A New Generation of Perspective API: Efficient Multilingual Character-level Transformers

Alyssa Lees; Vinh Q. Tran; Yi Tay; Jeffrey Sorensen; Jai Gupta; Donald Metzler; Lucy Vasserman

Source-linked AI summary

A New Generation of Perspective API: Efficient Multilingual Character-level Transformers

Alyssa Lees, Vinh Q. Tran, Yi Tay, Jeffrey Sorensen, Jai Gupta, Donald Metzler, Lucy Vasserman

arXiv:2202.11176v1Feb 22, 2022cs.CLcs.AIcs.CYcs.LG

TL;DR

Toxicity moderation needs models that work across multilingual, cross-cultural, and evolving web content, but specialized and vocabulary-bound systems have limited flexibility. The paper presents UTC, a compact multilingual token-free Charformer Transformer, and reports strong results across diverse robustness benchmarks and production deployment.

Problem
Existing moderation systems are often specialized by language or domain and constrained by static vocabularies, limiting applicability across multilingual, cross-cultural, and evolving web content.
Method
UTC is a single compact pretrained Charformer-based Transformer using vocabulary-free byte-level modeling, multilingual and comment-text pretraining, and production-efficiency techniques.
Results
UTC outperforms or matches strong baselines across multilingual toxicity, code-switching, covert-toxicity, emoji, obfuscation, distribution-shift, and bias evaluations.
Takeaways & Limitations
UTC supports a single language-agnostic moderation model that is robust to domain transfer and feasible for production serving.
Takeaways & Limitations
Biases in data collection, annotation, and society can manifest as unintended discriminatory biases in moderation models.

Abstract

from arXiv · show

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.

1 INTRODUCTION

UTC addresses the limits of specialized, vocabulary-bound moderation models with a single multilingual, token-free Charformer-based framework. The paper reports broad robustness, production efficiency, and competitive performance across languages, domains, and evaluation settings.

Monolingual and vocabulary-bound systems struggle with multilingual, cross-cultural, typographically diverse content, including code-switching and cross-lingual generalization.
UTC is a single compact pretrained Charformer-based Transformer for multilingual toxic content classification.It uses multilingual documents and online comment text with sequence-to-sequence denoising pretraining.
Vocabulary-free modeling makes UTC language agnostic and more robust to domain transfer while remaining competitive with domain- or language-specific subword models.
UTC is compact enough for production serving and is more than 10x more memory-efficient than the winning 2020 multilingual toxicity contest entries.
UTC combines Learnable Tokenizers, reconfigurable Seq2Seq Transformer architectures, and comment-based pretraining for efficient character-level multilingual moderation.
UTC outperforms or matches strong baselines across code-switching, covert toxicity, emoji, obfuscation, distribution-shift, and bias evaluations.

2 RELATED WORK

Prior moderation research faces methodological disagreement, multilingual gaps, and concerns that biased data can produce discriminatory model behavior. Related work motivates multilingual evaluation, bias analysis, and production-feasible standalone models.

Moderation research lacks consensus on sampling, annotation, annotator training, classifier design, and scoring metrics.
Training-data collection flaws, annotator biases, and societal biases can manifest as unintended discriminatory biases in moderation models.
The paper evaluates cross-language bias and uses data augmentation to address concerns about unintended bias.
English dominates offensive-content classification research, while multilingual models offer the practical advantage of not requiring language identification.

3 UTC: UNIFIED TOXIC CONTENT CLASSIFICATION

UTC converts UTF-8 byte sequences into learned latent subwords before applying a reconfigurable Transformer and task-specific prediction heads. Its pretraining and tokenizer components are designed to retain flexibility while supporting efficient multilingual modeling.

