10 Leading Language Models For NLP In 2022

UPDATE: We have published the updated version of this article with the top 10 transformative LLM research papers from 2023.

The introduction of transfer learning and pretrained language models in natural language processing (NLP) pushed forward the limits of language understanding and generation. Transfer learning and applying transformers to different downstream NLP tasks have become the main trend of the latest research advances.

At the same time, there is a controversy in the NLP community regarding the research value of the huge pretrained language models occupying the leaderboards. While lots of AI experts agree with Anna Rogers’s statement that getting state-of-the-art results just by using more data and computing power is not research news, other NLP opinion leaders point out some positive moments in the current trend, like, for example, the possibility of seeing the fundamental limitations of the current paradigm.

Anyway, the latest improvements in NLP language models seem to be driven not only by the massive boosts in computing capacity but also by the discovery of ingenious ways to lighten models while maintaining high performance.

To help you stay up to date with the latest breakthroughs in language modeling, we’ve summarized research papers featuring the key language models introduced during the last few years.

Subscribe to our AI Research mailing list at the bottom of this article to be alerted when we release new summaries.

If you’d like to skip around, here are the papers we featured:

1. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
2. GPT2: Language Models Are Unsupervised Multitask Learners
3. XLNet: Generalized Autoregressive Pretraining for Language Understanding
4. RoBERTa: A Robustly Optimized BERT Pretraining Approach
5. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations
6. T5: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
7. GPT3: Language Models Are Few-Shot Learners
8. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators
9. DeBERTa: Decoding-enhanced BERT with Disentangled Attention
10. PaLM: Scaling Language Modeling with Pathways

Important Pretrained Language Models

1. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, by Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova

Original Abstract

We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT representations can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications.

BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE benchmark to 80.4% (7.6% absolute improvement), MultiNLI accuracy to 86.7 (5.6% absolute improvement) and the SQuAD v1.1 question answering Test F1 to 93.2 (1.5% absolute improvement), outperforming human performance by 2.0%.

Our Summary

A Google AI team presents a new cutting-edge model for Natural Language Processing (NLP) – BERT, or Bidirectional Encoder Representations from Transformers. Its design allows the model to consider the context from both the left and the right sides of each word. While being conceptually simple, BERT obtains new state-of-the-art results on eleven NLP tasks, including question answering, named entity recognition and other tasks related to general language understanding.

What’s the core idea of this paper?

• Training a deep bidirectional model by randomly masking a percentage of input tokens – thus, avoiding cycles where words can indirectly “see themselves”.
• Also pre-training a sentence relationship model by building a simple binary classification task to predict whether sentence B immediately follows sentence A, thus allowing BERT to better understand relationships between sentences.
• Training a very big model (24 Transformer blocks, 1024-hidden, 340M parameters) with lots of data (3.3 billion word corpus).

What’s the key achievement?

• Advancing the state-of-the-art for 11 NLP tasks, including:
  • getting a GLUE score of 80.4%, which is 7.6% of absolute improvement from the previous best result;
  • achieving 93.2% accuracy on SQuAD 1.1 and outperforming human performance by 2%.
• Suggesting a pre-trained model, which doesn’t require any substantial architecture modifications to be applied to specific NLP tasks.

What does the AI community think?

• BERT model marks a new era of NLP.
• In a nutshell, two unsupervised tasks together (“fill in the blank” and “does sentence B comes after sentence A?” ) provide great results for many NLP tasks.
• Pre-training of language models becomes a new standard.

What are future research areas?

• Testing the method on a wider range of tasks.
• Investigating the linguistic phenomena that may or may not be captured by BERT.

What are possible business applications?

• BERT may assist businesses with a wide range of NLP problems, including:
  • chatbots for better customer experience;
  • analysis of customer reviews;
  • the search for relevant information, etc.

Where can you get implementation code?

• Google Research has released an official Github repository with Tensorflow code and pre-trained models for BERT.
• PyTorch implementation of BERT is also available on GitHub.

