acceptodds
Under review as a conference paper at ICLR 2027

DARE: Efficient Multilingual PII Detection

Abstract

Detecting personally identifiable information (PII) is a fundamental component of privacy-preserving natural language processing, requiring models that reliably identify sensitive information across languages and heterogeneous data distributions. However, current models either involve billions of parameters and achieve-state-of-the-art performance, or involve millions of parameters and fall short in performance. In this work, we introduce DARE, an efficient bidirectional encoder for multilingual PII detection with 300M parameters, trained on a heterogeneous mixture of publicly available datasets spanning multiple languages and annotation schemes. DARE achieves strong detection performance across well-established multilingual privacy benchmarks and outperforms state-of-the-art open-weight PII detection baselines with up to more total parameters under a common evaluation protocol, while incurring up to lower inference overhead.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
anonymization, pseudoanonymization, private identifiable information, privacy, pii detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RTRaOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DARE, an efficient bidirectional encoder for multilingual PII detection with 300M parameters. DARE is based on mmBERT-base and achieves strong performance on two multilingual privacy benchmarks.

Critical strengths
  1. The model features a lightweight architecture capable of rapidly responding to privacy spans, and it performs well on REDACT and AI4Privacy.

  2. The paper introduces marginal-likelihood training to learn from disjoint or overlapping label sets

Critical weaknesses
  1. The method does not feature specific optimizations for multilingual performance as the title might suggest; it is simply based on the multilingual mmBERT-base model and trained on a multilingual dataset.

  2. The paper lacks a discussion on the impact of PII detection on relevant downstream tasks. As the paper states that PII detection is used for text redaction, how do different masking methods affect downstream tasks—such as text comprehension by humans or models?

Questions for authors
  1. Please refer to the weaknesses detailed above.
  2. Could the authors evaluate additional language models in the main experiments, including both open- and closed-source options such as Qwen3.5-9B and GPT?
Reviewer_4d2POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DARE, a multilingual PII detection model based on mmBERT with approximately 300M parameters. It employs marginal likelihood training to handle heterogeneous annotations, language-balanced sampling and threshold-based decoding. Experiments on REDACT and AI4Privacy demonstrate strong detection performance and lower inference latency than several existing baselines.

Critical strengths
  1. This paper addresses inconsistent annotation granularities across multilingual datasets through marginal likelihood training. The additional evaluation of over-redaction is also relevant to real world privacy protection.
  2. DARE achieves the highest F2 scores among the evaluated baselines on both REDACT and AI4Privacy, while using a relatively compact architecture. Its significantly lower inference latency makes it promising for practical deployment.
Critical weaknesses
  1. The main components, including the mmBERT backbone, marginal likelihood training, are largely based on existing techniques. So the methodological novelty is somewhat limited, and the paper does not clearly demonstrate substantial technical advances beyond combining and adapting existing methods.
  2. The evaluation mainly relies on type-agnostic character-level F2, which ignores entity classification errors. Therefore, the reported results do not fully reflect the model's ability to identify different PII types accurately.
Questions for authors
  1. Could the authors clarify the main methodological novelty of DARE compared with existing multilingual PII detection methods, given that most of its components are based on established techniques?
  2. How are predictions from overlapping windows merged, and how are invalid BIO transitions and entities crossing window boundaries handled?
  3. How much of DARE's multilingual performance comes from the pretrained mmBERT backbone rather than the proposed training strategy?
Reviewer_PXKCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents DARE, a 300M-parameter bidirectional encoder (mmBERT-base backbone plus a token-classification head over 117 BIO classes covering 58 PII types) for multilingual PII detection. The model is trained on an undisclosed mixture of public PII corpora (53.5M tokens, 13 languages) using a per-token marginal-likelihood loss over admissible label sets to reconcile differing annotation granularities, with language-balanced sampling (alpha = 0.5). At inference, a thresholded decoder marks a token as sensitive when the pooled non-O probability exceeds tau = 0.175. DARE is compared against Presidio, GLiNER-PII, OpenAI Privacy Filter (1.5B), Qwen3.5-2B and Gemma4-E2B on REDACT and AI4Privacy using type-agnostic character-level F2, in a "zero-shot" setting and after small-sample fine-tuning. The paper also reports inference wall-clock time on an H100 and over-redaction on two internal AI-generated PII-free suites.

