Papers by Johannes Welbl
Achieving Verified Robustness to Symbol Substitutions via Interval Bound Propagation (D19-1)
Copied to clipboard
Po-Sen Huang, Robert Stanforth, Johannes Welbl, Chris Dyer, Dani Yogatama, Sven Gowal, Krishnamurthy Dvijotham, Pushmeet Kohli
| Challenge: | Recent work has exposed the vulnerabilities of neural NLP models, e.g. with small, semantically invariant input alterations. |
| Approach: | They propose to model text classification under synonym replacements or character flip perturbations and then use a formal model verification method to verify its robustness. |
| Outcome: | The proposed models show little difference in terms of nominal accuracy, but have much improved verified accuracy under perturbations and come with an efficiently computable formal guarantee on worst case adversaries. |
Beat the AI: Investigating Adversarial Human Annotation for Reading Comprehension (2020.tacl-1)
Copied to clipboard
| Challenge: | Innovations in annotation methodologies have been a catalyst for Reading Comprehension (RC) datasets and models. |
| Approach: | They propose to use a model-in-the-annotation-loop approach to train adversarial models in three different settings to explore reproducibility of the adversarial effect, transfer from data collected with varying model- in-the loop strengths, and generalization to data collected without a modeling model. |
| Outcome: | The proposed approach achieves 39.9F1 on questions it cannot answer when trained on SQUAD, but lower than when trained using RoBERTa itself (41.0F1). |
Jack the Reader – A Machine Reading Framework (P18-4)
Copied to clipboard
Dirk Weissenborn, Pasquale Minervini, Isabelle Augenstein, Johannes Welbl, Tim Rocktäschel, Matko Bošnjak, Jeff Mitchell, Thomas Demeester, Tim Dettmers, Pontus Stenetorp, Sebastian Riedel
| Challenge: | Many Machine Reading and Natural Language Understanding tasks require reading supporting text in order to answer questions. |
| Approach: | They propose a framework for Machine Reading that allows for quick prototyping by component reuse and evaluation of new models on existing datasets. |
| Outcome: | The proposed framework supports question answering, natural language inference and link prediction tasks. |
Undersensitivity in Neural Reading Comprehension (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing models generalise well to in-distribution test sets, yet perform poorly on adversarially selected data. |
| Approach: | They propose an adversarial attack which searches among semantic variations of the question for which a model erroneously predicts the same answer, and with even higher probability. |
| Outcome: | The proposed attack reduces the vulnerability of models trained on SQuAD2.0 and NewsQA, and outperforms a conventional model by as much as 10.9% F1. |
Challenges in Detoxifying Language Models (2021.findings-emnlp)
Copied to clipboard
Johannes Welbl, Amelia Glaese, Jonathan Uesato, Sumanth Dathathri, John Mellor, Lisa Anne Hendricks, Kirsty Anderson, Pushmeet Kohli, Ben Coppin, Po-Sen Huang
| Challenge: | Prior work often relies on automatic evaluation of LM toxicity. |
| Approach: | They evaluate toxicity mitigation strategies for automated and human evaluations . they find human raters disagree with high automatic toxicity scores after strong toxicity reduction interventions . |
| Outcome: | The proposed methods reduce LM toxicity but lower coverage for marginalized texts . human raters disagree with high toxicity scores after strong toxicity reduction interventions . |
Reducing Sentiment Bias in Language Models via Counterfactual Evaluation (2020.findings-emnlp)
Copied to clipboard
Po-Sen Huang, Huan Zhang, Ray Jiang, Robert Stanforth, Johannes Welbl, Jack Rae, Vishal Maini, Dani Yogatama, Pushmeet Kohli
| Challenge: | Language modeling has advanced rapidly due to efficient model architectures and the availability of large text corpora. |
| Approach: | They propose to embed and regularize sentiment prediction-derived regularizations on the language model’s latent representations to reduce bias in the sentiment of generated text. |
| Outcome: | The proposed methods reduce bias in the sentiment of generated text by adopting individual and group fairness metrics from the fair machine learning literature. |