Despite the rising popularity of saliency-based explanations, the research community remains at an impasse, facing doubts concerning their purpose, efficacy, and tendency to contradict each other. Seeking to unite the community’s efforts around common goals, several recent works have proposed evaluation metrics. In this paper, we critically examine two sets of metrics: the ERASER metrics (comprehensiveness and sufficiency) and the EVAL-X metrics, focusing our inquiry on natural language processing. First, we show that we can inflate a model’s comprehensiveness and sufficiency scores dramatically without altering its predictions or explanations on in-distribution test inputs. Our strategy exploits the tendency for extracted explanations and their complements to be “out-of-support” relative to each other and in-distribution inputs. Next, we demonstrate that the EVAL-X metrics can be inflated arbitrarily by a simple method that encodes the label, even though EVAL-X is precisely motivated to address such exploits. Our results raise doubts about the ability of current metrics to guide explainability research, underscoring the need for a broader reassessment of what precisely these metrics are intended to capture.

As corporations rush to integrate large language models (LLMs) it is critical that they provide factually accurate information, that is robust to any presuppositions that a user may express. In this work, we introduce UPHILL, a dataset consisting of health-related queries with varying degrees of presuppositions. Using UPHILL, we evaluate the factual accuracy and consistency of InstructGPT, ChatGPT, GPT-4 and Bing Copilot models. We find that while model responses rarely contradict true health claims (posed as questions), all investigated models fail to challenge false claims. Alarmingly, responses from these models agree with 23-32% of the existing false claims, and 49-55% with novel fabricated claims. As we increase the extent of presupposition in input queries, responses from all models except Bing Copilot agree with the claim considerably more often, regardless of its veracity. Given the moderate factual accuracy, and the inability of models to challenge false assumptions, our work calls for a careful assessment of current LLMs for use in high-stakes scenarios.

Large language models (LLMs) encode vast amounts of world knowledge. However, since these models are trained on large swaths of internet data, they are at risk of inordinately capturing information about dominant groups. This imbalance can propagate into generated language. In this work, we study and operationalise a form of geographical erasure wherein language models underpredict certain countries. We demonstrate consistent instances of erasure across a range of LLMs. We discover that erasure strongly correlates with low frequencies of country mentions in the training corpus. Lastly, we mitigate erasure by finetuning using a custom objective.

In recent years, NLP practitioners have converged on the following practice: (i) import an off-the-shelf pretrained (masked) language model; (ii) append a multilayer perceptron atop the CLS token’s hidden representation (with randomly initialized weights); and (iii) fine-tune the entire model on a downstream task (MLP-FT). This procedure has produced massive gains on standard NLP benchmarks, but these models remain brittle, even to mild adversarial perturbations. In this work, we demonstrate surprising gains in adversarial robustness enjoyed by Model-tuning Via Prompts (MVP), an alternative method of adapting to downstream tasks. Rather than appending an MLP head to make output prediction, MVP appends a prompt template to the input, and makes prediction via text infilling/completion. Across 5 NLP datasets, 4 adversarial attacks, and 3 different models, MVP improves performance against adversarial substitutions by an average of 8% over standard methods and even outperforms adversarial training-based state-of-art defenses by 3.5%. By combining MVP with adversarial training, we achieve further improvements in adversarial robustness while maintaining performance on unperturbed examples. Finally, we conduct ablations to investigate the mechanism underlying these gains. Notably, we find that the main causes of vulnerability of MLP-FT can be attributed to the misalignment between pre-training and fine-tuning tasks, and the randomly initialized MLP parameters.

Many adversarial attacks in NLP perturb text in puts to produce visually similar strings (‘ergo’, ‘εrgo’) which are legible to humans but degrade model performance. Although preserving legibility is a necessary condition for text perturbation, little work has been done to systematically characterize it; instead, legibility is typically loosely enforced via intuitions around the nature and extent of perturbations. Particularly, it is unclear to what extent can inputs be perturbed while preserving legibility, or how to quantify the legibility of a perturbed string. In this work, we address this gap by learning models that predict the legibility of a perturbed string, and rank candidate perturbations based on their legibility. To do so, we collect and release LEGIT, a human-annotated dataset comprising the legibility of visually perturbed text. Using this dataset, we build both text- and vision-based models which achieve up to 0.91 F score in predicting whether an input is legible, and an accuracy of 0.86 in predicting which of two given perturbations is more legible. Additionally, we discover that legible perturbations from the LEGIT dataset are more effective at lowering the performance of NLP models than best-known attack strategies, suggesting that current models may be vulnerable to a broad range of perturbations beyond what is captured by existing visual attacks.

While many methods purport to explain predictions by highlighting salient features, what aims these explanations serve and how they ought to be evaluated often go unstated. In this work, we introduce a framework to quantify the value of explanations via the accuracy gains that they confer on a student model trained to simulate a teacher model. Crucially, the explanations are available to the student during training, but are not available at test time. Compared with prior proposals, our approach is less easily gamed, enabling principled, automatic, model-agnostic evaluation of attributions. Using our framework, we compare numerous attribution methods for text classification and question answering, and observe quantitative differences that are consistent (to a moderate to high degree) across different student model architectures and learning strategies.1

Context-aware machine translation models are designed to leverage contextual information, but often fail to do so. As a result, they inaccurately disambiguate pronouns and polysemous words that require context for resolution. In this paper, we ask several questions: What contexts do human translators use to resolve ambiguous words? Are models paying large amounts of attention to the same context? What if we explicitly train them to do so? To answer these questions, we introduce SCAT (Supporting Context for Ambiguous Translations), a new English-French dataset comprising supporting context words for 14K translations that professional translators found useful for pronoun disambiguation. Using SCAT, we perform an in-depth analysis of the context used to disambiguate, examining positional and lexical characteristics of the supporting words. Furthermore, we measure the degree of alignment between the model’s attention scores and the supporting context from SCAT, and apply a guided attention strategy to encourage agreement between the two.

