Large Scale Author Obfuscation Using Siamese Variational Auto-Encoder: The SiamAO System
25333Authors
nan
Abstract
AbstractAuthor obfuscation is the task of masking the author of a piece of text, with applications in privacy. Recent advances in deep neural networks have boosted author identification performance making author obfuscation more challenging. Existing approaches to author obfuscation are largely heuristic. Obfuscation can, however, be thought of as the construction of adversarial examples to attack author identification, suggesting that the deep learning architectures used for adversarial attacks could have application here. Current architectures are proposed to construct adversarial examples against classification-based models, which in author identification would exclude the high-performing similarity-based models employed when facing large number of authorial classes. In this paper, we propose the first deep learning architecture for constructing adversarial examples against similarity-based learners, and explore its application to author obfuscation. We analyse the output from both success in obfuscation and language acceptability, as well as comparing the performance with some common baselines, and showing promising results in finding a balance between safety and soundness of the perturbed texts.
Keywords
Full Content
Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics (*SEM), pages 1–11 Barcelona, Spain (Online), December 12–13, 2020 1 Improving Medical NLI Using Context-Aware Domain Knowledge Shaika Chowdhury Department of Computer Science University of Illinois at Chicago schowd21@uic.edu Yuan Luo Department of Preventive Medicine Northwestern University yuan.luo@northwestern.edu Philip S. Yu Department of Computer Science University of Illinois at Chicago psyu@uic.edu Abstract Domain knowledge is important to understand both the lexical and relational associations of words in natural language text, especially for domain-specific tasks like Natural Language Inference (NLI) in the medical domain, where due to the lack of a large annotated dataset such knowledge cannot be implicitly learned during training. However, because of the lin- guistic idiosyncrasies of clinical texts (e.g., shorthand jargon), solely relying on domain knowledge from an external knowledge base (e.g., UMLS) can lead to wrong inference pre- dictions as it disregards contextual informa- tion and, hence, does not return the most rel- evant mapping. To remedy this, we devise a knowledge adaptive approach for medical NLI that encodes the premise/hypothesis texts by leveraging supplementary external knowl- edge, alongside the UMLS, based on the word contexts. By incorporating refined domain knowledge at both the lexical and relational lev- els through a multi-source attention mecha- nism, it is able to align the token-level interac- tions between the premise and hypothesis more effectively. Comprehensive experiments and case study on the recently released MedNLI dataset are conducted to validate the effective- ness of the proposed approach. 1 Introduction Natural Language Inference is a fundamentally im- portant but challenging task in Natural Language Processing (NLP) as it requires understanding and reasoning over natural language texts (MacCartney and Manning, 2009). As a result, a good performing This work is licensed under a Creative Commons Attri- bution 4.0 International License. License details: http: //creativecommons.org/licenses/by/4.0/. NLI system is considered indispensable for down- stream NLP applications such as question answer- ing and automatic text summarization (Harabagiu and Hickl, 2006; Lloret et al., 2008). Given a pair of sentences, a premise 푝 and a hypothesis ℎ, the goal of NLI is to determine whether the semantic relationship between 푝 and ℎ is among 푒푛푡푎푖푙푚푒푛푡, 푐표푛푡푟푎푑푖푐푡푖표푛 and 푛푒푢푡푟푎푙. The ability to understand natural language text innately requires to deal with background knowl- edge 1 (Long et al., 2017; Weissenborn et al., 2017). A robust NLI model usually needs to reason over two types of background knowledge - lexical and relational (Weissenborn et al., 2017). The former pertains to understanding the concepts expressed by the words in the text, while the latter learns the semantic relations between the different concepts. When performing NLI on open domain data, it is assumed that the background knowledge will be implicitly learned from the training corpora. Re- Premise:�-�DMII�complicated�by�DM�neuropathy�-�PVD�s/p�L CFA� w/balloon� angioplasty� of� SFA� and� AK� [**Doctor� Last Name�**]�artery�w/�persistent�non-healing�ulcer�at�the�lateral�and medial�malleolus,�non-healing�L�pedalulcer�-�Hypertension�-�h/o MDR� Pseudomonas� and� MRSA� skin� infections� -� h/o hemorrhagic� pancreatitis� ([**2857**])� -� h/o� cholecystitis� (still has�gallbladder) Hypothesis:�Patient�has�multiple�diabetes�related�comorbidities. Label:�Entailment Figure 1: Sample premise-hypothesis pair from MedNLI. The words in red, “DMII" and …
Content truncated to 500 words
Labeled Text
AbstractAuthor obfuscation is the task of masking the author of a piece of text, with applications in privacy. Recent advances in deep neural networks have boosted author identification performance making author obfuscation more challenging. Existing approaches to author obfuscation are largely heuristic. Obfuscation can, however, be thought of as the construction of adversarial examples to attack author identification, suggesting that the deep learning architectures used for adversarial attacks could have application here. Current architectures are proposed to construct adversarial examples against classification-based models, which in author identification would exclude the high-performing similarity-based models employed when facing large number of authorial classes. In this paper, we propose the first deep learning architecture for constructing adversarial examples against similarity-based learners, and explore its application to author obfuscation. We analyse the output from both success in obfuscation and language acceptability, as well as comparing the performance with some common baselines, and showing promising results in finding a balance between safety and soundness of the perturbed texts.
Publication Details
- Paper ID:
- 25333
- Year:
- 2020
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan