Large language models are able to perform a task by conditioning on a few input-output demonstrations - a paradigm known as in-context learning. We show that language models can explicitly infer an underlying task from a few demonstrations by prompting them to generate a natural language instruction that fits the examples. To explore this ability, we introduce the instruction induction challenge, compile a dataset consisting of 24 tasks, and define a novel evaluation metric based on executing the generated instruction. We discover that, to a large extent, the ability to generate instructions does indeed emerge when using a model that is both large enough and aligned to follow instructions; InstructGPT achieves 65.7% of human performance in our execution-based metric, while the original GPT-3 model reaches only 9.8% of human performance. This surprising result suggests that instruction induction might be a viable learning paradigm in and of itself, where instead of fitting a set of latent continuous parameters to the data, one searches for the best description in the natural language hypothesis space.

Naturally occurring information-seeking questions often contain questionable assumptions—assumptions that are false or unverifiable. Questions containing questionable assumptions are challenging because they require a distinct answer strategy that deviates from typical answers for information-seeking questions. For instance, the question “When did Marie Curie discover Uranium?” cannot be answered as a typical “when” question without addressing the false assumption “Marie Curie discovered Uranium”. In this work, we propose (QA)2 (Question Answering with Questionable Assumptions), an open-domain evaluation dataset consisting of naturally occurring search engine queries that may or may not contain questionable assumptions. To be successful on (QA)2, systems must be able to detect questionable assumptions and also be able to produce adequate responses for both typical information-seeking questions and ones with questionable assumptions. Through human rater acceptability on end-to-end QA with (QA)2, we find that current models do struggle with handling questionable assumptions, leaving substantial headroom for progress.

We present the results of the NLP Community Metasurvey. Run from May to June 2022, it elicited opinions on controversial issues, including industry influence in the field, concerns about AGI, and ethics. Our results put concrete numbers to several controversies: For example, respondents are split in half on the importance of artificial general intelligence, whether language models understand language, and the necessity of linguistic structure and inductive bias for solving NLP problems. In addition, the survey posed meta-questions, asking respondents to predict the distribution of survey responses. This allows us to uncover false sociological beliefs where the community’s predictions don’t match reality. Among other results, we find that the community greatly overestimates its own belief in the usefulness of benchmarks and the potential for scaling to solve real-world problems, while underestimating its belief in the importance of linguistic structure, inductive bias, and interdisciplinary science.

A number of recent benchmarks seek to assess how well models handle natural language negation. However, these benchmarks lack the controlled example paradigms that would allow us to infer whether a model had truly learned how negation morphemes semantically scope. To fill these analytical gaps, we present the Scoped Negation NLI (ScoNe-NLI) benchmark, which contains contrast sets of six examples with up to two negations where either zero, one, or both negative morphemes affect the NLI label. We use ScoNe-NLI to assess fine-tuning and in-context learning strategies. We find that RoBERTa and DeBERTa models solve ScoNe-NLI after many shot fine-tuning. For in-context learning, we test the latest InstructGPT models and find that most prompt strategies are not successful, including those using step-by-step reasoning. To better understand this result, we extend ScoNe with ScoNe-NLG, a sentence completion test set that embeds negation reasoning in short narratives. Here, InstructGPT is successful, which reveals the model can correctly reason about negation, but struggles to do so on NLI examples outside of its core pretraining regime.

As language models (LMs) scale, they develop many novel behaviors, good and bad, exacerbating the need to evaluate how they behave. Prior work creates evaluations with crowdwork (which is time-consuming and expensive) or existing data sources (which are not always available). Here, we automatically generate evaluations with LMs. We explore approaches with varying amounts of human effort, from instructing LMs to write yes/no questions to making complex Winogender schemas with multiple stages of LM-based generation and filtering. Crowdworkers rate the examples as highly relevant and agree with 90-100% of labels, sometimes more so than corresponding human-written datasets. We generate 154 datasets and discover new cases of inverse scaling where LMs get worse with size. Larger LMs repeat back a dialog user’s preferred answer (“sycophancy”) and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down. Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.

To enable building and testing models on long-document comprehension, we introduce QuALITY, a multiple-choice QA dataset with context passages in English that have an average length of about 5,000 tokens, much longer than typical current models can process. Unlike in prior work with passages, our questions are written and validated by contributors who have read the entire passage, rather than relying on summaries or excerpts. In addition, only half of the questions are answerable by annotators working under tight time constraints, indicating that skimming and simple search are not enough to consistently perform well. Our baseline models perform poorly on this task (55.4%) and significantly lag behind human performance (93.5%).

For a natural language understanding benchmark to be useful in research, it has to consist of examples that are diverse and difficult enough to discriminate among current and near-future state-of-the-art systems. However, we do not yet know how best to select text sources to collect a variety of challenging examples. In this study, we crowdsource multiple-choice reading comprehension questions for passages taken from seven qualitatively distinct sources, analyzing what attributes of passages contribute to the difficulty and question types of the collected examples. To our surprise, we find that passage source, length, and readability measures do not significantly affect question difficulty. Through our manual annotation of seven reasoning types, we observe several trends between passage sources and reasoning types, e.g., logical reasoning is more often required in questions written for technical passages. These results suggest that when creating a new benchmark dataset, selecting a diverse set of passages can help ensure a diverse range of question types, but that passage difficulty need not be a priority.

Researchers in NLP often frame and discuss research results in ways that serve to deemphasize the field’s successes, often in response to the field’s widespread hype. Though well-meaning, this has yielded many misleading or false claims about the limits of our best technology. This is a problem, and it may be more serious than it looks: It harms our credibility in ways that can make it harder to mitigate present-day harms, like those involving biased systems for content moderation or resume screening. It also limits our ability to prepare for the potentially enormous impacts of more distant future advances. This paper urges researchers to be careful about these claims and suggests some research directions and communication strategies that will make it easier to avoid or rebut them.

