@inproceedings{hoory-etal-2021-learning,
title = "Learning and Evaluating a Differentially Private Pre-trained Language Model",
author = "Hoory, Shlomo and
Feder, Amir and
Tendler, Avichai and
Cohen, Alon and
Erell, Sofia and
Laish, Itay and
Nakhost, Hootan and
Stemmer, Uri and
Benjamini, Ayelet and
Hassidim, Avinatan and
Matias, Yossi",
editor = "Feyisetan, Oluwaseyi and
Ghanavati, Sepideh and
Malmasi, Shervin and
Thaine, Patricia",
booktitle = "Proceedings of the Third Workshop on Privacy in Natural Language Processing",
month = jun,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.privatenlp-1.3",
doi = "10.18653/v1/2021.privatenlp-1.3",
pages = "21--29",
abstract = "Contextual language models have led to significantly better results on a plethora of language understanding tasks, especially when pre-trained on the same data as the downstream task. While this additional pre-training usually improves performance, it can lead to information leakage and therefore risks the privacy of individuals mentioned in the training data. One method to guarantee the privacy of such individuals is to train a differentially-private model, but this usually comes at the expense of model performance. Moreover, it is hard to tell given a privacy parameter $\epsilon$ what was the effect on the trained representation. In this work we aim to guide future practitioners and researchers on how to improve privacy while maintaining good model performance. We demonstrate how to train a differentially-private pre-trained language model (i.e., BERT) with a privacy guarantee of $\epsilon=1$ and with only a small degradation in performance. We experiment on a dataset of clinical notes with a model trained on a target entity extraction task, and compare it to a similar model trained without differential privacy. Finally, we present experiments showing how to interpret the differentially-private representation and understand the information lost and maintained in this process.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hoory-etal-2021-learning">
<titleInfo>
<title>Learning and Evaluating a Differentially Private Pre-trained Language Model</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shlomo</namePart>
<namePart type="family">Hoory</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Amir</namePart>
<namePart type="family">Feder</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Avichai</namePart>
<namePart type="family">Tendler</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alon</namePart>
<namePart type="family">Cohen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sofia</namePart>
<namePart type="family">Erell</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Itay</namePart>
<namePart type="family">Laish</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hootan</namePart>
<namePart type="family">Nakhost</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Uri</namePart>
<namePart type="family">Stemmer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayelet</namePart>
<namePart type="family">Benjamini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Avinatan</namePart>
<namePart type="family">Hassidim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yossi</namePart>
<namePart type="family">Matias</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2021-06</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Third Workshop on Privacy in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Oluwaseyi</namePart>
<namePart type="family">Feyisetan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sepideh</namePart>
<namePart type="family">Ghanavati</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shervin</namePart>
<namePart type="family">Malmasi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Patricia</namePart>
<namePart type="family">Thaine</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Online</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Contextual language models have led to significantly better results on a plethora of language understanding tasks, especially when pre-trained on the same data as the downstream task. While this additional pre-training usually improves performance, it can lead to information leakage and therefore risks the privacy of individuals mentioned in the training data. One method to guarantee the privacy of such individuals is to train a differentially-private model, but this usually comes at the expense of model performance. Moreover, it is hard to tell given a privacy parameter ε what was the effect on the trained representation. In this work we aim to guide future practitioners and researchers on how to improve privacy while maintaining good model performance. We demonstrate how to train a differentially-private pre-trained language model (i.e., BERT) with a privacy guarantee of ε=1 and with only a small degradation in performance. We experiment on a dataset of clinical notes with a model trained on a target entity extraction task, and compare it to a similar model trained without differential privacy. Finally, we present experiments showing how to interpret the differentially-private representation and understand the information lost and maintained in this process.</abstract>
<identifier type="citekey">hoory-etal-2021-learning</identifier>
<identifier type="doi">10.18653/v1/2021.privatenlp-1.3</identifier>
<location>
<url>https://aclanthology.org/2021.privatenlp-1.3</url>
</location>
<part>
<date>2021-06</date>
<extent unit="page">
<start>21</start>
<end>29</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Learning and Evaluating a Differentially Private Pre-trained Language Model
%A Hoory, Shlomo
%A Feder, Amir
%A Tendler, Avichai
%A Cohen, Alon
%A Erell, Sofia
%A Laish, Itay
%A Nakhost, Hootan
%A Stemmer, Uri
%A Benjamini, Ayelet
%A Hassidim, Avinatan
%A Matias, Yossi
%Y Feyisetan, Oluwaseyi
%Y Ghanavati, Sepideh
%Y Malmasi, Shervin
%Y Thaine, Patricia
%S Proceedings of the Third Workshop on Privacy in Natural Language Processing
%D 2021
%8 June
%I Association for Computational Linguistics
%C Online
%F hoory-etal-2021-learning
%X Contextual language models have led to significantly better results on a plethora of language understanding tasks, especially when pre-trained on the same data as the downstream task. While this additional pre-training usually improves performance, it can lead to information leakage and therefore risks the privacy of individuals mentioned in the training data. One method to guarantee the privacy of such individuals is to train a differentially-private model, but this usually comes at the expense of model performance. Moreover, it is hard to tell given a privacy parameter ε what was the effect on the trained representation. In this work we aim to guide future practitioners and researchers on how to improve privacy while maintaining good model performance. We demonstrate how to train a differentially-private pre-trained language model (i.e., BERT) with a privacy guarantee of ε=1 and with only a small degradation in performance. We experiment on a dataset of clinical notes with a model trained on a target entity extraction task, and compare it to a similar model trained without differential privacy. Finally, we present experiments showing how to interpret the differentially-private representation and understand the information lost and maintained in this process.
%R 10.18653/v1/2021.privatenlp-1.3
%U https://aclanthology.org/2021.privatenlp-1.3
%U https://doi.org/10.18653/v1/2021.privatenlp-1.3
%P 21-29
Markdown (Informal)
[Learning and Evaluating a Differentially Private Pre-trained Language Model](https://aclanthology.org/2021.privatenlp-1.3) (Hoory et al., PrivateNLP 2021)
ACL
- Shlomo Hoory, Amir Feder, Avichai Tendler, Alon Cohen, Sofia Erell, Itay Laish, Hootan Nakhost, Uri Stemmer, Ayelet Benjamini, Avinatan Hassidim, and Yossi Matias. 2021. Learning and Evaluating a Differentially Private Pre-trained Language Model. In Proceedings of the Third Workshop on Privacy in Natural Language Processing, pages 21–29, Online. Association for Computational Linguistics.