Викторина, связанная с руководством · Средний Уровень
Викторина по предвзятости в алгоритмах модерации контента
Проверяет, как предвзятость ярлыков и вариации диалектов проходят модерацию, чем ложноположительные результаты отличаются от ложноотрицательных и почему важны проверки на уровне конвейера.
Вопрос 1 из 8
How does a moderation false positive occur?
Предварительный просмотр каждого вопроса8 вопросы
Ответы и пояснения появляются по мере прохождения теста.
- How does a moderation false positive occur?
- Abusive content is correctly removed
- The model processes a post slowly
- A user appeals a correct decision
- Benign content is incorrectly flagged or restricted
- What did Sap et al. find in their 2019 study of hate-speech datasets?
- Every AAE post contained hate speech
- All platforms removed AAE from their guidelines
- AAE markers correlated with toxicity labels, and models learned that association in tested data
- The study evaluated only image classifiers
- What was the highest relative labeling difference reported by Sap et al. for AAE tweets in their experiments?
- Ten times less likely
- Exactly the same in all datasets
- The study did not compare labels
- Up to two times more likely to be labeled offensive
- How did dialect priming affect the annotators in the 2019 study?
- Making annotators aware that text used AAE reduced the chance they labeled it offensive
- It removed the need for human labels
- It made every post a violation
- It increased all toxicity ratings
- Why can a keyword-only moderation rule misclassify reclaimed language?
- It may ignore speaker, context, quotation and intended meaning
- It always improves recall
- It uses too many training examples
- It has no threshold
- Which evaluation provides stronger evidence than one overall accuracy number?
- Количество используемых графических процессоров
- Task-specific false-positive and false-negative rates on representative examples
- The average length of a post
- The model vendor’s market share
- Which set of stages should an audit include?
- Only posts removed by moderators
- Only model training time
- Score, threshold, human review, enforcement and appeal
- Only the platform’s public policy title
- What should platforms avoid when auditing dialect-related moderation outcomes?
- Offering an appeal route
- Treating race or dialect labels as automatic proxies for harmfulness
- Checking annotator disagreement
- Testing meaning-matched examples