@inproceedings{otto-etal-2026-evaluating,
title = "Evaluating Encoder- and {LLM}-Based Approaches for Robust Indirect Personal Identifier Detection",
author = {Otto, Christoph and
Baroud, Ibrahim and
Aizawa, Akiko and
M{\"o}ller, Sebastian and
Roller, Roland and
Raithel, Lisa},
editor = {Siegert, Ingo and
Szawerna, Maria Irena and
Choukri, Khalid and
Dobnik, Simon and
Kamocki, Pawe{\l} and
Lindstr{\"o}m Tiedemann, Therese and
Lison, Pierre and
Mu{\~n}oz S{\'a}nchez, Ricardo and
Pil{\'a}n, Ildik{\'o} and
S{\"o}derg{\r{a}}rd, Lisa and
Talmoudi, Kossay and
Volodina, Elena and
Vu, Xuan-Son},
booktitle = "Proceedings of the Joint Workshop on Legal and Ethical Issues in Human Language Technologies and Computational Approaches to Language Data Pseudonymization, Anonymization, De-identification, and Data Privacy ({LEGAL}2026 and {CALD}-pseudo 2026) @ {LREC} 2026",
month = may,
year = "2026",
address = "Palma, Mallorca (Spain)",
publisher = "ELRA",
url = "https://aclanthology.org/2026.legal-1.11/",
doi = "10.63317/3noxgknxv96m",
pages = "91--101",
abstract = "Removing explicit protected health information does not fully eliminate re-identification risk in clinical text. Contextual attributes such as socio-economic status, institutional affiliations or detailed life circumstances may still enable linkage attacks. These heterogeneous and sparsely distributed elements, termed Indirect Personal Identifiers, extend de-identification beyond fixed identifier lists and pose new modeling challenges. Therefore, we present the first systematic comparison of encoder-only models, prompt-based LLMs and hybrid pipelines for span-level IPI detection in English discharge summaries. A fine-tuned RoBERTa-large model improves on an existing baseline and substantially outperforms ChatGPT-5.2, achieving 0.906 micro-F1 and 0.724 macro-F1, compared to 0.509 micro-F1 and 0.487 macro-F1. Our findings indicate that IPI detection constitutes a distinct modeling regime characterized by class imbalance and high intra-class variability, where scaling model capacity alone does not guarantee macro-level robustness. We show that supervised encoder models currently provide the most reliable foundation for extending anonymization guarantees and future research."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="otto-etal-2026-evaluating">
<titleInfo>
<title>Evaluating Encoder- and LLM-Based Approaches for Robust Indirect Personal Identifier Detection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Christoph</namePart>
<namePart type="family">Otto</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ibrahim</namePart>
<namePart type="family">Baroud</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Akiko</namePart>
<namePart type="family">Aizawa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sebastian</namePart>
<namePart type="family">Möller</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Roland</namePart>
<namePart type="family">Roller</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lisa</namePart>
<namePart type="family">Raithel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Joint Workshop on Legal and Ethical Issues in Human Language Technologies and Computational Approaches to Language Data Pseudonymization, Anonymization, De-identification, and Data Privacy (LEGAL2026 and CALD-pseudo 2026) @ LREC 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ingo</namePart>
<namePart type="family">Siegert</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maria</namePart>
<namePart type="given">Irena</namePart>
<namePart type="family">Szawerna</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Khalid</namePart>
<namePart type="family">Choukri</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Simon</namePart>
<namePart type="family">Dobnik</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Paweł</namePart>
<namePart type="family">Kamocki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Therese</namePart>
<namePart type="family">Lindström Tiedemann</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pierre</namePart>
<namePart type="family">Lison</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ricardo</namePart>
<namePart type="family">Muñoz Sánchez</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ildikó</namePart>
<namePart type="family">Pilán</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lisa</namePart>
<namePart type="family">Södergård</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kossay</namePart>
<namePart type="family">Talmoudi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Elena</namePart>
