Papers by Eric Wallace
Imitation Attacks and Defenses for Black-box Machine Translation Systems (2020.emnlp-main)
Copied to clipboard
| Challenge: | Using simulated experiments, we demonstrate that MT systems can be stolen even when imitation models have different input data or architectures than their target models. |
| Approach: | They propose a defense that modifies translation outputs to misdirect optimization of imitation models. |
| Outcome: | The proposed defense degrades the adversary’s BLEU score and attack success rate at some cost in the defender’s performance and inference speed. |
AllenNLP Interpret: A Framework for Explaining Predictions of NLP Models (D19-3)
Copied to clipboard
| Challenge: | Existing interpretation codebases make it difficult to apply these methods to new models and tasks. |
| Approach: | They propose a framework for interpreting NLP models that provides explanations for specific models. |
| Outcome: | The proposed framework provides interpretation primitives for any AllenNLP model and task, a suite of built-in interpretation methods, and a library of front-end visualization components. |
Analyzing Dynamic Adversarial Training Data in the Limit (2022.findings-acl)
Copied to clipboard
| Challenge: | Dynamic adversarial data collection (DADC) can be used to build models that are robust across a wide range of test inputs. |
| Approach: | They propose to run Dynamic adversarial data collection over many rounds to maximize its training-time benefits. |
| Outcome: | The proposed model makes 26% fewer errors on the premise paragraphs compared to models trained on non-adversarial examples. |
Gradient-based Analysis of NLP Models is Manipulable (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Recent work has shown that explanation techniques can be unstable and can be manipulated to hide the actual reasoning behind the predictions of NLP models. |
| Approach: | They propose to merge a BERT-based sentiment classifier with a Facade Model that overwhelms the gradients without affecting the predictions. |
| Outcome: | The proposed model overwhelms the gradients without affecting the predictions on a variety of NLP tasks, such as sentiment analysis, NLI, and QA. |
Unfamiliar Finetuning Examples Control How Language Models Hallucinate (2025.naacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) generate plausible-sounding responses that are factually incorrect. |
| Approach: | They propose an approach to learn more reliable reward models by modifying how unfamiliar finetuning examples are supervised to influence model responses to unfamiliar queries. |
| Outcome: | The proposed approach improves the efficacy of RL factuality finetuning in long-form biography and book/movie plot generation tasks. |
Universal Adversarial Triggers for Attacking and Analyzing NLP (D19-1)
Copied to clipboard
| Challenge: | Using adversarial triggers, a model can produce a specific prediction . adversarial attacks are useful for evaluation and interpretation . |
| Approach: | They propose a gradient-guided search over tokens that finds short adversarial triggers that successfully trigger the target prediction. |
| Outcome: | The proposed algorithm finds short trigger sequences that successfully trigger the target prediction. |
Does BERT Pretrained on Clinical Notes Reveal Sensitive Data? (2021.naacl-main)
Copied to clipboard
| Challenge: | Pretraining large (masked) language models over EHR data has yielded consistent performance gains across tasks. |
| Approach: | They propose to use large Transformers to release pretraining models over EHRs . they propose to recover patient names and conditions associated with them . |
| Outcome: | The proposed models recover patient names and conditions associated with patients . the proposed models share the model parameters for use by other researchers . |
Interpreting Predictions of NLP Models (2020.emnlp-tutorials)
Copied to clipboard
| Challenge: | This tutorial will provide a background on interpretation techniques for neural NLP models. |
| Approach: | This tutorial will provide a background on interpretation techniques for NLP models . it will examine saliency maps, input perturbations, adversarial attacks and influence functions . |
| Outcome: | This tutorial will provide a background on interpretation techniques . examples-specific interpretations include saliency maps, input perturbations, adversarial attacks, influence functions . |
Do NLP Models Know Numbers? Probing Numeracy in Embeddings (D19-1)
Copied to clipboard
| Challenge: | Existing models cannot capture numeracy, but they can be useful for complex reasoning tasks. |
| Approach: | They investigate numerical reasoning capabilities of a question-answering model . they probe token embedding methods on synthetic list maximum, number decoding, and addition tasks. |
| Outcome: | The proposed model excels on questions that require numerical reasoning, i.e., it already captures numeracy. |
Detoxifying Language Models Risks Marginalizing Minority Voices (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing detoxification techniques have been proposed to mitigate toxic LM generations . e.g., detoxification makes LMs more brittle to distribution shift, especially on language used by marginalized groups . |
| Approach: | They propose to use detoxification techniques to reduce toxic LM generations without affecting perplexity or generation quality on nontoxic inputs. |
| Outcome: | The proposed methods hurt equity on language used by marginalized groups, the authors show . they show that detoxification makes LMs more brittle to distribution shift, they say . |
Concealed Data Poisoning Attacks on NLP Models (2021.naacl-main)
Copied to clipboard
| Challenge: | In contrast, adversarial attacks can cause model errors by modifying inputs, such as the universal triggers attack. |
| Approach: | They propose a data poisoning attack that allows an adversary to control model predictions whenever a desired trigger phrase is present in the input. |
| Outcome: | The proposed attack can cause model errors by modifying inputs, but it can also cause extra human annotation. |
Compositional Questions Do Not Necessitate Multi-hop Reasoning (P19-1)
Copied to clipboard
| Challenge: | a single-hop reasoning model can solve much more of the dataset than previously thought. |
| Approach: | They propose a single-hop BERT-based RC model that achieves 67 F1 . they propose an evaluation setting where humans are not shown all paragraphs . |
| Outcome: | The proposed model achieves 67 F1—comparable to state-of-the-art multi-hop models. |
