REVIEW 4 cited by
A New Generation of Perspective API: Efficient Multilingual Character-level Transformers
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
On the world wide web, toxic content detectors are a crucial line of defense against potentially hateful and offensive messages. As such, building highly effective classifiers that enable a safer internet is an important research area. Moreover, the web is a highly multilingual, cross-cultural community that develops its own lingo over time. As such, it is crucial to develop models that are effective across a diverse range of languages, usages, and styles. In this paper, we present the fundamentals behind the next version of the Perspective API from Google Jigsaw. At the heart of the approach is a single multilingual token-free Charformer model that is applicable across a range of languages, domains, and tasks. We demonstrate that by forgoing static vocabularies, we gain flexibility across a variety of settings. We additionally outline the techniques employed to make such a byte-level model efficient and feasible for productionization. Through extensive experiments on multilingual toxic comment classification benchmarks derived from real API traffic and evaluation on an array of code-switching, covert toxicity, emoji-based hate, human-readable obfuscation, distribution shift, and bias evaluation settings, we show that our proposed approach outperforms strong baselines. Finally, we present our findings from deploying this system in production.
Forward citations
Cited by 4 Pith papers
-
AEGIS: Awareness-Enhanced Guidance for Iterative Safeguard
Marking offensive spans changes — but does not consistently improve — the toxicity–meaning trade-off in multilingual detoxification; the effect depends on the generator backbone and the language.
-
Conditional Reliability of Toxicity Signals for Multilingual and Code-Mixed Abuse Detection
A gated fusion head that conditions English toxicity, Indic abuse, and rule-based severity scores on the text context improves code-mixed abuse detection in 10/12 in-domain and 7/8 transfer comparisons.
-
Catching Stray Balls: Football, fandom, and the impact on digital discourse
Football match outcomes shift sentiment in club subreddits and, within minutes, in unrelated subreddits where the same users post.
-
Towards Inclusive Toxic Content Moderation: Addressing Vulnerabilities to Adversarial Attacks in Toxicity Classifiers Tackling LLM-generated Content
Zeroing attack-vulnerable attention heads improves BERT/RoBERTa toxicity classifier accuracy on PGD-adversarial inputs, with distinct heads implicated per demographic group.
Discussion (0). Continue with ORCID to comment.