Crowdsourcing platforms are often used to collect datasets for training machine learning models, despite higher levels of inaccurate labeling compared to expert labeling. There are two common strategies to manage the impact of such noise: The first involves aggregating redundant annotations, but comes at the expense of labeling substantially fewer examples. Secondly, prior works have also considered using the entire annotation budget to label as many examples as possible and subsequently apply denoising algorithms to implicitly clean the dataset. We find a middle ground and propose an approach which reserves a fraction of annotations to explicitly clean up highly probable error samples to optimize the annotation process. In particular, we allocate a large portion of the labeling budget to form an initial dataset used to train a model. This model is then used to identify specific examples that appear most likely to be incorrect, which we spend the remaining budget to relabel. Experiments across three model variations and four natural language processing tasks show our approach outperforms or matches both label aggregation and advanced denoising methods designed to handle noisy labels when allocated the same finite annotation budget.

Summarization datasets are often assembled either by scraping naturally occurring public-domain summaries—which are nearly always in difficult-to-work-with technical domains—or by using approximate heuristics to extract them from everyday text—which frequently yields unfaithful summaries. In this work, we turn to a slower but more straightforward approach to developing summarization benchmark data: We hire highly-qualified contractors to read stories and write original summaries from scratch. To amortize reading time, we collect five summaries per document, with the first giving an overview and the subsequent four addressing specific questions. We use this protocol to collect SQuALITY, a dataset of question-focused summaries built on the same public-domain short stories as the multiple-choice dataset QuALITY (Pang et al., 2021). Experiments with state-of-the-art summarization systems show that our dataset is challenging and that existing automatic evaluation metrics are weak indicators of quality.

Pre-trained language models (PLMs) have outperformed other NLP models on a wide range of tasks. Opting for a more thorough understanding of their capabilities and inner workings, researchers have established the extend to which they capture lower-level knowledge like grammaticality, and mid-level semantic knowledge like factual understanding. However, there is still little understanding of their knowledge of higher-level aspects of language. In particular, despite the importance of sociodemographic aspects in shaping our language, the questions of whether, where, and how PLMs encode these aspects, e.g., gender or age, is still unexplored. We address this research gap by probing the sociodemographic knowledge of different single-GPU PLMs on multiple English data sets via traditional classifier probing and information-theoretic minimum description length probing. Our results show that PLMs do encode these sociodemographics, and that this knowledge is sometimes spread across the layers of some of the tested PLMs. We further conduct a multilingual analysis and investigate the effect of supplementary training to further explore to what extent, where, and with what amount of pre-training data the knowledge is encoded. Our overall results indicate that sociodemographic knowledge is still a major challenge for NLP. PLMs require large amounts of pre-training data to acquire the knowledge and models that excel in general language understanding do not seem to own more knowledge about these aspects.

It is well documented that NLP models learn social biases, but little work has been done on how these biases manifest in model outputs for applied tasks like question answering (QA). We introduce the Bias Benchmark for QA (BBQ), a dataset of question-sets constructed by the authors that highlight attested social biases against people belonging to protected classes along nine social dimensions relevant for U.S. English-speaking contexts. Our task evaluate model responses at two levels: (i) given an under-informative context, we test how strongly responses reflect social biases, and (ii) given an adequately informative context, we test whether the model’s biases override a correct answer choice. We find that models often rely on stereotypes when the context is under-informative, meaning the model’s outputs consistently reproduce harmful biases in this setting. Though models are more accurate when the context provides an informative answer, they still rely on stereotypes and average up to 3.4 percentage points higher accuracy when the correct answer aligns with a social bias than when it conflicts, with this difference widening to over 5 points on examples targeting gender for most models tested.

Current QA systems can generate reasonable-sounding yet false answers without explanation or evidence for the generated answer, which is especially problematic when humans cannot readily check the model’s answers. This presents a challenge for building trust in machine learning systems. We take inspiration from real-world situations where difficult questions are answered by considering opposing sides (see Irving et al., 2018). For multiple-choice QA examples, we build a dataset of single arguments for both a correct and incorrect answer option in a debate-style set-up as an initial step in training models to produce explanations for two candidate answers. We use long contexts—humans familiar with the context write convincing explanations for pre-selected correct and incorrect answers, and we test if those explanations allow humans who have not read the full context to more accurately determine the correct answer. We do not find that explanations in our set-up improve human accuracy, but a baseline condition shows that providing human-selected text snippets does improve accuracy. We use these findings to suggest ways of improving the debate set up for future data collection efforts.

Large language models increasingly saturate existing task benchmarks, in some cases outperforming humans, leaving little headroom with which to measure further progress. Adversarial dataset creation, which builds datasets using examples that a target system outputs incorrect predictions for, has been proposed as a strategy to construct more challenging datasets, avoiding the more serious challenge of building more precise benchmarks by conventional means. In this work, we study the impact of applying three common approaches for adversarial dataset creation: (1) filtering out easy examples (AFLite), (2) perturbing examples (TextFooler), and (3) model-in-the-loop data collection (ANLI and AdversarialQA), across 18 different adversary models. We find that all three methods can produce more challenging datasets, with stronger adversary models lowering the performance of evaluated models more. However, the resulting ranking of the evaluated models can also be unstable and highly sensitive to the choice of adversary model. Moreover, we find that AFLite oversamples examples with low annotator agreement, meaning that model comparisons hinge on the examples that are most contentious for humans. We recommend that researchers tread carefully when using adversarial methods for building evaluation datasets.