2. Language Models Are Unsupervised Multitask Learners, by Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever

Original Abstract

Natural language processing tasks, such as question answering, machine translation, reading comprehension, and summarization, are typically approached with supervised learning on task-specific datasets. We demonstrate that language models begin to learn these tasks without any explicit supervision when trained on a new dataset of millions of webpages called WebText. When conditioned on a document plus questions, the answers generated by the language model reach 55 F1 on the CoQA dataset – matching or exceeding the performance of 3 out of 4 baseline systems without using the 127,000+ training examples. The capacity of the language model is essential to the success of zero-shot task transfer and increasing it improves performance in a log-linear fashion across tasks. Our largest model, GPT-2, is a 1.5B parameter Transformer that achieves state of the art results on 7 out of 8 tested language modeling datasets in a zero-shot setting but still underfits WebText. Samples from the model reflect these improvements and contain coherent paragraphs of text. These findings suggest a promising path towards building language processing systems which learn to perform tasks from their naturally occurring demonstrations.

Our Summary

In this paper, the OpenAI team demonstrates that pre-trained language models can be used to solve downstream tasks without any parameter or architecture modifications. They have trained a very big model, a 1.5B-parameter Transformer, on a large and diverse dataset that contains text scraped from 45 million webpages. The model generates coherent paragraphs of text and achieves promising, competitive or state-of-the-art results on a wide variety of tasks.

What’s the core idea of this paper?

• Training the language model on the large and diverse dataset:
  • selecting webpages that have been curated/filtered by humans;
  • cleaning and de-duplicating the texts, and removing all Wikipedia documents to minimize overlapping of training and test sets;
  • using the resulting WebText dataset with slightly over 8 million documents for a total of 40 GB of text.
• Using a byte-level version of Byte Pair Encoding (BPE) for input representation.
• Building a very big Transformer-based model, GPT-2:
  • the largest model includes 1542M parameters and 48 layers;
  • the model mainly follows the OpenAI GPT model with few modifications (i.e., expanding vocabulary and context size, modifying initialization etc.).

What’s the key achievement?

• Getting state-of-the-art results on 7 out of 8 tested language modeling datasets.
• Showing quite promising results in commonsense reasoning, question answering, reading comprehension, and translation.
• Generating coherent texts, for example, a news article about the discovery of talking unicorns.

What does the AI community think?

• “The researchers built an interesting dataset, applying now-standard tools and yielding an impressive model.” – Zachary C. Lipton, an assistant professor at Carnegie Mellon University.

What are future research areas?

• Investigating fine-tuning on benchmarks such as decaNLP and GLUE to see whether the huge dataset and capacity of GPT-2 can overcome the inefficiencies of BERT’s unidirectional representations.

What are possible business applications?

• In terms of practical applications, the performance of the GPT-2 model without any fine-tuning is far from usable but it shows a very promising research direction.

Where can you get implementation code?

• Initially, OpenAI decided to release only a smaller version of GPT-2 with 117M parameters. The decision not to release larger models was taken “due to concerns about large language models being used to generate deceptive, biased, or abusive language at scale”.
• In November, OpenAI finally released its largest 1.5B-parameter model. The code is available here.
• Hugging Face has introduced a PyTorch implementation of the initially released GPT-2 model.

3. XLNet: Generalized Autoregressive Pretraining for Language Understanding, by Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le

Original Abstract

With the capability of modeling bidirectional contexts, denoising autoencoding based pretraining like BERT achieves better performance than pretraining approaches based on autoregressive language modeling. However, relying on corrupting the input with masks, BERT neglects dependency between the masked positions and suffers from a pretrain-finetune discrepancy. In light of these pros and cons, we propose XLNet, a generalized autoregressive pretraining method that (1) enables learning bidirectional contexts by maximizing the expected likelihood over all permutations of the factorization order and (2) overcomes the limitations of BERT thanks to its autoregressive formulation. Furthermore, XLNet integrates ideas from Transformer-XL, the state-of-the-art autoregressive model, into pretraining. Empirically, XLNet outperforms BERT on 20 tasks, often by a large margin, and achieves state-of-the-art results on 18 tasks including question answering, natural language inference, sentiment analysis, and document ranking.

Our Summary

The researchers from Carnegie Mellon University and Google have developed a new model, XLNet, for natural language processing (NLP) tasks such as reading comprehension, text classification, sentiment analysis, and others. XLNet is a generalized autoregressive pretraining method that leverages the best of both autoregressive language modeling (e.g., Transformer-XL) and autoencoding (e.g., BERT) while avoiding their limitations. The experiments demonstrate that the new model outperforms both BERT and Transformer-XL and achieves state-of-the-art performance on 18 NLP tasks.

What’s the core idea of this paper?

