Differential Privacy, Linguistic Fairness, and Training Data Influence: Impossibility and Possibility Theorems for Multilingual Language Models

Language models such as mBERT, XLM-R, and BLOOM aim to achieve multilingual generalization or compression to facilitate transfer to a large number of (potentially unseen) languages. However, these models should ideally also be private, linguistically fair, and transparent, by relating their predictions to training data. Can these requirements be simultaneously satisfied? We show that multilingual compression and linguistic fairness are compatible with differential privacy, but that differential privacy is at odds with training data influence sparsity, an objective for transparency. We further present a series of experiments on two common NLP tasks and evaluate multilingual compression and training data influence sparsity under different privacy guarantees, exploring these trade-offs in more detail. Our results suggest that we need to develop ways to jointly optimize for these objectives in order to find practical trade-offs.

[1]  Alexander M. Rush,et al.  BLOOM: A 176B-Parameter Open-Access Multilingual Language Model , 2022, ArXiv.

[2]  R. Shokri,et al.  Data Privacy and Trustworthy Machine Learning , 2022, IEEE Security & Privacy.

[3]  G. Chrysostomou.  Explainable Natural Language Processing , 2022, Computational Linguistics.

[4]  Ivan Vulic,et al.  Square One Bias in NLP: Towards a Multi-Dimensional Exploration of the Research Manifold , 2022, FINDINGS.

[5]  Hao Zhou,et al.  Enhancing Cross-lingual Transfer by Manifold Mixup , 2022, ICLR.

[6]  Victor Petrén Bach Hansen,et al.  The Impact of Differential Privacy on Group Disparity Mitigation , 2022, PRIVATENLP.

[7]  Samuel R. Bowman,et al.  One size does not fit all: Investigating strategies for differentially-private learning across NLP tasks , 2022 .

[8]  Anders Sogaard,et al.  Revisiting Methods for Finding Influential Examples , 2021, ArXiv.

[9]  Huseyin A. Inan,et al.  Differentially Private Fine-tuning of Language Models , 2021, ICLR.

[10]  Tatsunori B. Hashimoto,et al.  Large Language Models Can Be Strong Differentially Private Learners , 2021, ICLR.

[11]  Nicolas Papernot,et al.  Hyperparameter Tuning with Renyi Differential Privacy , 2021, ICLR.

[12]  Graham Cormode,et al.  Opacus: User-Friendly Differential Privacy Library in PyTorch , 2021, ArXiv.

[13]  Samuel R. Bowman,et al.  Fine-Tuned Transformers Show Clusters of Similar Representations Across Layers , 2021, BLACKBOXNLP.

[14]  Hinrich Schutze,et al.  Wine is Not v i n. - On the Compatibility of Tokenizations Across Languages , 2021, EMNLP.

[15]  Anders Sogaard,et al.  The Impact of Positional Encodings on Multilingual Compression , 2021, EMNLP.

[16]  Alexander M. Rush,et al.  Datasets: A Community Library for Natural Language Processing , 2021, EMNLP.

[17]  Ivan Habernal,et al.  When differential privacy meets NLP: The devil is in the detail , 2021, EMNLP.

[18]  Carsten Eickhoff,et al.  IsoScore: Measuring the Uniformity of Embedding Space Utilization , 2021, FINDINGS.

[19]  Fatemehsadat Mireshghallah,et al.  When Differential Privacy Meets Interpretability: A Case Study , 2021, ArXiv.

[20]  Ziming Huang,et al.  On Sample Based Explanation Methods for NLP: Faithfulness, Efficiency and Semantic Evaluation , 2021, ACL.

[21]  Lidong Bing,et al.  On the Effectiveness of Adapter-based Tuning for Pretrained Language Model Adaptation , 2021, ACL.

[22]  Mohammad Taher Pilehvar,et al.  A Cluster-based Approach for Improving Isotropy in Contextual Embedding Space , 2021, ACL.

[23]  Arianna Bisazza,et al.  Using Confidential Data for Domain Adaptation of Neural Machine Translation , 2021, PRIVATENLP.

[24]  Kamalika Chaudhuri,et al.  Understanding Instance-based Interpretability of Variational Auto-Encoders , 2021, NeurIPS.

[25]  Monojit Choudhury,et al.  How Linguistically Fair Are Multilingual Pre-Trained Language Models? , 2021, AAAI.

[26]  Serena Booth,et al.  Do Feature Attribution Methods Correctly Attribute Features? , 2021, AAAI.

[27]  Giacomo Spigler,et al.  Investigating Trade-offs in Utility, Fairness and Differential Privacy in Neural Networks , 2021, ArXiv.

[28]  Benjamin Muller,et al.  First Align, then Predict: Understanding the Cross-Lingual Ability of Multilingual BERT , 2021, EACL.

[29]  Khaled Shaalan,et al.  Self-Training Pre-Trained Language Models for Zero- and Few-Shot Multi-Dialectal Arabic Sequence Labeling , 2021, EACL.

