertices in a graph. A vertex's importance is determined by its similarity to other words and its relative frequency. Each vertex "votes" for its neighbors; the top-scoring words become the keywords for that tag.

3. Creating the Criterion Tag Set

The authors define a Popular Degree Function $g(r)=c\cdot r^{-\alpha }$, a power-law distribution that helps select a "Criterion Tag Set" $\Theta$—a standardized vocabulary of high-level interest categories.

Experimental Validation

The researchers tested the algorithm against standard baselines (TF-IDF, TextRank, LDA) using annotated news articles and real-world Sina Micro-blog data.

Performance Gains

In the keyword extraction task, the TNA approach reached 83.6% accuracy, surpassing TextRank's 80.9%. The physical intuition is that LDA and other probabilistic models struggle with the tiny "document" size of a single tag, whereas TNA leverages the dense information in Wikipedia.

Real-World Application: Social TV

The algorithm was integrated into a TV program recommendation system. By regularizing the user similarity matrix with normalized tags rather than raw ones, the system significantly reduced recommendation errors (Precision/Recall improvements).

In the figure above, specific tags like "DotA" are successfully mapped to "Game," and "NBA" to "Sports."

Critical Insight & Conclusion

This work demonstrates that external grounding is the antidote to the noise of social media. While the paper uses Wikipedia (a 2010s-era SOTA approach), the underlying logic—that short, noisy text must be expanded into a higher-dimensional conceptual space to be useful—remains a cornerstone of modern NLP.

Limitations: The manual classification of the top-level categories (21 categories) remains a bottleneck. Future work could automate this hierarchy construction using more dynamic ontologies or Large Language Models (LLMs) to handle emerging slang in real-time.