💻 computer science

Content Deviation Recognition and Recommendation Model for Sports and Health Information: Based on Improved Text Rank Algorithm and Deep Fusion Transformer Architecture

This study proposes an improved Content Deviation Recognition and Recommendation Model for sports and health information that integrates an enhanced TextRank algorithm with SBERT, a source credibility mechanism, and a multi-level BERT-Longformer Transformer architecture to effectively identify misleading content and prioritize high-credibility recommendations, achieving superior performance in bias detection and ranking metrics compared to existing baseline models.

Original authors: Xiaoxue Hu, Guoyu Wang, Fuxin Hong, SHINYEOL CHANG, Kitak Kim

Published 2026-10-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaoxue Hu, Guoyu Wang, Fuxin Hong, SHINYEOL CHANG, Kitak Kim

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the digital age, health advice travels faster than ever. A single post about a miracle diet or a quick-fix supplement can spread across social media and video platforms, reaching millions of users before experts have a chance to verify it. This speed creates a dangerous problem: algorithms designed to show us what we like often prioritize sensational claims over factual accuracy. When a user searches for ways to lose weight or boost immunity, the system might serve up unverified articles that promise impossible results, simply because those articles are written in an engaging way. The core challenge for scientists and engineers is to build a system that does not just match a user's interest, but also understands whether the information is trustworthy. This requires a new kind of intelligence, one that can read a long article, sense the subtle difference between professional advice and pseudoscience, and weigh the reputation of the person or organization that published it.

A team of researchers has developed a new model to solve this exact problem, creating a system that acts as both a filter and a guide for sports and health information. Their work focuses on two main tasks: identifying misleading content and recommending only the credible material to users. To do this, they built a computer program that learns from a massive collection of real-world health articles, some of which are known to be false. The system first breaks down the text to find the most important ideas, not just by counting how often words appear, but by understanding the context in which those words are used. It then evaluates the source of the information, assigning a trust score based on whether the article comes from a recognized medical institution or a less reliable website. Finally, it combines these insights to decide which articles should be shown to a user and which should be held back for further review.

The researchers tested their model using a large dataset of health misinformation that has been used by experts for years. They found that their new approach was significantly better at spotting false information than previous methods. When asked to identify biased or misleading content, the model achieved a high accuracy score of 92.7 percent, outperforming other advanced systems that rely on standard language processing techniques. This improvement is crucial because misleading health claims often hide in plain sight, using scientific-sounding language to disguise unproven ideas. By understanding the deeper meaning of sentences rather than just looking for keywords, the model could detect these subtle deceptions more effectively.

Beyond just spotting lies, the system also learned how to rank information so that the most reliable articles appear at the top of a user's feed. The researchers introduced a method to calculate a "source confidence score," which acts like a reputation badge for the publisher. If an article comes from a highly trusted medical organization, it receives a high score. If it comes from a source with a history of spreading false claims, the score drops. When the researchers added this reputation factor to their ranking system, the quality of the recommended articles improved noticeably. The system became better at pushing high-quality, relevant content to the front while pushing down articles that were both misleading and came from untrustworthy sources.

The study also looked at how the model performed on specific topics, such as exercise routines, nutritional supplements, and vaccination safety. The system proved particularly effective at handling queries about exercise methods, where it correctly identified biased content in nearly 94 percent of cases. For questions about supplements, the system was able to rank the most credible answers significantly higher than before. In one specific test, the system successfully flagged a group of high-risk articles that claimed things like "losing five pounds in one week with three pills" or "vaccines causing DNA mutations." These articles were automatically moved to a restricted list for human review, preventing them from being widely distributed to the public.

What makes this work stand out is how it combines different techniques to handle the complexity of long health articles. Traditional methods often struggle with texts that are very long or contain information spread across many paragraphs. The new model uses a dual-layer approach: one part focuses on the immediate details of a sentence, while another part looks at the connections between distant parts of the text. This allows the system to understand the full picture of an argument, even if the misleading claim is buried deep within the article. The researchers also found that the system worked well even when it had limited data to learn from, suggesting it could be adapted for new types of health topics as they emerge.

The results show that it is possible to build a recommendation system that values truth as much as popularity. By integrating a measure of source reliability directly into the way articles are ranked, the model reduces the chance that users will encounter dangerous health advice. The study did not claim to have solved the problem of misinformation entirely, noting that the system still relies on existing databases of trusted sources and does not yet account for individual user preferences. However, the findings provide a strong foundation for the next generation of health information tools. The model demonstrated that by teaching computers to weigh the credibility of a source alongside the relevance of the content, we can create a digital environment where accurate health information is easier to find and harder to ignore.

Technical Summary: Content Deviation Recognition and Recommendation Model for Sports and Health Information

Problem Statement
The rapid proliferation of sports and health information on digital platforms has led to the widespread circulation of misleading and unsubstantiated content, such as pseudoscientific claims regarding "rapid weight-loss methods" or "side-effect-free supplements." Current recommendation systems primarily optimize for user interest relevance, often lacking mechanisms to account for content authenticity or source reliability. This gap allows biased health information to be amplified, potentially influencing individual exercise routines, nutritional choices, and health beliefs. Existing detection methods struggle with implicit bias signals in long-form texts and fail to systematically integrate source authority into ranking logic, while bias detection and recommendation modeling are often treated as separate tasks.