NLP is currently dominated by language models like RoBERTa which are pretrained on billions of words. But what exact knowledge or skills do Transformer LMs learn from large-scale pretraining that they cannot learn from less data? To explore this question, we adopt five styles of evaluation: classifier probing, information-theoretic probing, unsupervised relative acceptability judgments, unsupervised language model knowledge probing, and fine-tuning on NLU tasks. We then draw learning curves that track the growth of these different measures of model ability with respect to pretraining data volume using the MiniBERTas, a group of RoBERTa models pretrained on 1M, 10M, 100M and 1B words. We find that these LMs require only about 10M to 100M words to learn to reliably encode most syntactic and semantic features we test. They need a much larger quantity of data in order to acquire enough commonsense knowledge and other skills required to master typical downstream NLU tasks. The results suggest that, while the ability to encode linguistic features is almost certainly necessary for language understanding, it is likely that other, unidentified, forms of knowledge are the major drivers of recent improvements in language understanding among large pretrained models.

Recent years have seen numerous NLP datasets introduced to evaluate the performance of fine-tuned models on natural language understanding tasks. Recent results from large pretrained models, though, show that many of these datasets are largely saturated and unlikely to be able to detect further progress. What kind of datasets are still effective at discriminating among strong models, and what kind of datasets should we expect to be able to detect future improvements? To measure this uniformly across datasets, we draw on Item Response Theory and evaluate 29 datasets using predictions from 18 pretrained Transformer models on individual test examples. We find that Quoref, HellaSwag, and MC-TACO are best suited for distinguishing among state-of-the-art models, while SNLI, MNLI, and CommitmentBank seem to be saturated for current strong models. We also observe span selection task format, which is used for QA datasets like QAMR or SQuAD2.0, is effective in differentiating between strong and weak models.

Crowdsourcing is widely used to create data for common natural language understanding tasks. Despite the importance of these datasets for measuring and refining model understanding of language, there has been little focus on the crowdsourcing methods used for collecting the datasets. In this paper, we compare the efficacy of interventions that have been proposed in prior work as ways of improving data quality. We use multiple-choice question answering as a testbed and run a randomized trial by assigning crowdworkers to write questions under one of four different data collection protocols. We find that asking workers to write explanations for their examples is an ineffective stand-alone strategy for boosting NLU example difficulty. However, we find that training crowdworkers, and then using an iterative process of collecting data, sending feedback, and qualifying workers based on expert judgments is an effective means of collecting challenging data. But using crowdsourced, instead of expert judgments, to qualify workers and send feedback does not prove to be effective. We observe that the data from the iterative protocol with expert assessments is more challenging by several measures. Notably, the human–model gap on the unanimous agreement portion of this data is, on average, twice as large as the gap for the baseline protocol data.

Evaluation for many natural language understanding (NLU) tasks is broken: Unreliable and biased systems score so highly on standard benchmarks that there is little room for researchers who develop better systems to demonstrate their improvements. The recent trend to abandon IID benchmarks in favor of adversarially-constructed, out-of-distribution test sets ensures that current models will perform poorly, but ultimately only obscures the abilities that we want our benchmarks to measure. In this position paper, we lay out four criteria that we argue NLU benchmarks should meet. We argue most current benchmarks fail at these criteria, and that adversarial data collection does not meaningfully address the causes of these failures. Instead, restoring a healthy evaluation ecosystem will require significant progress in the design of benchmark datasets, the reliability with which they are annotated, their size, and the ways they handle social bias.

Understanding language requires grasping not only the overtly stated content, but also making inferences about things that were left unsaid. These inferences include presuppositions, a phenomenon by which a listener learns about new information through reasoning about what a speaker takes as given. Presuppositions require complex understanding of the lexical and syntactic properties that trigger them as well as the broader conversational context. In this work, we introduce the Naturally-Occurring Presuppositions in English (NOPE) Corpus to investigate the context-sensitivity of 10 different types of presupposition triggers and to evaluate machine learning models’ ability to predict human inferences. We find that most of the triggers we investigate exhibit moderate variability. We further find that transformer-based models draw correct inferences in simple cases involving presuppositions, but they fail to capture the minority of exceptional cases in which human judgments reveal complex interactions between context and triggers.

Despite the success of fine-tuning pretrained language encoders like BERT for downstream natural language understanding (NLU) tasks, it is still poorly understood how neural networks change after fine-tuning. In this work, we use centered kernel alignment (CKA), a method for comparing learned representations, to measure the similarity of representations in task-tuned models across layers. In experiments across twelve NLU tasks, we discover a consistent block diagonal structure in the similarity of representations within fine-tuned RoBERTa and ALBERT models, with strong similarity within clusters of earlier and later layers, but not between them. The similarity of later layer representations implies that later layers only marginally contribute to task performance, and we verify in experiments that the top few layers of fine-tuned Transformers can be discarded without hurting performance, even with no further tuning.

Many crowdsourced NLP datasets contain systematic artifacts that are identified only after data collection is complete. Earlier identification of these issues should make it easier to create high-quality training and evaluation data. We attempt this by evaluating protocols in which expert linguists work ‘in the loop’ during data collection to identify and address these issues by adjusting task instructions and incentives. Using natural language inference as a test case, we compare three data collection protocols: (i) a baseline protocol with no linguist involvement, (ii) a linguist-in-the-loop intervention with iteratively-updated constraints on the writing task, and (iii) an extension that adds direct interaction between linguists and crowdworkers via a chatroom. We find that linguist involvement does not lead to increased accuracy on out-of-domain test sets compared to baseline, and adding a chatroom has no effect on the data. Linguist involvement does, however, lead to more challenging evaluation data and higher accuracy on some challenge sets, demonstrating the benefits of integrating expert analysis during data collection.

Crowdsourcing from non-experts is one of the most common approaches to collecting data and annotations in NLP. Even though it is such a fundamental tool in NLP, crowdsourcing use is largely guided by common practices and the personal experience of researchers. Developing a theory of crowdsourcing use for practical language problems remains an open challenge. However, there are various principles and practices that have proven effective in generating high quality and diverse data. This tutorial exposes NLP researchers to such data collection crowdsourcing methods and principles through a detailed discussion of a diverse set of case studies. The selection of case studies focuses on challenging settings where crowdworkers are asked to write original text or otherwise perform relatively unconstrained work. Through these case studies, we discuss in detail processes that were carefully designed to achieve data with specific properties, for example to require logical inference, grounded reasoning or conversational understanding. Each case study focuses on data collection crowdsourcing protocol details that often receive limited attention in research presentations, for example in conferences, but are critical for research success.