[30]  Bo Liu,et al.  When Machine Learning Meets Privacy , 2020, ACM Comput. Surv..

[31]  Michael A. Lepori,et al.  Picking BERT’s Brain: Probing for Linguistic Dependencies in Contextualized Embeddings Using Representational Similarity Analysis , 2020, COLING.

[32]  Dan Boneh,et al.  Differentially Private Learning Needs Better Features (or Much More Data) , 2020, ICLR.

[33]  R. Shokri,et al.  On the Privacy Risks of Algorithmic Fairness , 2020, 2021 IEEE European Symposium on Security and Privacy (EuroS&P).

[34]  Goran Glavaš,et al.  From Zero to Hero: On the Limitations of Zero-Shot Language Transfer with Multilingual Transformers , 2020, EMNLP.

[35]  Hinrich Schütze,et al.  Identifying Elements Essential for BERT’s Multilinguality , 2020, EMNLP.

[36]  Lingjuan Lyu,et al.  Differentially Private Representation for NLP: Formal Guarantee and An Empirical Study on Privacy and Fairness , 2020, FINDINGS.

[37]  Ekaterina Shutova,et al.  What does it mean to be language-agnostic? Probing multilingual sentence encoders for typological properties , 2020, ArXiv.

[38]  Dylan Slack,et al.  Differentially Private Language Models Benefit from Public Pre-training , 2020, PRIVATENLP.

[39]  N. Arun,et al.  Assessing the (Un)Trustworthiness of Saliency Maps for Localizing Abnormalities in Medical Imaging , 2020, medRxiv.

[40]  Yulia Tsvetkov,et al.  Explaining Black Box Predictions and Unveiling Data Artifacts through Influence Functions , 2020, ACL.

[41]  Bill Yuchen Lin,et al.  IsoBN: Fine-Tuning BERT with Isotropic Batch Normalization , 2020, AAAI.

[42]  Jing Huang,et al.  Improving Neural Language Generation with Spectrum Control , 2020, ICLR.

[43]  Sampo Pyysalo,et al.  Universal Dependencies v2: An Evergrowing Multilingual Treebank Collection , 2020, LREC.

[44]  Iryna Gurevych,et al.  Making Monolingual Sentence Embeddings Multilingual Using Knowledge Distillation , 2020, EMNLP.

[45]  Malvina Nissim,et al.  What’s so special about BERT’s layers? A closer look at the NLP pipeline in monolingual and multilingual models , 2020, FINDINGS.

[46]  Alexander M. Fraser,et al.  On the Language Neutrality of Pre-trained Multilingual Representations , 2020, FINDINGS.

[47]  Thomas Steinke,et al.  The Discrete Gaussian for Differential Privacy , 2020, NeurIPS.

[48]  知秀 柴田.  5分で分かる!? 有名論文ナナメ読み：Jacob Devlin et al. : BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding , 2020 .

[49]  Frederick Liu,et al.  Estimating Training Data Influence by Tracking Gradient Descent , 2020, NeurIPS.

[50]  Dan Roth,et al.  Cross-Lingual Ability of Multilingual BERT: An Empirical Study , 2019, ICLR.

[51]  Guillaume Charpiat,et al.  Input Similarity from the Neural Network Perspective , 2019, NeurIPS.

[52]  Natalia Gimelshein,et al.  PyTorch: An Imperative Style, High-Performance Deep Learning Library , 2019, NeurIPS.

[53]  Myle Ott,et al.  Unsupervised Cross-lingual Representation Learning at Scale , 2019, ACL.

[54]  Luke Zettlemoyer,et al.  Emerging Cross-lingual Structure in Pretrained Language Models , 2019, ACL.

[55]  S. Feizi,et al.  Second-Order Group Influence Functions for Black-Box Predictions , 2019, ArXiv.

[56]  Richard Socher,et al.  BERT is Not an Interlingua and the Bias of Tokenization , 2019, EMNLP.

[57]  Lysandre Debut,et al.  HuggingFace's Transformers: State-of-the-art Natural Language Processing , 2019, ArXiv.

[58]  Kawin Ethayarajh,et al.  How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings , 2019, EMNLP.

[59]  Li Zhang,et al.  Rényi Differential Privacy of the Sampled Gaussian Mechanism , 2019, ArXiv.

[60]  Kristina Lerman,et al.  A Survey on Bias and Fairness in Machine Learning , 2019, ACM Comput. Surv..

[61]  Jordan Rodu,et al.  Getting in Shape: Word Embedding SubSpaces , 2019, IJCAI.

[62]  Di He,et al.  Representation Degeneration Problem in Training Natural Language Generation Models , 2019, ICLR.

[63]  Holger Schwenk,et al.  WikiMatrix: Mining 135M Parallel Sentences in 1620 Language Pairs from Wikipedia , 2019, EACL.

[64]  Reza Shokri,et al.  On the Privacy Risks of Model Explanations , 2019, AIES.