Methodology
The study proposes a unified modeling framework that integrates content bias detection with credible recommendation, utilizing the TREC Health Misinformation Track dataset. The architecture consists of four core components:

Improved TextRank Keyword Extraction:
To overcome the semantic limitations of traditional TextRank, the authors enhanced the algorithm by incorporating Sentence Bidirectional Encoder Representations from Transformers (SBERT). Sentences are encoded into context-aware semantic vectors. A graph structure is constructed where nodes represent candidate keywords and edge weights are defined by the cosine similarity of the sentence vectors containing those words. This semantic-aware graph allows for better capture of implicit bias information compared to frequency-based methods.
Source Credibility Modeling Mechanism:
A Source Credibility Score (SCS) is calculated to dynamically regulate document dissemination. The score is a weighted combination of three factors:
- Domain Authority ( $D$ ): Derived from trusted domain databases.
- Platform Credibility ( $P$ ): Mapped based on verified media lists.
- Historical Quality ( $H$ ): Calculated based on the historical bias rates of documents published by the source.
  The final score is compressed using a hyperbolic tangent function ($tanh$) and integrated into the ranking layer. Documents meeting specific criteria (high predicted bias probability and low source credibility) are assigned to a Restricted Dissemination Candidate Set (RES) for further verification.
Multi-Level Transformer Architecture:
To handle both local semantic details and long-range dependencies in health texts, the model fuses BERT and Longformer:
- Local Encoder (BERT): Extracts features from short text windows.
- Global Encoder (Longformer): Captures cross-segment dependencies using sliding window and global attention mechanisms.
- Fusion: A gated mechanism combines outputs from both encoders.
- Multi-Task Learning: The fused representation feeds into a dual-output structure performing Bias Classification (binary classification via cross-entropy loss) and Recommendation Ranking (pairwise ranking loss). The total loss function balances these two objectives.
Joint Recommendation Strategy:
The final recommendation score combines the content relevance prediction with the source credibility score, allowing the system to prioritize high-credibility content while suppressing low-credibility, high-bias material.

Key Contributions

Semantic-Enhanced Keyword Extraction: The integration of SBERT vectors into the TextRank framework improves the detection of implicit bias signals that traditional keyword extraction methods miss.
Credibility-Aware Ranking: The introduction of a dynamic Source Credibility Score allows the model to adjust recommendation priorities based on the publisher's historical reliability and domain authority.
Unified Multi-Task Architecture: The proposed HCR-Net (Health Credibility-aware Recommendation Network) jointly optimizes bias detection and relevance ranking using a BERT-Longformer fusion, addressing the limitations of treating these tasks separately.
Restricted Dissemination Mechanism: A specific protocol (RES) is established to flag and isolate high-risk content (high bias + low credibility) for manual review or restricted exposure.

Experimental Results
The model was evaluated on the TREC Health Misinformation Track dataset and compared against baselines including RoBERTa-base, XLNet-large, ERNIE-Health, DeBERTa, and BioLinkBERT.

Bias Detection: The proposed model achieved an F1-score of 92.7%, outperforming RoBERTa-base (89.1%), XLNet-large (87.6%), and ERNIE-Health (90.3%). It also achieved an AUC of 94.8%.
Recommendation Performance: In the recommendation task, the model showed relative improvements of 8.3% in nDCG@10 and 6.5% in Mean Average Precision (MAP) compared to baseline models.
Impact of Source Credibility: Adding the SCS feature improved the identification of high-bias, low-weight documents by 3.4% in F1-score and increased nDCG@10 by 4.7%.
Topic-Specific Performance: The model achieved an F1-score of 94.2% for "exercise methods" queries and a 9.1% increase in nDCG@10 for "supplements and nutritional interventions."
Restricted Dissemination: The RES mechanism successfully filtered 938 documents, with 91.3% manually confirmed as misleading, demonstrating effective proactive identification of high-risk content.

Significance and Claims
The paper claims that the proposed framework effectively addresses the dual challenges of detecting implicit bias in health information and mitigating its spread through intelligent recommendation. By combining semantic understanding with source credibility modeling, the system enhances the accuracy of biased information detection and regulates content exposure without sacrificing relevance. The study highlights that balancing content semantic scores with source credibility (optimal at $\lambda=0.5$ ) is critical for ranking quality.

Limitations and Future Directions
The authors acknowledge that the source credibility modeling relies on external knowledge bases, which may limit effectiveness for unstructured content or emerging platforms. Additionally, the current recommendation module does not incorporate user profiles or personalized health preferences. Future work is suggested to integrate knowledge graphs for better interpretability and to incorporate user interaction behaviors for enhanced personalization.

Technical Summary: Content Deviation Recognition and Recommendation Model for Sports and Health Information

More like this