Critical strengths

-The problem (compact, fast, multilingual PII detection) is practically relevant, and a 300M encoder that beats a 1.5B model on two public benchmarks would be a useful artifact if the comparison holds.

-The evaluation includes bootstrap confidence intervals on paired differences, per-language recall breakdowns, a threshold ablation (Appendix B/C) and an uncalibrated comparison, which is more careful than usual for this kind of system paper.

-Weights and evaluation splits are released anonymously.

Critical weaknesses

-Unspecified training data and likely benchmark overlap. Section 3.2 describes the mixture only by token counts and 13 languages. It never lists the sources, their licences, their label inventories, or how the 58-type taxonomy was derived.

-Table 6 states that tau was chosen as "the largest recall with precision at least 65.62%". 65.62 is exactly the uncalibrated precision of OpenAI Privacy Filter in Table 7. The operating point was therefore chosen with reference to a competitor's result on the same benchmark distribution, and F2 rewards exactly this trade. The same tau is then used on AI4Privacy with no stated recalibration.

Questions for authors

Q1. List all training datasets with per-language, per-source token and annotation counts. Is any portion of AI4Privacy/OpenPII (any version) or REDACT in the training or validation mixture? Was deduplication against the two evaluation sets performed, and with what method?

Q2. What exactly is the "500-document selection split" used to fit tau (Table 6) and the validation set used for early stopping? Are they drawn from REDACT or AI4Privacy?

Q3. How were baseline thresholds for GLiNER-PII and OpenAI Privacy Filter chosen, on which data, and what values were used?

Q4. Which entity label list was passed to GLiNER-PII at inference, and what prompt (with output format) was given to Qwen3.5-2B and Gemma4-E2B?

Comments on the paper

7 comments

Dear Reviewers,

Thank you for your time and effort in reviewing this submission.

As the rebuttal discussion period is approaching its deadline, I would like to kindly encourage you to review the authors' responses and engage in the discussion. Please indicate whether your concerns have been adequately addressed or if any questions remain. If appropriate, please also update your ratings accordingly.

Your timely feedback would be greatly appreciated and will help ensure a fair and constructive evaluation.

Best regards,
Area Chair

Dear Reviewers,

We would like to thank you for your initial reviews of our submission. We wanted to kindly remind you that we are approaching the last 24 hours of the rebuttal and so far, no reviewer has participated in the rebuttal, acknowledged our replies and reflected the rebuttal in the score.

We would like to kindly point out that we invested significant effort in providing our answers as early as the 8th of November. On top of it, we have written the reviews ourselves and have not incorporated large language models. We would like to additionally remind you, that the Program Chairs have written the following on the 13th of November:

We’re about halfway through the discussion period. This is a friendly reminder to check all of your assigned papers. If the authors have answered your questions or responded to critiques, please leave a comment letting them know if your concerns have been addressed. Thank you! – the PCs

We would like to invite the reviewers to honor their agreement when they decided to join the conference and participate in the discussion as they would expect other reviewers that review their submissions to participate.

Regards, The Authors

Dear Reviewers RTRa and 4d2P,

You have made a commitment when you joined the conference as reviewers and until the moment of writing you have not participated in the discussion, acknowledged our replies and reflected our feedback into the review score. Something which goes against what is expected and to what the Program Chairs urged.

Following the closure of the rebuttal we will request an investigation into our submission and more concretely the reviewer’s behaviour. We will request for disciplinary measures to be taken against both reviewers.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.