To create models that are robust across a wide range of test inputs, training datasets should include diverse examples that span numerous phenomena. Dynamic adversarial data collection (DADC), where annotators craft examples that challenge continually improving models, holds promise as an approach for generating such diverse training sets. Prior work has shown that running DADC over 1-3 rounds can help models fix some error types, but it does not necessarily lead to better generalization beyond adversarial test data. We argue that running DADC over many rounds maximizes its training-time benefits, as the different rounds can together cover many of the task-relevant phenomena. We present the first study of longer-term DADC, where we collect 20 rounds of NLI examples for a small set of premise paragraphs, with both adversarial and non-adversarial approaches. Models trained on DADC examples make 26% fewer errors on our expert-curated test set compared to models trained on non-adversarial data. Our analysis shows that DADC yields examples that are more difficult, more lexically and syntactically diverse, and contain fewer annotation artifacts compared to non-adversarial examples.
This paper proposes a pre-training objective based on question answering (QA) for learning general-purpose contextual representations, motivated by the intuition that the representation of a phrase in a passage should encode all questions that the phrase can answer in context. We accomplish this goal by training a bi-encoder QA model, which independently encodes passages and questions, to match the predictions of a more accurate cross-encoder model on 80 million synthesized QA pairs. By encoding QA-relevant information, the bi-encoder's token-level representations are useful for non-QA downstream tasks without extensive (or in some cases, any) fine-tuning. We show large improvements over both RoBERTa-large and previous state-of-the-art results on zero-shot and few-shot paraphrase detection on four datasets, few-shot named entity recognition on two datasets, and zero-shot sentiment analysis on three datasets.
We release a new benchmark for lexical substitution, the task of finding appropriate substitutes for a target word in a context. To assist humans with writing, lexical substitution systems can suggest words that humans cannot easily think of. However, existing benchmarks depend on human recall as the only source of data, and therefore lack coverage of the substitutes that would be most helpful to humans. Furthermore, annotators often provide substitutes of low quality, which are not actually appropriate in the given context. We collect higher-coverage and higher-quality data by framing lexical substitution as a classification problem, guided by the intuition that it is easier for humans to judge the appropriateness of candidate substitutes than conjure them from memory. To this end, we use a context-free thesaurus to produce candidates and rely on human judgement to determine contextual appropriateness. Compared to the previous largest benchmark, our Swords benchmark has 4.1x more substitutes per target word for the same level of quality, and its substitutes are 1.5x more appropriate (based on human judgement) for the same number of substitutes.
Estimating the expected output quality of generation systems is central to NLG. This paper qualifies the notion that automatic metrics are not as good as humans in estimating system-level quality. Statistically, humans are unbiased, high variance estimators, while metrics are biased, low variance estimators. We compare these estimators by their error in pairwise prediction (which generation system is better?) using the bootstrap. Measuring this error is complicated: predictions are evaluated against noisy, human predicted labels instead of the ground truth, and metric predictions fluctuate based on the test sets they were calculated on. By applying a bias-variance-noise decomposition, we adjust this error to a noise-free, infinite test set setting. Our analysis compares the adjusted error of metrics to humans and a derived, perfect segment-level annotator, both of which are unbiased estimators dependent on the number of judgments collected. In MT, we identify two settings where metrics outperform humans due to a statistical advantage in variance: when the number of human judgments used is small, and when the quality difference between compared systems is small. The data and code to reproduce our analyses are available at https://github.com/johntzwei/metric-statistical-advantage .
We introduce Dynaboard, an evaluation-as-a-service framework for hosting benchmarks and conducting holistic model comparison, integrated with the Dynabench platform. Our platform evaluates NLP models directly instead of relying on self-reported metrics or predictions on a single dataset. Under this paradigm, models are submitted to be evaluated in the cloud, circumventing the issues of reproducibility, accessibility, and backwards compatibility that often hinder benchmarking in NLP. This allows users to interact with uploaded models in real time to assess their quality, and permits the collection of additional metrics such as memory use, throughput, and robustness, which -- despite their importance to practitioners -- have traditionally been absent from leaderboards. On each task, models are ranked according to the Dynascore, a novel utility-based aggregation of these statistics, which users can customize to better reflect their preferences, placing more/less weight on a particular axis of evaluation or dataset. As state-of-the-art NLP models push the limits of traditional benchmarks, Dynaboard offers a standardized solution for a more diverse and comprehensive evaluation of model quality.
