Large language models (LM) generate remarkably fluent text and can be efficiently adapted across NLP tasks. Measuring and guaranteeing the quality of generated text in terms of safety is imperative for deploying LMs in the real world; to this end, prior work often relies on automatic evaluation of LM toxicity. We critically discuss this approach, evaluate several toxicity mitigation strategies with respect to both automatic and human evaluation, and analyze consequences of toxicity mitigation in terms of model bias and LM quality. We demonstrate that while basic intervention strategies can effectively optimize previously established automatic metrics on the REALTOXICITYPROMPTS dataset, this comes at the cost of reduced LM coverage for both texts about, and dialects of, marginalized groups. Additionally, we find that human raters often disagree with high automatic toxicity scores after strong toxicity reduction interventions—highlighting further the nuances involved in careful evaluation of LM toxicity.

Innovations in annotation methodology have been a catalyst for Reading Comprehension (RC) datasets and models. One recent trend to challenge current RC models is to involve a model in the annotation process: Humans create questions adversarially, such that the model fails to answer them correctly. In this work we investigate this annotation methodology and apply it in three different settings, collecting a total of 36,000 samples with progressively stronger models in the annotation loop. This allows us to explore questions such as the reproducibility of the adversarial effect, transfer from data collected with varying model-in-the-loop strengths, and generalization to data collected without a model. We find that training on adversarially collected samples leads to strong generalization to non-adversarially collected datasets, yet with progressive performance deterioration with increasingly stronger models-in-the-loop. Furthermore, we find that stronger models can still learn from datasets collected with substantially weaker models-in-the-loop. When trained on data collected with a BiDAF model in the loop, RoBERTa achieves 39.9F1 on questions that it cannot answer when trained on SQuAD—only marginally lower than when trained on data collected using RoBERTa itself (41.0F1).

Advances in language modeling architectures and the availability of large text corpora have driven progress in automatic text generation. While this results in models capable of generating coherent texts, it also prompts models to internalize social biases present in the training corpus. This paper aims to quantify and reduce a particular type of bias exhibited by language models: bias in the sentiment of generated text. Given a conditioning context (e.g., a writing prompt) and a language model, we analyze if (and how) the sentiment of the generated text is affected by changes in values of sensitive attributes (e.g., country names, occupations, genders) in the conditioning context using a form of counterfactual evaluation. We quantify sentiment bias by adopting individual and group fairness metrics from the fair machine learning literature, and demonstrate that large-scale models trained on two different corpora (news articles, and Wikipedia) exhibit considerable levels of bias. We then propose embedding and sentiment prediction-derived regularization on the language model’s latent representations. The regularizations improve fairness metrics while retaining comparable levels of perplexity and semantic similarity.

Current reading comprehension methods generalise well to in-distribution test sets, yet perform poorly on adversarially selected data. Prior work on adversarial inputs typically studies model oversensitivity: semantically invariant text perturbations that cause a model’s prediction to change. Here we focus on the complementary problem: excessive prediction undersensitivity, where input text is meaningfully changed but the model’s prediction does not, even though it should. We formulate 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. We demonstrate that models trained on both SQuAD2.0 and NewsQA are vulnerable to this attack, and then investigate data augmentation and adversarial training as defences. Both substantially decrease adversarial vulnerability, which generalises to held-out data and held-out attack spaces. Addressing undersensitivity furthermore improves model robustness on the previously introduced ADDSENT and ADDONESENT datasets, and models generalise better when facing train / evaluation distribution mismatch: they are less prone to overly rely on shallow predictive cues present only in the training set, and outperform a conventional model by as much as 10.9% F1.

Neural networks are part of many contemporary NLP systems, yet their empirical successes come at the price of vulnerability to adversarial attacks. Previous work has used adversarial training and data augmentation to partially mitigate such brittleness, but these are unlikely to find worst-case adversaries due to the complexity of the search space arising from discrete text perturbations. In this work, we approach the problem from the opposite direction: to formally verify a system’s robustness against a predefined class of adversarial attacks. We study text classification under synonym replacements or character flip perturbations. We propose modeling these input perturbations as a simplex and then using Interval Bound Propagation – a formal model verification method. We modify the conventional log-likelihood training objective to train models that can be efficiently verified, which would otherwise come with exponential search complexity. The resulting models show only 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.

Many Machine Reading and Natural Language Understanding tasks require reading supporting text in order to answer questions. For example, in Question Answering, the supporting text can be newswire or Wikipedia articles; in Natural Language Inference, premises can be seen as the supporting text and hypotheses as questions. Providing a set of useful primitives operating in a single framework of related tasks would allow for expressive modelling, and easier model comparison and replication. To that end, we present Jack the Reader (JACK), a framework for Machine Reading that allows for quick model prototyping by component reuse, evaluation of new models on existing datasets as well as integrating new datasets and applying them on a growing set of implemented baseline models. JACK is currently supporting (but not limited to) three tasks: Question Answering, Natural Language Inference, and Link Prediction. It is developed with the aim of increasing research efficiency and code reuse.

Most Reading Comprehension methods limit themselves to queries which can be answered using a single sentence, paragraph, or document. Enabling models to combine disjoint pieces of textual evidence would extend the scope of machine comprehension methods, but currently no resources exist to train and test this capability. We propose a novel task to encourage the development of models for text understanding across multiple documents and to investigate the limits of existing methods. In our task, a model learns to seek and combine evidence — effectively performing multihop, alias multi-step, inference. We devise a methodology to produce datasets for this task, given a collection of query-answer pairs and thematically linked documents. Two datasets from different domains are induced, and we identify potential pitfalls and devise circumvention strategies. We evaluate two previously proposed competitive models and find that one can integrate information across documents. However, both models struggle to select relevant information; and providing documents guaranteed to be relevant greatly improves their performance. While the models outperform several strong baselines, their best accuracy reaches 54.5% on an annotated test set, compared to human performance at 85.0%, leaving ample room for improvement.

In this paper we describe our 2nd place FEVER shared-task system that achieved a FEVER score of 62.52% on the provisional test set (without additional human evaluation), and 65.41% on the development set. Our system is a four stage model consisting of document retrieval, sentence retrieval, natural language inference and aggregation. Retrieval is performed leveraging task-specific features, and then a natural language inference model takes each of the retrieved sentences paired with the claimed fact. The resulting predictions are aggregated across retrieved sentences with a Multi-Layer Perceptron, and re-ranked corresponding to the final prediction.

We present a novel method for obtaining high-quality, domain-targeted multiple choice questions from crowd workers. Generating these questions can be difficult without trading away originality, relevance or diversity in the answer options. Our method addresses these problems by leveraging a large corpus of domain-specific text and a small set of existing questions. It produces model suggestions for document selection and answer distractor choice which aid the human question generation process. With this method we have assembled SciQ, a dataset of 13.7K multiple choice science exam questions. We demonstrate that the method produces in-domain questions by providing an analysis of this new dataset and by showing that humans cannot distinguish the crowdsourced questions from original questions. When using SciQ as additional training data to existing questions, we observe accuracy improvements on real science exams.