Attention mechanisms are ubiquitous components in neural architectures applied to natural language processing. In addition to yielding gains in predictive accuracy, attention weights are often claimed to confer interpretability, purportedly useful both for providing insights to practitioners and for explaining why a model makes its decisions to stakeholders. We call the latter use of attention mechanisms into question by demonstrating a simple method for training models to produce deceptive attention masks. Our method diminishes the total weight assigned to designated impermissible tokens, even when the models can be shown to nevertheless rely on these features to drive predictions. Across multiple models and tasks, our approach manipulates attention weights while paying surprisingly little cost in accuracy. Through a human study, we show that our manipulated attention-based explanations deceive people into thinking that predictions from a model biased against gender minorities do not rely on the gender. Consequently, our results cast doubt on attention’s reliability as a tool for auditing algorithms in the context of fairness and accountability.

For many prediction tasks, stakeholders desire not only predictions but also supporting evidence that a human can use to verify its correctness. However, in practice, evidence annotations may only be available for a minority of training examples (if available at all). In this paper, we propose new methods to combine few evidence annotations (strong semi-supervision) with abundant document-level labels (weak supervision) for the task of evidence extraction. Evaluating on two classification tasks that feature evidence annotations, we find that our methods outperform baselines adapted from the interpretability literature to our task. Our approach yields gains with as few as hundred evidence annotations.

Pooling-based recurrent neural architectures consistently outperform their counterparts without pooling on sequence classification tasks. However, the reasons for their enhanced performance are largely unexamined. In this work, we examine three commonly used pooling techniques (mean-pooling, max-pooling, and attention, and propose *max-attention*, a novel variant that captures interactions among predictive tokens in a sentence. Using novel experiments, we demonstrate that pooling architectures substantially differ from their non-pooling equivalents in their learning ability and positional biases: (i) pooling facilitates better gradient flow than BiLSTMs in initial training epochs, and (ii) BiLSTMs are biased towards tokens at the beginning and end of the input, whereas pooling alleviates this bias. Consequently, we find that pooling yields large gains in low resource scenarios, and instances when salient words lie towards the middle of the input. Across several text classification tasks, we find max-attention to frequently outperform other pooling techniques.

We introduce NeuSpell, an open-source toolkit for spelling correction in English. Our toolkit comprises ten different models, and benchmarks them on naturally occurring misspellings from multiple sources. We find that many systems do not adequately leverage the context around the misspelt token. To remedy this, (i) we train neural models using spelling errors in context, synthetically constructed by reverse engineering isolated misspellings; and (ii) use richer representations of the context. By training on our synthetic examples, correction rates improve by 9% (absolute) compared to the case when models are trained on randomly sampled character perturbations. Using richer contextual representations boosts the correction rate by another 3%. Our toolkit enables practitioners to use our proposed and existing spelling correction systems, both via a simple unified command line, as well as a web interface. Among many potential applications, we demonstrate the utility of our spell-checkers in combating adversarial misspellings. The toolkit can be accessed at neuspell.github.io.

To combat adversarial spelling mistakes, we propose placing a word recognition model in front of the downstream classifier. Our word recognition models build upon the RNN semi-character architecture, introducing several new backoff strategies for handling rare and unseen words. Trained to recognize words corrupted by random adds, drops, swaps, and keyboard mistakes, our method achieves 32% relative (and 3.3% absolute) error reduction over the vanilla semi-character model. Notably, our pipeline confers robustness on the downstream classifier, outperforming both adversarial training and off-the-shelf spell checkers. Against a BERT model fine-tuned for sentiment analysis, a single adversarially-chosen character attack lowers accuracy from 90.3% to 45.8%. Our defense restores accuracy to 75%. Surprisingly, better word recognition does not always entail greater robustness. Our analysis reveals that robustness also depends upon a quantity that we denote the sensitivity.

In this paper, we describe compare-mt, a tool for holistic analysis and comparison of the results of systems for language generation tasks such as machine translation. The main goal of the tool is to give the user a high-level and coherent view of the salient differences between systems that can then be used to guide further analysis or system improvement. It implements a number of tools to do so, such as analysis of accuracy of generation of particular types of words, bucketed histograms of sentence accuracies or counts based on salient characteristics, and extraction of characteristic n-grams for each system. It also has a number of advanced features such as use of linguistic labels, source side data, or comparison of log likelihoods for probabilistic models, and also aims to be easily extensible by users to new types of analysis. compare-mt is a pure-Python open source package, that has already proven useful to generate analyses that have been used in our published papers. Demo Video: https://youtu.be/NyJEQT7t2CA

Recent success of deep learning models for the task of extractive Question Answering (QA) is hinged on the availability of large annotated corpora. However, large domain specific annotated corpora are limited and expensive to construct. In this work, we envision a system where the end user specifies a set of base documents and only a few labelled examples. Our system exploits the document structure to create cloze-style questions from these base documents; pre-trains a powerful neural network on the cloze style questions; and further fine-tunes the model on the labeled examples. We evaluate our proposed system across three diverse datasets from different domains, and find it to be highly effective with very little labeled data. We attain more than 50% F1 score on SQuAD and TriviaQA with less than a thousand labelled examples. We are also releasing a set of 3.2M cloze-style questions for practitioners to use while building QA systems.