<namePart type="family">Volodina</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuan-Son</namePart>
<namePart type="family">Vu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA</publisher>
<place>
<placeTerm type="text">Palma, Mallorca (Spain)</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Removing explicit protected health information does not fully eliminate re-identification risk in clinical text. Contextual attributes such as socio-economic status, institutional affiliations or detailed life circumstances may still enable linkage attacks. These heterogeneous and sparsely distributed elements, termed Indirect Personal Identifiers, extend de-identification beyond fixed identifier lists and pose new modeling challenges. Therefore, we present the first systematic comparison of encoder-only models, prompt-based LLMs and hybrid pipelines for span-level IPI detection in English discharge summaries. A fine-tuned RoBERTa-large model improves on an existing baseline and substantially outperforms ChatGPT-5.2, achieving 0.906 micro-F1 and 0.724 macro-F1, compared to 0.509 micro-F1 and 0.487 macro-F1. Our findings indicate that IPI detection constitutes a distinct modeling regime characterized by class imbalance and high intra-class variability, where scaling model capacity alone does not guarantee macro-level robustness. We show that supervised encoder models currently provide the most reliable foundation for extending anonymization guarantees and future research.</abstract>
<identifier type="citekey">otto-etal-2026-evaluating</identifier>
<identifier type="doi">10.63317/3noxgknxv96m</identifier>
<location>
<url>https://aclanthology.org/2026.legal-1.11/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>91</start>
<end>101</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Evaluating Encoder- and LLM-Based Approaches for Robust Indirect Personal Identifier Detection
%A Otto, Christoph
%A Baroud, Ibrahim
%A Aizawa, Akiko
%A Möller, Sebastian
%A Roller, Roland
%A Raithel, Lisa
%Y Siegert, Ingo
%Y Szawerna, Maria Irena
%Y Choukri, Khalid
%Y Dobnik, Simon
%Y Kamocki, Paweł
%Y Lindström Tiedemann, Therese
%Y Lison, Pierre
%Y Muñoz Sánchez, Ricardo
%Y Pilán, Ildikó
%Y Södergård, Lisa
%Y Talmoudi, Kossay
%Y Volodina, Elena
%Y Vu, Xuan-Son
%S Proceedings of the Joint Workshop on Legal and Ethical Issues in Human Language Technologies and Computational Approaches to Language Data Pseudonymization, Anonymization, De-identification, and Data Privacy (LEGAL2026 and CALD-pseudo 2026) @ LREC 2026
%D 2026
%8 May
%I ELRA
%C Palma, Mallorca (Spain)
%F otto-etal-2026-evaluating
%X Removing explicit protected health information does not fully eliminate re-identification risk in clinical text. Contextual attributes such as socio-economic status, institutional affiliations or detailed life circumstances may still enable linkage attacks. These heterogeneous and sparsely distributed elements, termed Indirect Personal Identifiers, extend de-identification beyond fixed identifier lists and pose new modeling challenges. Therefore, we present the first systematic comparison of encoder-only models, prompt-based LLMs and hybrid pipelines for span-level IPI detection in English discharge summaries. A fine-tuned RoBERTa-large model improves on an existing baseline and substantially outperforms ChatGPT-5.2, achieving 0.906 micro-F1 and 0.724 macro-F1, compared to 0.509 micro-F1 and 0.487 macro-F1. Our findings indicate that IPI detection constitutes a distinct modeling regime characterized by class imbalance and high intra-class variability, where scaling model capacity alone does not guarantee macro-level robustness. We show that supervised encoder models currently provide the most reliable foundation for extending anonymization guarantees and future research.
%R 10.63317/3noxgknxv96m
%U https://aclanthology.org/2026.legal-1.11/
%U https://doi.org/10.63317/3noxgknxv96m
%P 91-101
Markdown (Informal)
[Evaluating Encoder- and LLM-Based Approaches for Robust Indirect Personal Identifier Detection](https://aclanthology.org/2026.legal-1.11/) (Otto et al., LEGAL-CALD-pseudo 2026)
ACL
- Christoph Otto, Ibrahim Baroud, Akiko Aizawa, Sebastian Möller, Roland Roller, and Lisa Raithel. 2026. Evaluating Encoder- and LLM-Based Approaches for Robust Indirect Personal Identifier Detection. In Proceedings of the Joint Workshop on Legal and Ethical Issues in Human Language Technologies and Computational Approaches to Language Data Pseudonymization, Anonymization, De-identification, and Data Privacy (LEGAL2026 and CALD-pseudo 2026) @ LREC 2026, pages 91–101, Palma, Mallorca (Spain). ELRA.