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 present a model for semantic role labeling that effectively captures the linguistic intuition that a semantic argument frame is a joint structure, with strong dependencies among the arguments. We show how to incorporate these strong dependencies in a statistical joint model with a rich set of features over multiple argument phrases. The proposed model substantially outperforms a similar state-of-the-art local model that does not include dependencies among different arguments.

We evaluate the gains from incorporating this joint information on the Propbank corpus, when using correct syntactic parse trees as input, and when using automatically derived parse trees. The gains amount to 24.1% error reduction on all arguments and 36.8% on core arguments for gold-standard parse trees on Propbank. For automatic parse trees, the error reductions are 8.3% and 10.3% on all and core arguments, respectively. We also present results on the CoNLL 2005 shared task data set. Additionally, we explore considering multiple syntactic analyses to cope with parser noise and uncertainty.

/pdf/a-global-joint-model-for-semantic-role-labeling-4tni921j4b.pdf

A global joint model for semantic role labeling

Recent developments in NLP have been accompanied by large, expensive models. Knowledge distillation is the standard method to realize these gains in applications with limited resources: a compact student is trained to recover the outputs of a powerful teacher. While most prior work investigates student architectures and transfer techniques, we focus on an often-neglected aspect---student initialization. We argue that a random starting point hinders students from fully leveraging the teacher expertise, even in the presence of a large transfer set. We observe that applying language model pre-training to students unlocks their generalization potential, surprisingly even for very compact networks. We conduct experiments on 4 NLP tasks and 24 sizes of Transformer-based students; for sentiment classification on the Amazon Book Reviews dataset, pre-training boosts size reduction and TPU speed-up from 3.1x/1.25x to 31x/16x. Extensive ablation studies dissect the interaction between pre-training and distillation, revealing a compound effect even when they are applied on the same unlabeled dataset.

Well-Read Students Learn Better: The Impact of Student Initialization on Knowledge Distillation

The LinGO Redwoods initiative is a seed activity in the design and development of a new type of treebank. A treebank is a (typically hand-built) collection of natural language utterances and associated linguistic analyses; typical treebanks—as for example the widely recognized Penn Treebank (Marcus, Santorini, & Marcinkiewicz, 1993), the Prague Dependency Treebank (Hajic, 1998), or the German TiGer Corpus (Skut, Krenn, Brants, & Uszkoreit, 1997)—assign syntactic phrase structure or tectogrammatical dependency trees over sentences taken from a naturallyoccuring source, often newspaper text. Applications of existing treebanks fall into two broad categories: (i) use of an annotated corpus in empirical linguistics as a source of structured language data and distributional patterns and (ii) use of the treebank for the acquisition (e.g. using stochastic or machine learning approaches) and evaluation of parsing systems. While several mediumto large-scale treebanks exist for English (and some for other major languages), all pre-existing publicly available resources exhibit the following limitations: (i) the depth of linguistic information recorded in these treebanks is comparatively shallow, (ii) the design and format of linguistic representation in the treebank hard-wires a small, predefined range of ways in which information can be extracted from the treebank, and (iii) representations in existing treebanks are static and over the (often yearor decade-long) evolution of a large-scale treebank tend to fall behind theoretical advances in formal linguistics and grammatical representation. LinGO Redwoods aims at the development of a novel treebanking methodology, (i) rich in nature and dynamic in both (ii) the ways linguistic data can be retrieved from the treebank in varying granularity and (iii) the constant evolution and regular updating of the treebank itself, synchronized to the development of ideas in syntactic theory. Starting in October 2001, the project is aiming to build the foundations for this new type of treebank, develop a basic set of tools required for treebank construction and maintenance, and construct an initial set of 10,000 annotated trees to be distributed together with the tools under an open-source license. Building a largescale treebank, disseminating it, and positioning the corpus as a widely-accepted resource is a multi-year effort; the results of this seeding activity will serve as a proof of concept for the novel approach that is expected to enable the LinGO group at CSLI both to disseminate the approach to the wider academic and industrial audience and to secure appropriate funding for the realization and exploitation of a larger treebank. The purpose of publication at this early stage is three-fold: (i) to encourage feedback on the Redwoods approach from a broader academic audience, (ii) to facilitate exchange with related work at other sites, and (iii) to invite additional collaborators to contribute to the construction of the Redwoods treebank or start its exploitation as early-access versions become available. This paper is an updated version of an earlier project report published by Oepen, Callahan, Flickinger, and Manning (2002); changes over that version include more recent numbers on the current Redwoods development status, inclusion of an example of discriminator-based disambiguation, and minor adaptations and corrections in various parts of the discussion.

/pdf/lingo-redwoods-a-rich-and-dynamic-treebank-for-hpsg-49mdu31cyd.pdf

LinGO Redwoods: A Rich and Dynamic Treebank for HPSG

Recent work on open domain question answering (QA) assumes strong supervision of the supporting evidence and/or assumes a blackbox information retrieval (IR) system to retrieve evidence candidates. We argue that both are suboptimal, since gold evidence is not always available, and QA is fundamentally different from IR. We show for the first time that it is possible to jointly learn the retriever and reader from question-answer string pairs and without any IR system. In this setting, evidence retrieval from all of Wikipedia is treated as a latent variable. Since this is impractical to learn from scratch, we pre-train the retriever with an Inverse Cloze Task. We evaluate on open versions of five QA datasets. On datasets where the questioner already knows the answer, a traditional IR system such as BM25 is sufficient. On datasets where a user is genuinely seeking an answer, we show that learned retrieval is crucial, outperforming BM25 by up to 19 points in exact match.

Latent Retrieval for Weakly Supervised Open Domain Question Answering

A word is selected from a received text and features are identified from the word. The features are applied to a model to identify probabilities for sets of part-of-speech tags. The probabilities for the sets of part-of-speech tags are used to weight scores for possible part-of-speech tags for the selected word to form weighted scores. The weighted scores are used to select a part-of-speech tag for the word and the selected part of speech tag is stored or output. The scores for the possible part-of-speech tags are based on variational approximation parameters trained from a sparse prior over probability distributions describing the probability of a part-of-speech tag given a word.

Kristina Toutanova

Papers

A global joint model for semantic role labeling

Well-Read Students Learn Better: The Impact of Student Initialization on Knowledge Distillation

LinGO Redwoods: A Rich and Dynamic Treebank for HPSG

Latent Retrieval for Weakly Supervised Open Domain Question Answering

Semi-supervised part-of-speech tagging