User:Biasbot AS
From Wikipedia, the free encyclopedia
| This user is a bot | |
|---|---|
| (talk · contribs) | |
 Biasbot in action, flagging biased wording in Wikipedia articles. | |
| Status | Semi-active |
| Operator | asoundd |
| Flagged? | No |
| Task(s) | Identifying and suggesting neutral alternatives to stigmatizing language |
| Automatic or manual? | Semi-automatic |
| Programming language | Python |
| Exclusion compliant? | No |
Introduction
Biasbot AS is a bot that assists with enforcement of Wikipedia's neutral point of view policy by classifying stigmatizing language. The bot flags potentially biased sentences, generates neutral alternatives, and presents them to human editors for review.
Detection Algorithm
Model
Biasbot AS is built on BERTBASE fine-tuned for binary classification (stigmatizing vs. neutral). The model uses a single dense layer with dropout (p=0.1) on top of BERT's [CLS] token representation. Training: 8 epochs, AdamW optimizer (lr=2e-5), early stopping on validation loss.
Process
Biasbot AS scans articles periodically. It scores each sentence (0.0-1.0 scale) and flags sentences above 0.65 threshold. If possible, it generates a neutral alternative and presents to editors for review.
Threshold Selection
| Threshold | Precision | Recall | FP Rate |
|---|---|---|---|
| 0.50 | 62% | 68% | ~4.2% |
| 0.65 | 76% | 51% | ~1.8% |
| 0.80 | 85% | 32% | ~0.6% |
Generating Suggestions
Template-based:
Common patterns use rule-based substitutions:
"suffers from X" → "has X" / "lives with X"
"is bipolar" → "has bipolar disorder"
"committed suicide" → "died by suicide"
ML-based:
Complex cases use fine-tuned T5-small trained on 847 paired examples from Wikipedia edits and mental health style guides.
Performance
At the 0.65 confidence threshold:
- Precision: 76%
- Recall: 51%
- F1 Score: 0.61
- False positive rate: ~1.8% on validation set
Dataset: 2,847 sentences from 183 articles, annotated with κ = 0.68 inter-annotator agreement. The threshold prioritizes precision to minimize editor workload, similar to ClueBot NG's approach.
False Positives
Approximately 1.8% of neutral sentences are incorrectly flagged. This is not a judgment on your edit; review the suggestion and ignore if inappropriate. To report false positives, navigate to Asoundd's talk page.
Emergency Measure
Emergency bot shutoff button
Administrators: Use this button if the bot is malfunctioning. (direct link)
Non-administrators can a malfunctioning bot to Wikipedia:Administrators' noticeboard/Incidents.