Intermediate-task training—fine-tuning a pretrained model on an intermediate task before fine-tuning again on the target task—often improves model performance substantially on language understanding tasks in monolingual English settings. We investigate whether English intermediate-task training is still helpful on non-English target tasks. Using nine intermediate language-understanding tasks, we evaluate intermediate-task transfer in a zero-shot cross-lingual setting on the XTREME benchmark. We see large improvements from intermediate training on the BUCC and Tatoeba sentence retrieval tasks and moderate improvements on question-answering target tasks. MNLI, SQuAD and HellaSwag achieve the best overall results as intermediate tasks, while multi-task intermediate offers small additional improvements. Using our best intermediate-task models for each target task, we obtain a 5.4 point improvement over XLM-R Large on the XTREME benchmark, setting the state of the art as of June 2020. We also investigate continuing multilingual MLM during intermediate-task training and using machine-translated intermediate-task data, but neither consistently outperforms simply performing English intermediate-task training.

Large-scale natural language inference (NLI) datasets such as SNLI or MNLI have been created by asking crowdworkers to read a premise and write three new hypotheses, one for each possible semantic relationships (entailment, contradiction, and neutral). While this protocol has been used to create useful benchmark data, it remains unclear whether the writing-based annotation protocol is optimal for any purpose, since it has not been evaluated directly. Furthermore, there is ample evidence that crowdworker writing can introduce artifacts in the data. We investigate two alternative protocols which automatically create candidate (premise, hypothesis) pairs for annotators to label. Using these protocols and a writing-based baseline, we collect several new English NLI datasets of over 3k examples each, each using a fixed amount of annotator time, but a varying number of examples to fit that time budget. Our experiments on NLI and transfer learning show negative results: None of the alternative protocols outperforms the baseline in evaluations of generalization within NLI or on transfer to outside target tasks. We conclude that crowdworker writing still the best known option for entailment data, highlighting the need for further data collection work to focus on improving writing-based annotation processes.

While pretrained models such as BERT have shown large gains across natural language understanding tasks, their performance can be improved by further training the model on a data-rich intermediate task, before fine-tuning it on a target task. However, it is still poorly understood when and why intermediate-task training is beneficial for a given target task. To investigate this, we perform a large-scale study on the pretrained RoBERTa model with 110 intermediate-target task combinations. We further evaluate all trained models with 25 probing tasks meant to reveal the specific skills that drive transfer. We observe that intermediate tasks requiring high-level inference and reasoning abilities tend to work best. We also observe that target task performance is strongly correlated with higher-level abilities such as coreference resolution. However, we fail to observe more granular correlations between probing and target task performance, highlighting the need for further work on broad-coverage probing benchmarks. We also observe evidence that the forgetting of knowledge learned during pretraining may limit our analysis, highlighting the need for further work on transfer learning methods in these settings.

We introduce jiant, an open source toolkit for conducting multitask and transfer learning experiments on English NLU tasks. jiant enables modular and configuration driven experimentation with state-of-the-art models and a broad set of tasks for probing, transfer learning, and multitask training experiments. jiant implements over 50 NLU tasks, including all GLUE and SuperGLUE benchmark tasks. We demonstrate that jiant reproduces published performance on a variety of tasks and models, e.g., RoBERTa and BERT.

A growing body of work shows that models exploit annotation artifacts to achieve state-of-the-art performance on standard crowdsourced benchmarks—datasets collected from crowdworkers to create an evaluation task—while still failing on out-of-domain examples for the same task. Recent work has explored the use of counterfactually-augmented data—data built by minimally editing a set of seed examples to yield counterfactual labels—to augment training data associated with these benchmarks and build more robust classifiers that generalize better. However, Khashabi et al. (2020) find that this type of augmentation yields little benefit on reading comprehension tasks when controlling for dataset size and cost of collection. We build upon this work by using English natural language inference data to test model generalization and robustness and find that models trained on a counterfactually-augmented SNLI dataset do not generalize better than unaugmented datasets of similar size and that counterfactual augmentation can hurt performance, yielding models that are less robust to challenge examples. Counterfactual augmentation of natural language understanding data through standard crowdsourcing techniques does not appear to be an effective way of collecting training data and further innovation is required to make this general line of work viable.

Neural unsupervised parsing (UP) models learn to parse without access to syntactic annotations, while being optimized for another task like language modeling. In this work, we propose self-training for neural UP models: we leverage aggregated annotations predicted by copies of our model as supervision for future copies. To be able to use our model’s predictions during training, we extend a recent neural UP architecture, the PRPN (Shen et al., 2018a), such that it can be trained in a semi-supervised fashion. We then add examples with parses predicted by our model to our unlabeled UP training data. Our self-trained model outperforms the PRPN by 8.1% F1 and the previous state of the art by 1.6% F1. In addition, we show that our architecture can also be helpful for semi-supervised parsing in ultra-low-resource settings.

We introduce The Benchmark of Linguistic Minimal Pairs (BLiMP),1 a challenge set for evaluating the linguistic knowledge of language models (LMs) on major grammatical phenomena in English. BLiMP consists of 67 individual datasets, each containing 1,000 minimal pairs—that is, pairs of minimally different sentences that contrast in grammatical acceptability and isolate specific phenomenon in syntax, morphology, or semantics. We generate the data according to linguist-crafted grammar templates, and human aggregate agreement with the labels is 96.4%. We evaluate n-gram, LSTM, and Transformer (GPT-2 and Transformer-XL) LMs by observing whether they assign a higher probability to the acceptable sentence in each minimal pair. We find that state-of-the-art models identify morphological contrasts related to agreement reliably, but they struggle with some subtle semantic and syntactic phenomena, such as negative polarity items and extraction islands.

