Bilibili Personal Attack Content Governance: Background, Goals, Methods, and Effectiveness

Governance Background and Objectives

Bilibili, as a comprehensive video community, faces challenges from personal attacks, insults, and defamatory comments that can damage the community atmosphere. The goal is to reduce exposure of such negative content, promote positive interaction, and create a friendly environment.

Current Situation of Personal Attack Content

Analysis of June 2022 reports shows that the main sources of negative comments are "trolling" and personal attacks. Short abusive terms are limited but their variants (homophones, initial‑letter matches, special characters/emojis) make detection difficult.

Sector‑Specific Issues

Eight sectors (life, entertainment, film, knowledge, technology, sports, games, music) were examined. Entertainment, film, and game sectors have the highest volume and concentration of personal‑attack reports, making them priority targets.

Specialized Governance Process

The process focuses on personal attacks (while also considering trolling). It includes:

Word‑matching identification using techniques such as pinyin recognition, numeric homophone detection, character similarity, and transformation mapping.

Model‑based recognition employing various NLP models (FastText, DPCNN, TextRCNN, Attention, BERT, tiny_bert) and graph‑based methods (GCN/GAT) to capture nuanced abusive language.

Examples of character‑similarity detection use the "phonetic‑shape code" concept to convert characters into machine‑readable numeric strings, enabling recognition of similar‑sounding or similarly‑shaped characters.

Key Strategies

The strategy is built on three dimensions:

Keyword dimension: A knowledge base linking entities to abusive keywords is maintained per sector, allowing precise matching of comments to risky entities.

User‑group dimension: Users are categorized by target (e.g., non‑real entities like celebrities or games vs. platform users/UPs). High‑risk user groups are identified through interaction patterns and relationships.

Content & UP dimension: High‑risk videos or UPs are flagged, and multi‑model fusion (linear regression, logistic regression, tree models) is used to improve recall of abusive content.

Governance Outcomes

After implementation, the proportion of personal‑attack reports in the entertainment, film, and game sectors dropped by 31.97% (from June to December 2023), and trolling reports decreased by 24.77%.

Summary and Outlook

While the reduction trend is evident, further improvements are needed, such as optimizing content ranking, automating negative‑keyword extraction, refining reporting workflows, and accelerating short‑cycle model training and deployment.