Despite the availability of very large datasets and pretrained models, state-of-the-art question answering models remain susceptible to a variety of adversarial attacks and are still far from obtaining human-level language understanding. One proposed way forward is dynamic adversarial data collection, in which a human annotator attempts to create examples for which a model-in-the-loop fails. However, this approach comes at a higher cost per sample and slower pace of annotation, as model-adversarial data requires more annotator effort to generate. In this work, we investigate several answer selection, question generation, and filtering methods that form a synthetic adversarial data generation pipeline that takes human-generated adversarial samples and unannotated text to create synthetic question-answer pairs. Models trained on both synthetic and human-generated data outperform models not trained on synthetic adversarial data, and obtain state-of-the-art results on the AdversarialQA dataset with overall performance gains of 3.7F1. Furthermore, we find that training on the synthetic adversarial data improves model generalisation across domains for non-adversarial data, demonstrating gains on 9 of the 12 datasets for MRQA. Lastly, we find that our models become considerably more difficult to beat by human adversaries, with a drop in macro-averaged validated model error rate from 17.6% to 8.8% when compared to non-augmented models.
A possible explanation for the impressive performance of masked language model (MLM) pre-training is that such models have learned to represent the syntactic structures prevalent in classical NLP pipelines. In this paper, we propose a different explanation: MLMs succeed on downstream tasks almost entirely due to their ability to model higher-order word co-occurrence statistics. To demonstrate this, we pre-train MLMs on sentences with randomly shuffled word order, and show that these models still achieve high accuracy after fine-tuning on many downstream tasks -- including on tasks specifically designed to be challenging for models that ignore word order. Our models perform surprisingly well according to some parametric syntactic probes, indicating possible deficiencies in how we test representations for syntactic information. Overall, our results show that purely distributional information largely explains the success of pre-training, and underscore the importance of curating challenging evaluation datasets that require deeper linguistic knowledge.
We introduce Dynabench, an open-source platform for dynamic dataset creation and model benchmarking. Dynabench runs in a web browser and supports human-and-model-in-the-loop dataset creation: annotators seek to create examples that a target model will misclassify, but that another person will not. In this paper, we argue that Dynabench addresses a critical need in our community: contemporary models quickly achieve outstanding performance on benchmark tasks but nonetheless fail on simple challenge examples and falter in real-world scenarios. With Dynabench, dataset creation, model development, and model assessment can directly inform each other, leading to more robust and informative benchmarks. We report on four initial NLP tasks, illustrating these concepts and highlighting the promise of the platform, and address potential objections to dynamic benchmarking as a new standard for the field.
Datasets are not only resources for training accurate, deployable systems, but are also benchmarks for developing new modeling approaches. While large, natural datasets are necessary for training accurate systems, are they necessary for driving modeling innovation? For example, while the popular SQuAD question answering benchmark has driven the development of new modeling approaches, could synthetic or smaller benchmarks have led to similar innovations? This counterfactual question is impossible to answer, but we can study a necessary condition: the ability for a benchmark to recapitulate findings made on SQuAD. We conduct a retrospective study of 20 SQuAD modeling approaches, investigating how well 32 existing and synthesized benchmarks concur with SQuAD -- i.e., do they rank the approaches similarly? We carefully construct small, targeted synthetic benchmarks that do not resemble natural language, yet have high concurrence with SQuAD, demonstrating that naturalness and size are not necessary for reflecting historical modeling improvements on SQuAD. Our results raise the intriguing possibility that small and carefully designed synthetic benchmarks may be useful for driving the development of new modeling approaches.