One reason pretraining on self-supervised linguistic tasks is effective is that it teaches models features that are helpful for language understanding. However, we want pretrained models to learn not only to represent linguistic features, but also to use those features preferentially during fine-turning. With this goal in mind, we introduce a new English-language diagnostic set called MSGS (the Mixed Signals Generalization Set), which consists of 20 ambiguous binary classification tasks that we use to test whether a pretrained model prefers linguistic or surface generalizations during finetuning. We pretrain RoBERTa from scratch on quantities of data ranging from 1M to 1B words and compare their performance on MSGS to the publicly available RoBERTa_BASE. We find that models can learn to represent linguistic features with little pretraining data, but require far more data to learn to prefer linguistic generalizations over surface ones. Eventually, with about 30B words of pretraining data, RoBERTa_BASE does consistently demonstrate a linguistic bias with some regularity. We conclude that while self-supervised pretraining is an effective way to learn helpful inductive biases, there is likely room to improve the rate at which models learn which features matter.

Pretrained language models, especially masked language models (MLMs) have seen success across many NLP tasks. However, there is ample evidence that they use the cultural biases that are undoubtedly present in the corpora they are trained on, implicitly creating harm with biased representations. To measure some forms of social bias in language models against protected demographic groups in the US, we introduce the Crowdsourced Stereotype Pairs benchmark (CrowS-Pairs). CrowS-Pairs has 1508 examples that cover stereotypes dealing with nine types of bias, like race, religion, and age. In CrowS-Pairs a model is presented with two sentences: one that is more stereotyping and another that is less stereotyping. The data focuses on stereotypes about historically disadvantaged groups and contrasts them with advantaged groups. We find that all three of the widely-used MLMs we evaluate substantially favor sentences that express stereotypes in every category in CrowS-Pairs. As work on building less biased models advances, this dataset can be used as a benchmark to evaluate progress.

Natural language inference (NLI) data has proven useful in benchmarking and, especially, as pretraining data for tasks requiring language understanding. However, the crowdsourcing protocol that was used to collect this data has known issues and was not explicitly optimized for either of these purposes, so it is likely far from ideal. We propose four alternative protocols, each aimed at improving either the ease with which annotators can produce sound training examples or the quality and diversity of those examples. Using these alternatives and a fifth baseline protocol, we collect and compare five new 8.5k-example training sets. In evaluations focused on transfer learning applications, our results are solidly negative, with models trained on our baseline dataset yielding good transfer performance to downstream tasks, but none of our four new methods (nor the recent ANLI) showing any improvements over that baseline. In a small silver lining, we observe that all four new protocols, especially those where annotators edit *pre-filled* text boxes, reduce previously observed issues with annotation artifacts.

Performance on the Winograd Schema Challenge (WSC), a respected English commonsense reasoning benchmark, recently rocketed from chance accuracy to 89% on the SuperGLUE leaderboard, with relatively little corroborating evidence of a correspondingly large improvement in reasoning ability. We hypothesize that much of this improvement comes from recent changes in task formalization—the combination of input specification, loss function, and reuse of pretrained parameters—by users of the dataset, rather than improvements in the pretrained model’s reasoning ability. We perform an ablation on two Winograd Schema datasets that interpolates between the formalizations used before and after this surge, and find (i) framing the task as multiple choice improves performance dramatically and (ii)several additional techniques, including the reuse of a pretrained language modeling head, can mitigate the model’s extreme sensitivity to hyperparameters. We urge future benchmark creators to impose additional structure to minimize the impact of formalization decisions on reported results.

We introduce a set of nine challenge tasks that test for the understanding of function words. These tasks are created by structurally mutating sentences from existing datasets to target the comprehension of specific types of function words (e.g., prepositions, wh-words). Using these probing tasks, we explore the effects of various pretraining objectives for sentence encoders (e.g., language modeling, CCG supertagging and natural language inference (NLI)) on the learned representations. Our results show that pretraining on CCG—our most syntactic objective—performs the best on average across our probing tasks, suggesting that syntactic knowledge helps function word comprehension. Language modeling also shows strong performance, supporting its widespread use for pretraining state-of-the-art NLP models. Overall, no pretraining objective dominates across the board, and our function word probing tasks highlight several intuitive differences between pretraining objectives, e.g., that NLI helps the comprehension of negation.

Natural language understanding has recently seen a surge of progress with the use of sentence encoders like ELMo (Peters et al., 2018a) and BERT (Devlin et al., 2019) which are pretrained on variants of language modeling. We conduct the first large-scale systematic study of candidate pretraining tasks, comparing 19 different tasks both as alternatives and complements to language modeling. Our primary results support the use language modeling, especially when combined with pretraining on additional labeled-data tasks. However, our results are mixed across pretraining tasks and show some concerning trends: In ELMo’s pretrain-then-freeze paradigm, random baselines are worryingly strong and results vary strikingly across target tasks. In addition, fine-tuning BERT on an intermediate task often negatively impacts downstream transfer. In a more positive trend, we see modest gains from multitask training, suggesting the development of more sophisticated multitask and transfer learning techniques as an avenue for further research.

The GLUE benchmark (Wang et al., 2019b) is a suite of language understanding tasks which has seen dramatic progress in the past year, with average performance moving from 70.0 at launch to 83.9, state of the art at the time of writing (May 24, 2019). Here, we measure human performance on the benchmark, in order to learn whether significant headroom remains for further progress. We provide a conservative estimate of human performance on the benchmark through crowdsourcing: Our annotators are non-experts who must learn each task from a brief set of instructions and 20 examples. In spite of limited training, these annotators robustly outperform the state of the art on six of the nine GLUE tasks and achieve an average score of 87.1. Given the fast pace of progress however, the headroom we observe is quite limited. To reproduce the data-poor setting that our annotators must learn in, we also train the BERT model (Devlin et al., 2019) in limited-data regimes, and conclude that low-resource sentence classification remains a challenge for modern neural network approaches to text understanding.