[65]  Varun Gupta,et al.  On the Compatibility of Privacy and Fairness , 2019, UMAP.

[66]  Eva Schlinger,et al.  How Multilingual is Multilingual BERT? , 2019, ACL.

[67]  Vitaly Shmatikov,et al.  Differential Privacy Has Disparate Impact on Model Accuracy , 2019, NeurIPS.

[68]  Vitalii Zhelezniak,et al.  Correlation Coefficients and Semantic Textual Similarity , 2019, NAACL.

[69]  Afra Alishahi,et al.  Correlating Neural and Symbolic Representations of Language , 2019, ACL.

[70]  Geoffrey E. Hinton,et al.  Similarity of Neural Network Representations Revisited , 2019, ICML.

[71]  Percy Liang,et al.  On the Accuracy of Influence Functions for Measuring Group Effects , 2019, NeurIPS.

[72]  Mark Dredze,et al.  Beto, Bentz, Becas: The Surprising Cross-Lingual Effectiveness of BERT , 2019, EMNLP.

[73]  R. C. Williamson,et al.  Fairness risk measures , 2019, ICML.

[74]  Holger Schwenk,et al.  Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond , 2018, Transactions of the Association for Computational Linguistics.

[75]  Grzegorz Chrupala,et al.  Symbolic Inductive Bias for Visually Grounded Learning of Spoken Language , 2018, ACL.

[76]  Aaron Roth,et al.  Differentially Private Fair Learning , 2018, ICML.

[77]  Pradeep Ravikumar,et al.  Representer Point Selection for Explaining Deep Neural Networks , 2018, NeurIPS.

[78]  Kunal Talwar,et al.  Private selection from private candidates , 2018, STOC.

[79]  Wei Ding,et al.  Tight Analysis of Privacy and Utility Tradeoff in Approximate Differential Privacy , 2018, AISTATS.

[80]  Guillaume Lample,et al.  XNLI: Evaluating Cross-lingual Sentence Representations , 2018, EMNLP.

[81]  Marco Baroni,et al.  How agents see things: On visual representations in an emergent language game , 2018, EMNLP.

[82]  Julia Rubin,et al.  Fairness Definitions Explained , 2018, 2018 IEEE/ACM International Workshop on Software Fairness (FairWare).

[83]  Samuel R. Bowman,et al.  GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding , 2018, BlackboxNLP@EMNLP.

[84]  Frank Hutter,et al.  Decoupled Weight Decay Regularization , 2017, ICLR.

[85]  Dumitru Erhan,et al.  The (Un)reliability of saliency methods , 2017, Explainable AI.

[86]  H. Brendan McMahan,et al.  Learning Differentially Private Recurrent Language Models , 2017, ICLR.

[87]  Thomas Miconi,et al.  The impossibility of "fairness": a generalized impossibility result for decisions , 2017, 1707.01195.

[88]  M. Kearns,et al.  Fairness in Criminal Justice Risk Assessments: The State of the Art , 2017, Sociological Methods & Research.

[89]  Percy Liang,et al.  Understanding Black-box Predictions via Influence Functions , 2017, ICML.

[90]  Ilya Mironov,et al.  Rényi Differential Privacy , 2017, 2017 IEEE 30th Computer Security Foundations Symposium (CSF).

[91]  Jörn Diedrichsen,et al.  Representational models: A common framework for understanding encoding, pattern-component, and representational-similarity analysis , 2017, bioRxiv.

[92]  Abhishek Das,et al.  Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization , 2016, 2017 IEEE International Conference on Computer Vision (ICCV).

[93]  Ian Goodfellow,et al.  Deep Learning with Differential Privacy , 2016, CCS.

[94]  Jimmy Ba,et al.  Adam: A Method for Stochastic Optimization , 2014, ICLR.

[95]  Aaron Roth,et al.  The Algorithmic Foundations of Differential Privacy , 2014, Found. Trends Theor. Comput. Sci..

[96]  Ninghui Li,et al.  On sampling, anonymization, and differential privacy or, k-anonymization meets differential privacy , 2011, ASIACCS '12.

[97]  Nikolaus Kriegeskorte,et al.  Frontiers in Systems Neuroscience Systems Neuroscience , 2022 .

[98]  Cynthia Dwork,et al.  Differential Privacy , 2006, ICALP.

[99]  S Edelman,et al.  Representation is representation of similarities , 1996, Behavioral and Brain Sciences.

[100]  Christopher M. Bishop,et al.  Current address: Microsoft Research, , 2022 .

[101]  Michele Banko,et al.  Practical Transformer-based Multilingual Text Classification , 2021, NAACL.

[102]  Goran Glavas,et al.  Is Supervised Syntactic Parsing Beneficial for Language Understanding Tasks? An Empirical Investigation , 2021, EACL.

[103]  Proceedings of the First Workshop on Trustworthy Natural Language Processing , 2021 .

[104]  Genta Indra Winata,et al.  Preserving Cross-Linguality of Pre-trained Models via Continual Learning , 2021, REPL4NLP.