Types of Out-of-Distribution Texts and How to Detect Them

Arora, Udit; Huang, William; He, He

doi:10.18653/v1/2021.emnlp-main.835

Cited by 28 publications

(51 citation statements)

References 34 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This might increase the risk of finding artefacts due to changes in the patient population. Third, to identify changes in writing style and writing conventions, one could use methods for out-of-distribution detection such as perplexity of the texts over time 28 . This, however, requires the use of a GPU as well as a generative language model, neither of which were available to us.…”

Section: Limitationsmentioning

confidence: 99%

Lexical Stability of Psychiatric Clinical Notes from Electronic Health Records over a Decade

Hansen

Enevoldsen

Bernstorff

et al. 2022

Preprint

View full text Add to dashboard Cite

Natural Language Processing methods hold promise for improving clinical prediction by utilising information otherwise hidden in the clinical notes of electronic health records. However, clinical practice-as well as the systems and databases in which clinical notes are recorded and stored-change over time. As a consequence, the content of clinical notes may also change over time, which could degrade the performance of prediction models. Despite its importance, the stability of clinical notes over time has rarely been tested. Therefore, in this study, we examined the lexical stability of clinical notes from the Psychiatric Services of the Central Denmark Region in the period from January 1, 2011, to November 22, 2021 (a total of 14,811,551 clinical notes describing 129,570 patients) by quantifying sentence length, readability, syntactic complexity and clinical content - and estimating changepoints in these metrics. We find lexical and syntactic stability over time, which bodes well for the use of Natural Language Processing for predictive modelling in clinical practice.

show abstract

Section: Limitationsmentioning

confidence: 99%

Lexical Stability of Psychiatric Clinical Notes from Electronic Health Records over a Decade

Hansen

Enevoldsen

Bernstorff

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Therefore, the model is reasonable to misclassify such data, which is irrelevant to backdoor attacks. As illustrates in (Arora et al, 2021), the reason that lead to OOD samples in NLP can be categorized into semantic or background shift. Following (Arora et al, 2021), we utilize the density estimation (PPL) 1 for OOD text detection.…”

Section: Some Empirical Analysis Of the Gap Between Asr And Asrdmentioning

confidence: 99%

“…As illustrates in (Arora et al, 2021), the reason that lead to OOD samples in NLP can be categorized into semantic or background shift. Following (Arora et al, 2021), we utilize the density estimation (PPL) 1 for OOD text detection. Then we launch OOD detection on the poisoned test set generated by StyAtk.…”

Section: Some Empirical Analysis Of the Gap Between Asr And Asrdmentioning

confidence: 99%

“…Then we launch OOD detection on the poisoned test set generated by StyAtk. Specifically, we choose the SST-2 dataset since the PPL method performs well on SST-2 (Arora et al, 2021) which shows the best reliability. The results are shown in the Figure 3, where we can see that 21.1% of the poison testset samples belong to OOD samples.…”

Section: Some Empirical Analysis Of the Gap Between Asr And Asrdmentioning

confidence: 99%

See 1 more Smart Citation

Rethink the Evaluation for Attack Strength of Backdoor Attacks in Natural Language Processing

Shen¹,

Jiang²,

Liu³

et al. 2022

Preprint

View full text Add to dashboard Cite

Recently, it has been shown that natural language processing (NLP) models are vulnerable to a kind of security threat called the Backdoor Attack, which utilizes a 'backdoor trigger' paradigm to mislead the models. The most threatening backdoor attack is the stealthy backdoor, which defines the triggers as text style or syntactic. Although they have achieved an incredible high attack success rate (ASR), we find that the principal factor contributing to their ASR is not the 'backdoor trigger' paradigm. Thus the capacity of these stealthy backdoor attacks is overestimated when categorized as backdoor attacks. Therefore, to evaluate the real attack power of backdoor attacks, we propose a new metric called attack successful rate difference (ASRD), which measures the ASR difference between clean state and poison state models. Besides, since the defenses against stealthy backdoor attacks are absent, we propose Trigger Breaker, consisting of two too simple tricks that can defend against stealthy backdoor attacks effectively. Experiments on text classification tasks show that our method achieves significantly better performance than state-of-the-art defense methods against stealthy backdoor attacks.

show abstract

“…Language learning with distribution shift (a.k.a., OOD, short for out-of-distribution) has drawn a growing attention in the NLP community Arora et al, 2021). Most previous work focuses on OOD in different domains (Muandet et al, 2013;Ganin et al, 2015) and studies how to learn generalizable cross-domain features.…”

Section: Introductionmentioning

confidence: 99%

Time Will Change Things: An Empirical Study on Dynamic Language Understanding in Social Media Classification

Zhang¹,

Li²

2022

Preprint

View full text Add to dashboard Cite

Language features are ever-evolving in the real-world social media environment. Many trained models in natural language understanding (NLU), ineffective in semantic inference for unseen features, might consequently struggle with the deteriorating performance in dynamicity. To address this challenge, we empirically study social media NLU in a dynamic setup, where models are trained on the past data and test on the future. It better reflects the realistic practice compared to the commonly-adopted static setup of random data split. To further analyze model adaption to the dynamicity, we explore the usefulness of leveraging some unlabeled data created after a model is trained. The performance of unsupervised domain adaption baselines based on auto-encoding and pseudo-labeling and a joint framework coupling them both are examined in the experiments. Substantial results on four social media tasks imply the universally negative effects of evolving environments over classification accuracy, while auto-encoding and pseudo-labeling collaboratively show the best robustness in dynamicity.

show abstract

Types of Out-of-Distribution Texts and How to Detect Them

Cited by 28 publications

References 34 publications

Lexical Stability of Psychiatric Clinical Notes from Electronic Health Records over a Decade

Lexical Stability of Psychiatric Clinical Notes from Electronic Health Records over a Decade

Rethink the Evaluation for Attack Strength of Backdoor Attacks in Natural Language Processing

Time Will Change Things: An Empirical Study on Dynamic Language Understanding in Social Media Classification

Contact Info

Product

Resources

About