This paper investigates the ability of artificial neural networks to judge the grammatical acceptability of a sentence, with the goal of testing their linguistic competence. We introduce the Corpus of Linguistic Acceptability (CoLA), a set of 10,657 English sentences labeled as grammatical or ungrammatical from published linguistics literature. As baselines, we train several recurrent neural network models on acceptability classification, and find that our models outperform unsupervised models by Lau et al. (2016) on CoLA. Error-analysis on specific grammatical phenomena reveals that both Lau et al.’s models and ours learn systematic generalizations like subject-verb-object order. However, all models we test perform far below human level on a wide range of grammatical constructions.

The Word Embedding Association Test shows that GloVe and word2vec word embeddings exhibit human-like implicit biases based on gender, race, and other social constructs (Caliskan et al., 2017). Meanwhile, research on learning reusable text representations has begun to explore sentence-level texts, with some sentence encoders seeing enthusiastic adoption. Accordingly, we extend the Word Embedding Association Test to measure bias in sentence encoders. We then test several sentence encoders, including state-of-the-art methods such as ELMo and BERT, for the social biases studied in prior work and two important biases that are difficult or impossible to test at the word level. We observe mixed results including suspicious patterns of sensitivity that suggest the test’s assumptions may not hold in general. We conclude by proposing directions for future work on measuring bias in sentence encoders.

Many text corpora exhibit socially problematic biases, which can be propagated or amplified in the models trained on such data. For example, doctor cooccurs more frequently with male pronouns than female pronouns. In this study we (i) propose a metric to measure gender bias; (ii) measure bias in a text corpus and the text generated from a recurrent neural network language model trained on the text corpus; (iii) propose a regularization loss term for the language model that minimizes the projection of encoder-trained embeddings onto an embedding subspace that encodes gender; (iv) finally, evaluate efficacy of our proposed method on reducing gender bias. We find this regularization method to be effective in reducing gender bias up to an optimal weight assigned to the loss term, beyond which the model becomes unstable as the perplexity increases. We replicate this study on three training corpora—Penn Treebank, WikiText-2, and CNN/Daily Mail—resulting in similar conclusions.

This tutorial discusses cutting-edge research on NLI, including recent advance on dataset development, cutting-edge deep learning models, and highlights from recent research on using NLI to understand capabilities and limits of deep learning models for language understanding and reasoning.

Though state-of-the-art sentence representation models can perform tasks requiring significant knowledge of grammar, it is an open question how best to evaluate their grammatical knowledge. We explore five experimental methods inspired by prior work evaluating pretrained sentence representation models. We use a single linguistic phenomenon, negative polarity item (NPI) licensing, as a case study for our experiments. NPIs like any are grammatical only if they appear in a licensing environment like negation (Sue doesn’t have any cats vs. *Sue has any cats). This phenomenon is challenging because of the variety of NPI licensing environments that exist. We introduce an artificially generated dataset that manipulates key features of NPI licensing for the experiments. We find that BERT has significant knowledge of these features, but its success varies widely across different experimental methods. We conclude that a variety of methods is necessary to reveal all relevant aspects of a model’s grammatical knowledge in a given domain.

Development sets are impractical to obtain for real low-resource languages, since using all available data for training is often more effective. However, development sets are widely used in research papers that purport to deal with low-resource natural language processing (NLP). Here, we aim to answer the following questions: Does using a development set for early stopping in the low-resource setting influence results as compared to a more realistic alternative, where the number of training epochs is tuned on development languages? And does it lead to overestimation or underestimation of performance? We repeat multiple experiments from recent work on neural models for low-resource NLP and compare results for models obtained by training with and without development sets. On average over languages, absolute accuracy differs by up to 1.4%. However, for some languages and tasks, differences are as big as 18.0% accuracy. Our results highlight the importance of realistic experimental setups in the publication of low-resource NLP research results.

Recently, neural network models which automatically infer syntactic structure from raw text have started to achieve promising results. However, earlier work on unsupervised parsing shows large performance differences between non-neural models trained on corpora in different languages, even for comparable amounts of data. With that in mind, we train instances of the PRPN architecture (Shen et al., 2018)—one of these unsupervised neural network parsers—for Arabic, Chinese, English, and German. We find that (i) the model strongly outperforms trivial baselines and, thus, acquires at least some parsing ability for all languages; (ii) good hyperparameter values seem to be universal; (iii) how the model benefits from larger training set sizes depends on the corpus, with the model achieving the largest performance gains when increasing the number of sentences from 2,500 to 12,500 for English. In addition, we show that, by sharing parameters between the related languages German and English, we can improve the model’s unsupervised parsing F1 score by up to 4% in the low-resource setting.

This paper introduces the Multi-Genre Natural Language Inference (MultiNLI) corpus, a dataset designed for use in the development and evaluation of machine learning models for sentence understanding. At 433k examples, this resource is one of the largest corpora available for natural language inference (a.k.a. recognizing textual entailment), improving upon available resources in both its coverage and difficulty. MultiNLI accomplishes this by offering data from ten distinct genres of written and spoken English, making it possible to evaluate systems on nearly the full complexity of the language, while supplying an explicit setting for evaluating cross-genre domain adaptation. In addition, an evaluation using existing machine learning models designed for the Stanford NLI corpus shows that it represents a substantially more difficult task than does that corpus, despite the two showing similar levels of inter-annotator agreement.