3 UTC: UNIFIED TOXIC CONTENT CLASSIFICATION: UTC receives UTF-8 bytes, embeds them, and uses Charformer encoders to learn subwords automatically.
3.1 Learnable Tokenizer: GBST dynamically downsamples byte embeddings into latent subword embeddings through differentiable, position-wise block scoring.
3.1.1 Learning Latent Subwords Automatically: GBST forms candidate blocks with strided pooling, scores them, upsamples their representations, normalizes scores across block sizes, and combines them by weighted summation.
3.1.1 Learning Latent Subwords Automatically: The tokenizer enumerates block sizes 1 through 4 and mean-pools the resulting locally composed sequence.
3.2 Transformer: The Transformer consumes latent subwords, while its standard self-attention and feed-forward stack supports the remaining computation.
3.3 Seq2Seq Pretraining: Seq2Seq denoising pretraining uses an encoder-decoder architecture, while task-specific uses may remove the decoder and retain a universal encoder.
3.3.2 Multi-Regression Heads and Loss Function: The model predicts toxicity probabilities or k-way toxicity subtypes through regression heads applied to pooled encoder outputs.
3.4 Pretraining: Pretraining mixes the Perspective Pretraining Corpus with mC4 and applies span-based denoising with 20-byte mean spans and 15% corruption.

4 EXPERIMENTAL SETTINGS

The experiments cover multilingual toxicity classification, robustness, and adaptation to new toxicity types, using public and production-derived data. Models include Perspective API, Custom mBERT, mT5, and the proposed UTC variants.

Experimental scope: The evaluation spans core multilingual classification, robustness, and adaptation to new types of toxicity.Core classification uses public and live-traffic-derived datasets.
Datasets: The setup includes the Multilingual Toxic Comments Challenge and a labeled Production-Multilingual dataset derived from live API traffic.The production-derived dataset is human-labeled for toxicity.
Models: Compared models include Perspective API, Custom mBERT, mT5, and the proposed UTC and UTC† variants.Custom mBERT uses a 200K SentencePiece vocabulary and comment-domain pretraining.
Models: UTC† uses approximately 268M parameters with d_model=768, d_ff=3072, d_kv=64, 12 heads, and 28 encoder layers.These settings distinguish UTC† from the vanilla UTC configuration described in the model setup.

5 EXPERIMENTS: MULTILINGUAL

The multilingual experiments evaluate the models on production-derived and public toxicity benchmarks against established baselines and highly engineered competition systems. UTC performs strongly overall, including on non-English languages, while UTC† approaches the competition’s top result as a single deployable model.

Production-Multilingual: The Production-Multilingual dataset is derived from live Perspective API traffic and is labeled offline by human annotators for toxicity.It covers multiple languages, including Arabic, Chinese, English, French, Hindi, Japanese, Korean, Portuguese, Russian, and Spanish.
Production-Multilingual: UTC outperformed all baselines on the Production-Multilingual dataset, with a smaller model surpassing mT5_base despite having more than twice fewer parameters.UTC’s main advantage appeared across many non-English languages, although Custom mBERT was stronger on English.
JMTCC: The JMTCC benchmark contains English plus Spanish, French, Italian, Portuguese, Russian, and Turkish, although Turkish is omitted from Perspective API results.The competition provided training data only in English, encouraging participants to use translation data.
JMTCC: The comparison includes competition-winning systems as highly engineered upper bounds, despite their infeasibility for the paper’s standalone production goal.The champion entry is estimated to exceed 5B parameters because it ensembles multiple large multilingual and monolingual models.
JMTCC: 0.9367 AUC-ROC is achieved by UTC† on JMTCC, outperforming all considered single-model baselines, especially mT5.This result is only slightly below the top Kaggle result, which used XLM-RoBERTa ensembles, pseudo-labeling, and other techniques.

6 EXPERIMENTS: ROBUSTNESS

The robustness experiments evaluate UTC under code-switching, human-readable obfuscation, domain and labeling shifts, and subpopulation shift. UTC generally outperforms strong baselines, with especially strong results under code-switching and obfuscation.

Code-Switching: UTC outperforms the best baseline on both JMTCC-CS and Production-CS code-switching evaluations.The authors attribute this advantage partly to UTC’s byte-level vocabulary, while noting that these sets are dominated by English and Latin-based languages.
Human-Readable Obfuscation: 86.0 AUC-ROC is recovered by UTC after fine-tuning on 30% obfuscated Civil Comments, versus 84.5 AUC-ROC for mT5-small.Without additional obfuscated training, UTC outperforms every baseline at every nonzero obfuscation rate, although all models degrade as obfuscation increases.
Distribution and Task Shift: UTC and multilingual mT5 perform on par with an English-only RoBERTa model additionally pretrained on Twitter data and fine-tuned on TweetEval.The evaluation is zero-shot on a differently labeled, Twitter-focused hate-classification task.
Distribution and Task Shift: UTC and mT5-small significantly outperform prior DistilBERT results on all CivilComments-WILDS metrics.Their gaps between average and worst-group performance are approximately half to one-third those of DistilBERT models using robust optimization techniques.
Bias Evaluation: UTC shows strong cross-language bias metrics above .7 and performs more strongly on more subgroup-language combinations than mT5-small.The analysis aggregates results with corresponding English identity terms across language splits.

