Language model pretraining has led to significant performance gains but careful comparison between different approaches is challenging. Training is computationally expensive, often done on private datasets of different sizes, and, as we will show, hyperparameter choices have significant impact on the final results. We present a replication study of BERT pretraining (Devlin et al., 2019) that carefully measures the impact of many key hyperparameters and training data size. We find that BERT was significantly undertrained, and can match or exceed the performance of every model published after it. Our best model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results highlight the importance of previously overlooked design choices, and raise questions about the source of recently reported improvements. We release our models and code.

/pdf/roberta-a-robustly-optimized-bert-pretraining-approach-2rj0bdyhim.pdf

RoBERTa: A Robustly Optimized BERT Pretraining Approach

Machine Learning is the study of methods for programming computers to learn. Computers are applied to a wide range of tasks, and for most of these it is relatively easy for programmers to design and implement the necessary software. However, there are many tasks for which this is difficult or impossible. These can be divided into four general categories. First, there are problems for which there exist no human experts. For example, in modern automated manufacturing facilities, there is a need to predict machine failures before they occur by analyzing sensor readings. Because the machines are new, there are no human experts who can be interviewed by a programmer to provide the knowledge necessary to build a computer system. A machine learning system can study recorded data and subsequent machine failures and learn prediction rules. Second, there are problems where human experts exist, but where they are unable to explain their expertise. This is the case in many perceptual tasks, such as speech recognition, hand-writing recognition, and natural language understanding. Virtually all humans exhibit expert-level abilities on these tasks, but none of them can describe the detailed steps that they follow as they perform them. Fortunately, humans can provide machines with examples of the inputs and correct outputs for these tasks, so machine learning algorithms can learn to map the inputs to the outputs. Third, there are problems where phenomena are changing rapidly. In finance, for example, people would like to predict the future behavior of the stock market, of consumer purchases, or of exchange rates. These behaviors change frequently, so that even if a programmer could construct a good predictive computer program, it would need to be rewritten frequently. A learning program can relieve the programmer of this burden by constantly modifying and tuning a set of learned prediction rules. Fourth, there are applications that need to be customized for each computer user separately. Consider, for example, a program to filter unwanted electronic mail messages. Different users will need different filters. It is unreasonable to expect each user to program his or her own rules, and it is infeasible to provide every user with a software engineer to keep the rules up-to-date. A machine learning system can learn which mail messages the user rejects and maintain the filtering rules automatically. Machine learning addresses many of the same research questions as the fields of statistics, data mining, and psychology, but with differences of emphasis. Statistics focuses on understanding the phenomena that have generated the data, often with the goal of testing different hypotheses about those phenomena. Data mining seeks to find patterns in the data that are understandable by people. Psychological studies of human learning aspire to understand the mechanisms underlying the various learning behaviors exhibited by people (concept learning, skill acquisition, strategy change, etc.).

Machine learning

While the Transformer architecture has become the de-facto standard for natural language processing tasks, its applications to computer vision remain limited. In vision, attention is either applied in conjunction with convolutional networks, or used to replace certain components of convolutional networks while keeping their overall structure in place. We show that this reliance on CNNs is not necessary and a pure transformer applied directly to sequences of image patches can perform very well on image classification tasks. When pre-trained on large amounts of data and transferred to multiple mid-sized or small image recognition benchmarks (ImageNet, CIFAR-100, VTAB, etc.), Vision Transformer (ViT) attains excellent results compared to state-of-the-art convolutional networks while requiring substantially fewer computational resources to train.

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous experimental effort1–4, the structures of around 100,000 unique proteins have been determined5, but this represents a small fraction of the billions of known protein sequences6,7. Structural coverage is bottlenecked by the months to years of painstaking effort required to determine a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the three-dimensional structure that a protein will adopt based solely on its amino acid sequence—the structure prediction component of the ‘protein folding problem’8—has been an important open research problem for more than 50 years9. Despite recent progress10–14, existing methods fall far short of atomic accuracy, especially when no homologous structure is available. Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14)15, demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates physical and biological knowledge about protein structure, leveraging multi-sequence alignments, into the design of the deep learning algorithm. AlphaFold predicts protein structures with an accuracy competitive with experimental structures in the majority of cases using a novel deep learning architecture.

/pdf/highly-accurate-protein-structure-prediction-with-alphafold-2xjbzfwu3n.pdf

Highly accurate protein structure prediction with AlphaFold

Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task. While typically task-agnostic in architecture, this method still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples. By contrast, humans can generally perform a new language task from only a few examples or from simple instructions - something which current NLP systems still largely struggle to do. Here we show that scaling up language models greatly improves task-agnostic, few-shot performance, sometimes even reaching competitiveness with prior state-of-the-art fine-tuning approaches. Specifically, we train GPT-3, an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model, and test its performance in the few-shot setting. For all tasks, GPT-3 is applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model. GPT-3 achieves strong performance on many NLP datasets, including translation, question-answering, and cloze tasks, as well as several tasks that require on-the-fly reasoning or domain adaptation, such as unscrambling words, using a novel word in a sentence, or performing 3-digit arithmetic. At the same time, we also identify some datasets where GPT-3's few-shot learning still struggles, as well as some datasets where GPT-3 faces methodological issues related to training on large web corpora. Finally, we find that GPT-3 can generate samples of news articles which human evaluators have difficulty distinguishing from articles written by humans. We discuss broader societal impacts of this finding and of GPT-3 in general.