Large-scale datasets for natural language inference are created by presenting crowd workers with a sentence (premise), and asking them to generate three new sentences (hypotheses) that it entails, contradicts, or is logically neutral with respect to. We show that, in a significant portion of such data, this protocol leaves clues that make it possible to identify the label by looking only at the hypothesis, without observing the premise. Specifically, we show that a simple text categorization model can correctly classify the hypothesis alone in about 67% of SNLI (Bowman et. al, 2015) and 53% of MultiNLI (Williams et. al, 2017). Our analysis reveals that specific linguistic phenomena such as negation and vagueness are highly correlated with certain inference classes. Our findings suggest that the success of natural language inference models to date has been overestimated, and that the task remains a hard open problem.

Latent tree learning models learn to parse a sentence without syntactic supervision, and use that parse to build the sentence representation. Existing work on such models has shown that, while they perform well on tasks like sentence classification, they do not learn grammars that conform to any plausible semantic or syntactic formalism (Williams et al., 2018a). Studying the parsing ability of such models in natural language can be challenging due to the inherent complexities of natural language, like having several valid parses for a single sentence. In this paper we introduce ListOps, a toy dataset created to study the parsing ability of latent tree models. ListOps sequences are in the style of prefix arithmetic. The dataset is designed to have a single correct parsing strategy that a system needs to learn to succeed at the task. We show that the current leading latent tree models are unable to learn to parse and succeed at ListOps. These models achieve accuracies worse than purely sequential RNNs.

In recent years, there have been amazing advances in deep learning methods for machine reading. In machine reading, the machine reader has to extract the answer from the given ground truth paragraph. Recently, the state-of-the-art machine reading models achieve human level performance in SQuAD which is a reading comprehension-style question answering (QA) task. The success of machine reading has inspired researchers to combine Information Retrieval with machine reading to tackle open-domain QA. However, these systems perform poorly compared to reading comprehension-style QA because it is difficult to retrieve the pieces of paragraphs that contain the answer to the question. In this study, we propose two neural network rankers that assign scores to different passages based on their likelihood of containing the answer to a given question. Additionally, we analyze the relative importance of semantic similarity and word level relevance matching in open-domain QA.

Tree-structured neural network architectures for sentence encoding draw inspiration from the approach to semantic composition generally seen in formal linguistics, and have shown empirical improvements over comparable sequence models by doing so. Moreover, adding multiplicative interaction terms to the composition functions in these models can yield significant further improvements. However, existing compositional approaches that adopt such a powerful composition function scale poorly, with parameter counts exploding as model dimension or vocabulary size grows. We introduce the Lifted Matrix-Space model, which uses a global transformation to map vector word embeddings to matrices, which can then be composed via an operation based on matrix-matrix multiplication. Its composition function effectively transmits a larger number of activations across layers with relatively few model parameters. We evaluate our model on the Stanford NLI corpus, the Multi-Genre NLI corpus, and the Stanford Sentiment Treebank and find that it consistently outperforms TreeLSTM (Tai et al., 2015), the previous best known composition function for tree-structured models.

Recent work on the problem of latent tree learning has made it possible to train neural networks that learn to both parse a sentence and use the resulting parse to interpret the sentence, all without exposure to ground-truth parse trees at training time. Surprisingly, these models often perform better at sentence understanding tasks than models that use parse trees from conventional parsers. This paper aims to investigate what these latent tree learning models learn. We replicate two such models in a shared codebase and find that (i) only one of these models outperforms conventional tree-structured models on sentence classification, (ii) its parsing strategies are not especially consistent across random restarts, (iii) the parses it produces tend to be shallower than standard Penn Treebank (PTB) parses, and (iv) they do not resemble those of PTB or any other semantic or syntactic formalism that the authors are aware of.

To answer the question in machine comprehension (MC) task, the models need to establish the interaction between the question and the context. To tackle the problem that the single-pass model cannot reflect on and correct its answer, we present Ruminating Reader. Ruminating Reader adds a second pass of attention and a novel information fusion component to the Bi-Directional Attention Flow model (BiDAF). We propose novel layer structures that construct a query aware context vector representation and fuse encoding representation with intermediate representation on top of BiDAF model. We show that a multi-hop attention mechanism can be applied to a bi-directional attention structure. In experiments on SQuAD, we find that the Reader outperforms the BiDAF baseline by 2.1 F1 score and 2.7 EM score. Our analysis shows that different hops of the attention have different responsibilities in selecting answers.

Human ability to understand language is general, flexible, and robust. In contrast, most NLU models above the word level are designed for a specific task and struggle with out-of-domain data. If we aspire to develop models with understanding beyond the detection of superficial correspondences between inputs and outputs, then it is critical to develop a unified model that can execute a range of linguistic tasks across different domains. To facilitate research in this direction, we present the General Language Understanding Evaluation (GLUE, gluebenchmark.com): a benchmark of nine diverse NLU tasks, an auxiliary dataset for probing models for understanding of specific linguistic phenomena, and an online platform for evaluating and comparing models. For some benchmark tasks, training data is plentiful, but for others it is limited or does not match the genre of the test set. GLUE thus favors models that can represent linguistic knowledge in a way that facilitates sample-efficient learning and effective knowledge-transfer across tasks. While none of the datasets in GLUE were created from scratch for the benchmark, four of them feature privately-held test data, which is used to ensure that the benchmark is used fairly. We evaluate baselines that use ELMo (Peters et al., 2018), a powerful transfer learning technique, as well as state-of-the-art sentence representation models. The best models still achieve fairly low absolute scores. Analysis with our diagnostic dataset yields similarly weak performance over all phenomena tested, with some exceptions.

Recently, researchers have found that deep LSTMs trained on tasks like machine translation learn substantial syntactic and semantic information about their input sentences, including part-of-speech. These findings begin to shed light on why pretrained representations, like ELMo and CoVe, are so beneficial for neural language understanding models. We still, though, do not yet have a clear understanding of how the choice of pretraining objective affects the type of linguistic information that models learn. With this in mind, we compare four objectives—language modeling, translation, skip-thought, and autoencoding—on their ability to induce syntactic and part-of-speech information, holding constant the quantity and genre of the training data, as well as the LSTM architecture.