7 EXPERIMENTS: ADAPTING

The adaptation experiments test whether UTC checkpoints transfer to covert toxicity and emoji-based hate after further fine-tuning. UTC achieves competitive or superior results against specialized monolingual and task-specific baselines.

Covert Toxicity: UTC and UTC† outperform Toxic-BERT and Covert-BERT on the CovertToxicity task.UTC† is competitive with mT5-base, and multilingual mT5 and UTC outperform the specialized monolingual English T5 model.
Emoji-Based Hate: Fine-tuned UTC significantly outperforms the best reported English-only DeBERTa model on HatemojiCheck.The authors posit that the learned tokenizer adapts during fine-tuning to parse emojis; mT5-small underperforms the comparison baseline.

8 DEPLOYMENT RESULTS

UTC was deployed in Perspective API to support ten new languages, meeting production standards where earlier candidates were too inaccurate, biased, or slow. The deployment operated without reported issues, with measured throughput and latency suitable for real-time serving.

Public Launch: 10 new languages were launched in Perspective API powered by UTC† on December 9, 2021.The launch covered Arabic, Simplified Chinese, Czech, Dutch, Indonesian, Japanese, Korean, Polish, Hindi, and Hinglish.
Public Launch: UTC† was the first model to reach production standards for these ten languages.Previous CNN and BERT-based candidates had low overall performance, low bias-evaluation performance, or excessive serving time.
Serving Performance: The deployed model averaged ∼15 QPS and ∼200ms median latency for the ten newly launched languages.Load testing found that the smaller 102M-parameter UTC model achieved 45ms median latency at 1K QPS on one TPUv2 chip with batching.
Efficiency: Charformer downsampling and decoder removal reduced sequence length and model depth by half, supporting efficiency in production.The paper attributes much of the model’s performance to these architectural modifications and evaluates their speed impact in an ablation study.

9 CONCLUSION

The paper presents a token-free Charformer approach for toxic comment classification and productionizes it for multilingual deployment. Its effectiveness is supported by rigorous evaluations on real-world and academic benchmarks.

UTC applies a state-of-the-art token-free Charformer model to toxic comment classification.
The authors describe efficiency techniques that enable serving the byte-level model in production.
UTC is deployed in production for 10 new languages in the Perspective API.

A REPRODUCTION DETAILS

The implementation follows a T5.1.1-style Transformer architecture with relative attention biases and uses established distributed-training software components. Charformer is integrated through its official implementation.

UTC is implemented in Mesh TensorFlow with the T5 library for distributed model parallelism.
The Charformer component uses the official implementation released by its authors.
The overarching architecture follows T5.1.1 and uses T5-styled relative attention biases instead of position embeddings.

A.0.1 Implementation.

The appendix documents optimization settings, production language coverage, and supplementary dataset and evaluation figures. These materials include training hyperparameters, API routing details, and captions for code-switching, obfuscation, and bias evaluations.

Implementation: Finetuning uses a fixed learning rate of 10^-3 and dropout of 0.1, while pretraining uses an inverse-square-root learning-rate schedule.Pretraining uses Adafactor and is conducted with 64 TPU-v3 devices.
Implementation: A single UTC model serves Arabic, Chinese (Simplified), Czech, Dutch, Hindi, Hinglish, Indonesian, Japanese, Korean, Polish, and Russian through the Perspective API.English access is released under the TOXICITY_EXPERIMENTAL attribute.
Evaluation materials: The supplementary materials include language breakdowns for production and JMTCC code-switching evaluation examples.The tables note that examples can contain multiple languages, so column totals do not sum to the overall total.
Evaluation materials: The figures document character substitutions, increasing character-obfuscation rates, and per-language unintended-bias metrics for mT5 and UTC.The obfuscation examples range from 0 to 50% character obfuscation rate.