Trick Me If You Can: Adversarial Writing of Trivia Challenge Questions (P18-3)
Copied to clipboard
| Challenge: | Existing question answering datasets are imperfect tests that do not expose model limitations. |
| Approach: | They develop an adversarial writing setting where humans interact with trained models and try to break them. |
| Outcome: | The proposed model-driven annotation process systematically stumps automated question answering systems. |
What Evidence Do Language Models Find Convincing? (2024.acl-long)
Copied to clipboard
| Challenge: | Current retrieval-augmented language models are tasked with subjective, contentious, and conflicting queries. |
| Approach: | They construct a dataset that pairs controversial queries with real-world evidence documents . they find current models rely heavily on relevance of a website to the query . |
| Outcome: | The proposed dataset pairs controversial queries with real-world evidence documents that contain different facts, arguments, and answers. |
Inferring Which Medical Treatments Work from Reports of Clinical Trials (N19-1)
Copied to clipboard
| Challenge: | Ideally, one would consult all available evidence from relevant clinical trials. however, these results are primarily disseminated in natural language scientific articles. |
| Approach: | They propose a task that involves inferring results from a full-text article describing randomized controlled trials with respect to a given intervention, comparator, and outcome of interest. |
| Outcome: | The proposed task consists of 10,000+ prompts coupled with full-text articles describing randomized controlled trials. |
Automated Crossword Solving (2022.acl-long)
Copied to clipboard
| Challenge: | Using neural question answering models, our system generates answer candidates and then combines loopy belief propagation with local search to find full puzzle solutions. |
| Approach: | They propose a new approach to automatically solving crossword puzzles that uses neural question answering models and loopy belief propagation with local search to find full puzzle solutions. |
| Outcome: | The proposed system outperforms even the best human solvers and can solve crosswords from a wide range of domains with perfect accuracy. |
Misleading Failures of Partial-input Baselines (P19-1)
Copied to clipboard
| Challenge: | Recent work establishes dataset difficulty and removes annotation artifacts via partial-input baselines. |
| Approach: | They propose to use partial-input baselines to establish dataset difficulty . they show how trivial patterns only visible in the full input can evade partial-output baseline . |
| Outcome: | The proposed model can solve 15% of previously-thought "hard" examples. |
ERASER: A Benchmark to Evaluate Rationalized NLP Models (2020.acl-main)
Copied to clipboard
Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, Byron C. Wallace
| Challenge: | State-of-the-art models in NLP are opaque in terms of how they come to make predictions. |
| Approach: | They propose to release a benchmark to measure the quality of rationales extracted by models and how faithful these rationale are to human annotators. |
| Outcome: | The proposed benchmark will enable researchers to compare models and track progress on interpretable models for NLP. |
Evaluating Models’ Local Decision Boundaries via Contrast Sets (2020.findings-emnlp)
Copied to clipboard
Matt Gardner, Yoav Artzi, Victoria Basmov, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, Nitish Gupta, Hannaneh Hajishirzi, Gabriel Ilharco, Daniel Khashabi, Kevin Lin, Jiangming Liu, Nelson F. Liu, Phoebe Mulcaire, Qiang Ning, Sameer Singh, Noah A. Smith, Sanjay Subramanian, Reut Tsarfaty, Eric Wallace, Ally Zhang, Ben Zhou
| Challenge: | Standard test sets for supervised learning evaluate in-distribution generalization but are misleading when a dataset has systematic gaps. |
| Approach: | They propose a more rigorous annotation paradigm for NLP that helps to close systematic gaps in the test data. |
| Outcome: | The proposed model performs significantly lower on contrast sets than on the original test sets—up to 25% in some cases. |
AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts (2020.emnlp-main)
Copied to clipboard
| Challenge: | Pretrained language models have been successful when finetuned to downstream tasks . however, it is difficult to determine whether the knowledge that finetuning LMs contain is learned during the pretraining or the finetailing process. |
| Approach: | They propose a method to create prompts for a diverse set of tasks using a gradient-guided search. |
| Outcome: | The proposed method performs sentiment analysis and natural language inference without additional parameters and finetuning. |
Pathologies of Neural Models Make Interpretations Difficult (D18-1)
Copied to clipboard
| Challenge: | Existing methods for NLP use input reduction to determine a word's importance . human accuracy degrades when shown the reduced examples instead of the original . |
| Approach: | They propose a process that iteratively removes the least important word from an input . they show human models make the same predictions with high confidence . |
| Outcome: | The proposed methods expose pathological behaviors of neural models . human experiments show that reduced examples lack information to support the prediction of any label . |
Cutting Down on Prompts and Parameters: Simple Few-Shot Learning with Language Models (2022.findings-acl)
Copied to clipboard
| Challenge: | Prompting language models (LMs) with training examples and task descriptions has been seen as critical to recent successes in few-shot learning. |
| Approach: | They propose to fine tune masked language models with training examples and task descriptions to reduce prompt engineering by using null prompts. |
| Outcome: | The proposed prompts can be used to improve few-shot learning by finetuning only the bias terms while updating only 0.1% of the parameters. |
Pretrained Transformers Improve Out-of-Distribution Robustness (2020.acl-main)
Copied to clipboard
| Challenge: | Pretrained Transformers are more effective at detecting anomalous or OOD examples, while many previous models are frequently worse than chance. |
| Approach: | They construct a new robustness benchmark with real distribution shifts to measure out-of-distribution generalization for seven NLP datasets and compare them to previous models. |
| Outcome: | The proposed model generalizations for seven datasets show that pretrained Transformers are significantly less effective at detecting anomalous or OOD examples, while many previous models are often worse than chance. |