Grammar induction is the task of learning syntactic structure without the expert-labeled treebanks (Charniak and Carroll, 1992; Klein and Manning, 2002). Recent work on latent tree learning offers a new family of approaches to this problem by inducing syntactic structure using the supervision from a downstream NLP task (Yogatama et al., 2017; Maillard et al., 2017; Choi et al., 2018). In a recent paper published at ICLR, Shen et al. (2018) introduce such a model and report near state-of-the-art results on the target task of language modeling, and the first strong latent tree learning result on constituency parsing. During the analysis of this model, we discover issues that make the original results hard to trust, including tuning and even training on what is effectively the test set. Here, we analyze the model under different configurations to understand what it learns and to identify the conditions under which it succeeds. We find that this model represents the first empirical success for neural network latent tree learning, and that neural language modeling warrants further study as a setting for grammar induction.

Beam search is a widely used approximate search strategy for neural network decoders, and it generally outperforms simple greedy decoding on tasks like machine translation. However, this improvement comes at substantial computational cost. In this paper, we propose a flexible new method that allows us to reap nearly the full benefits of beam search with nearly no additional computational cost. The method revolves around a small neural network actor that is trained to observe and manipulate the hidden state of a previously-trained decoder. To train this actor network, we introduce the use of a pseudo-parallel corpus built using the output of beam search on a base model, ranked by a target quality metric like BLEU. Our method is inspired by earlier work on this problem, but requires no reinforcement learning, and can be trained reliably on a range of models. Experiments on three parallel corpora and three architectures show that the method yields substantial improvements in translation quality and speed over each base system.

State-of-the-art natural language processing systems rely on supervision in the form of annotated data to learn competent models. These models are generally trained on data in a single language (usually English), and cannot be directly used beyond that language. Since collecting data in every language is not realistic, there has been a growing interest in cross-lingual language understanding (XLU) and low-resource cross-language transfer. In this work, we construct an evaluation set for XLU by extending the development and test sets of the Multi-Genre Natural Language Inference Corpus (MultiNLI) to 14 languages, including low-resource languages such as Swahili and Urdu. We hope that our dataset, dubbed XNLI, will catalyze research in cross-lingual sentence understanding by providing an informative standard evaluation task. In addition, we provide several baselines for multilingual sentence understanding, including two based on machine translation systems, and two that use parallel data to train aligned multilingual bag-of-words and LSTM encoders. We find that XNLI represents a practical and challenging evaluation suite, and that directly translating the test data yields the best performance among available baselines.

A substantial thread of recent work on latent tree learning has attempted to develop neural network models with parse-valued latent variables and train them on non-parsing tasks, in the hope of having them discover interpretable tree structure. In a recent paper, Shen et al. (2018) introduce such a model and report near-state-of-the-art results on the target task of language modeling, and the first strong latent tree learning result on constituency parsing. In an attempt to reproduce these results, we discover issues that make the original results hard to trust, including tuning and even training on what is effectively the test set. Here, we attempt to reproduce these results in a fair experiment and to extend them to two new datasets. We find that the results of this work are robust: All variants of the model under study outperform all latent tree learning baselines, and perform competitively with symbolic grammar induction systems. We find that this model represents the first empirical success for latent tree learning, and that neural network language modeling warrants further study as a setting for grammar induction.

In this paper we propose a neural network model with a novel Sequential Attention layer that extends soft attention by assigning weights to words in an input sequence in a way that takes into account not just how well that word matches a query, but how well surrounding words match. We evaluate this approach on the task of reading comprehension (on the Who did What and CNN datasets) and show that it dramatically improves a strong baseline—the Stanford Reader—and is competitive with the state of the art.

Individuals on social media may reveal themselves to be in various states of crisis (e.g. suicide, self-harm, abuse, or eating disorders). Detecting crisis from social media text automatically and accurately can have profound consequences. However, detecting a general state of crisis without explaining why has limited applications. An explanation in this context is a coherent, concise subset of the text that rationalizes the crisis detection. We explore several methods to detect and explain crisis using a combination of neural and non-neural techniques. We evaluate these techniques on a unique data set obtained from Koko, an anonymous emotional support network available through various messaging applications. We annotate a small subset of the samples labeled with crisis with corresponding explanations. Our best technique significantly outperforms the baseline for detection and explanation.

This paper presents the results of the RepEval 2017 Shared Task, which evaluated neural network sentence representation learning models on the Multi-Genre Natural Language Inference corpus (MultiNLI) recently introduced by Williams et al. (2017). All of the five participating teams beat the bidirectional LSTM (BiLSTM) and continuous bag of words baselines reported in Williams et al. The best single model used stacked BiLSTMs with residual connections to extract sentence features and reached 74.5% accuracy on the genre-matched test set. Surprisingly, the results of the competition were fairly consistent across the genre-matched and genre-mismatched test sets, and across subsets of the test data representing a variety of linguistic phenomena, suggesting that all of the submitted systems learned reasonably domain-independent representations for sentence meaning.

We present a gold standard annotation of syntactic dependencies in the English Web Treebank corpus using the Stanford Dependencies formalism. This resource addresses the lack of a gold standard dependency treebank for English, as well as the limited availability of gold standard syntactic annotations for English informal text genres. We also present experiments on the use of this resource, both for training dependency parsers and for evaluating the quality of different versions of the Stanford Parser, which includes a converter tool to produce dependency annotation from constituency trees. We show that training a dependency parser on a mix of newswire and web data leads to better performance on that type of data without hurting performance on newswire text, and therefore gold standard annotations for non-canonical text can be a valuable resource for parsing. Furthermore, the systematic annotation effort has informed both the SD formalism and its implementation in the Stanford Parser’s dependency converter. In response to the challenges encountered by annotators in the EWT corpus, the formalism has been revised and extended, and the converter has been improved.