/pdf/language-models-are-few-shot-learners-2fg8gvia7m.pdf

Language Models are Few-Shot Learners

We study approaches to improve fine-grained short answer Question Answering models by integrating coarse-grained data annotated for paragraph-level relevance and show that coarsely annotated data can bring significant performance gains. Experiments demonstrate that the standard multi-task learning approach of sharing representations is not the most effective way to leverage coarse-grained annotations. Instead, we can explicitly model the latent fine-grained short answer variables and optimize the marginal log-likelihood directly or use a newly proposed \emph{posterior distillation} learning objective. Since these latent-variable methods have explicit access to the relationship between the fine and coarse tasks, they result in significantly larger improvements from coarse supervision.

Improving Span-based Question Answering Systems with Coarsely Labeled Data.

We study multi-answer retrieval, an under-explored problem that requires retrieving passages to cover multiple distinct answers for a given question. This task requires joint modeling of retrieved passages, as models should not repeatedly retrieve passages containing the same answer at the cost of missing a different valid answer. Prior work focusing on single-answer retrieval is limited as it cannot reason about the set of passages jointly. In this paper, we introduce JPR, a joint passage retrieval model focusing on reranking. To model the joint probability of the retrieved passages, JPR makes use of an autoregressive reranker that selects a sequence of passages, equipped with novel training and decoding algorithms. Compared to prior approaches, JPR achieves significantly better answer coverage on three multi-answer datasets. When combined with downstream question answering, the improved retrieval enables larger answer generation models since they need to consider fewer passages, establishing a new state-of-the-art.

Joint Passage Ranking for Diverse Multi-Answer Retrieval

Tables in Web documents are pervasive and can be directly used to answer many of the queries searched on the Web, motivating their integration in question answering. Very often information presented in tables is succinct and hard to interpret with standard language representations. On the other hand, tables often appear within textual context, such as an article describing the table. Using the information from an article as additional context can potentially enrich table representations. In this work we aim to improve question answering from tables by refining table representations based on information from surrounding text. We also present an effective method to combine text and table-based predictions for question answering from full documents, obtaining significant improvements on the Natural Questions dataset.

Representations for Question Answering from Documents with Tables and Text

Current pre-trained language models have enabled remarkable improvements in 1 downstream tasks, but it remains difficult to distinguish effects of statistical correla- 2 tion from more systematic logical reasoning grounded on understanding of the real 3 world. In this paper we tease these factors apart by leveraging counterfactual con- 4 ditionals , which force language models to predict unusual consequences based on 5 hypothetical propositions. We introduce a set of tests drawn from psycholinguistic 6 experiments, as well as larger-scale controlled datasets, to probe counterfactual 7 predictions from a variety of popular pre-trained language models. We find that 8 models are consistently able to override real-world knowledge in counterfactual 9 scenarios, and that this effect is more robust in case of stronger baseline world 10 knowledge—however, we also find that for most models this effect appears largely 11 to be driven by simple lexical cues. When we mitigate effects of both world knowl- 12 edge and lexical cues to test knowledge of linguistic nuances of counterfactuals, 13 we find that only GPT-3 shows sensitivity to these nuances, though this sensitivity 14 is also non-trivially impacted by lexical associative factors.

Counterfactual reasoning: Do Language Models need world knowledge for causal inference?

1 Mei Yang was an intern with MSR in the summer of 2007 The system combination approach combining system outputs at the word level is similar to the one described in (Rosti et al., 2007). Compared to the previous work, we developed a new method to generate a better alignment between multiple MT hypotheses from different individual systems, which is used to construct a high-quality confusion network. The details of our method will be elaborated in a future paper (He et al., 2008). First, a minimum Bayes risk (MBR) based method is used to select a backbone from the multiple hypotheses, then all the hypotheses are aligned to that backbone to form a confusion network, i.e., a word lattice in which each word is aligned to a list of alternative words (including null). Then, a set of features, including language model scores, word count, and normalized system voting score, are used to decode the confusion network. In training, a confusion network is constructed based on the multiple hypotheses of each sentence in a dev set. Then the corresponding feature weights are trained using Powell’s search to maximize the BLEU score on that dev set. In testing, a confusion network for each sentence in the test set is constructed and these feature weights are applied to decode the final MT output from the confusion network. In this entry, two language models are used, including a 3-gram LM trained on the English part of the parallel training data, and a 5gram LM trained on the whole English Gigaword corpus using a scalable LM toolkit (Nguyen et al., 2007).

/pdf/the-msr-nrc-sri-mt-system-for-nist-open-machine-translation-3we910hpma.pdf

Kristina Toutanova

Papers

Improving Span-based Question Answering Systems with Coarsely Labeled Data.

Joint Passage Ranking for Diverse Multi-Answer Retrieval

Representations for Question Answering from Documents with Tables and Text

Counterfactual reasoning: Do Language Models need world knowledge for causal inference?

The MSR-NRC-SRI MT System for NIST Open Machine Translation 2008 Evaluation