• XLNet combines the bidirectional capability of BERT with the autoregressive technology of Transformer-XL:
  • Like BERT, XLNet uses a bidirectional context, which means it looks at the words before and after a given token to predict what it should be. To this end, XLNet maximizes the expected log-likelihood of a sequence with respect to all possible permutations of the factorization order.
  • As an autoregressive language model, XLNet doesn’t rely on data corruption, and thus avoids BERT’s limitations due to masking – i.e., pretrain-finetune discrepancy and the assumption that unmasked tokens are independent of each other.
• To further improve architectural designs for pretraining, XLNet integrates the segment recurrence mechanism and relative encoding scheme of Transformer-XL.

What’s the key achievement?

• XLnet outperforms BERT on 20 tasks, often by a large margin.
• The new model achieves state-of-the-art performance on 18 NLP tasks including question answering, natural language inference, sentiment analysis, and document ranking.

What does the AI community think?

• The paper was accepted for oral presentation at NeurIPS 2019, the leading conference in artificial intelligence.
• “The king is dead. Long live the king. BERT’s reign might be coming to an end. XLNet, a new model by people from CMU and Google outperforms BERT on 20 tasks.” – Sebastian Ruder, a research scientist at Deepmind.
• “XLNet will probably be an important tool for any NLP practitioner for a while…[it is] the latest cutting-edge technique in NLP.” – Keita Kurita, Carnegie Mellon University.

What are future research areas?

• Extending XLNet to new areas, such as computer vision and reinforcement learning.

What are possible business applications?

• XLNet may assist businesses with a wide range of NLP problems, including:
  • chatbots for first-line customer support or answering product inquiries;
  • sentiment analysis for gauging brand awareness and perception based on customer reviews and social media;
  • the search for relevant information in document bases or online, etc.

Where can you get implementation code?

• The authors have released the official Tensorflow implementation of XLNet.
• PyTorch implementation of the model is also available on GitHub.

4. RoBERTa: A Robustly Optimized BERT Pretraining Approach, by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov

Original Abstract

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.

Our Summary

Natural language processing models have made significant advances thanks to the introduction of pretraining methods, but the computational expense of training has made replication and fine-tuning parameters difficult. In this study, Facebook AI and the University of Washington researchers analyzed the training of Google’s Bidirectional Encoder Representations from Transformers (BERT) model and identified several changes to the training procedure that enhance its performance. Specifically, the researchers used a new, larger dataset for training, trained the model over far more iterations, and removed the next sequence prediction training objective. The resulting optimized model, RoBERTa (Robustly Optimized BERT Approach), matched the scores of the recently introduced XLNet model on the GLUE benchmark.

What’s the core idea of this paper?

• The Facebook AI research team found that BERT was significantly undertrained and suggested an improved recipe for its training, called RoBERTa:
  • More data: 160GB of text instead of the 16GB dataset originally used to train BERT.
  • Longer training: increasing the number of iterations from 100K to 300K and then further to 500K.
  • Larger batches: 8K instead of 256 in the original BERT base model.
  • Larger byte-level BPE vocabulary with 50K subword units instead of character-level BPE vocabulary of size 30K.
  • Removing the next sequence prediction objective from the training procedure.
  • Dynamically changing the masking pattern applied to the training data.

What’s the key achievement?

• RoBERTa outperforms BERT in all individual tasks on the General Language Understanding Evaluation (GLUE) benchmark.
• The new model matches the recently introduced XLNet model on the GLUE benchmark and sets a new state of the art in four out of nine individual tasks.

What are future research areas?

• Incorporating more sophisticated multi-task finetuning procedures.

What are possible business applications?

• Big pretrained language frameworks like RoBERTa can be leveraged in the business setting for a wide range of downstream tasks, including dialogue systems, question answering, document classification, etc.

Where can you get implementation code?

• The models and code used in this study are available on GitHub.

5. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations, by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut

Original Abstract

Increasing model size when pretraining natural language representations often results in improved performance on downstream tasks. However, at some point further model increases become harder due to GPU/TPU memory limitations, longer training times, and unexpected model degradation. To address these problems, we present two parameter-reduction techniques to lower memory consumption and increase the training speed of BERT. Comprehensive empirical evidence shows that our proposed methods lead to models that scale much better compared to the original BERT. We also use a self-supervised loss that focuses on modeling inter-sentence coherence, and show it consistently helps downstream tasks with multi-sentence inputs. As a result, our best model establishes new state-of-the-art results on the GLUE, RACE, and SQuAD benchmarks while having fewer parameters compared to BERT-large.

Our Summary

The Google Research team addresses the problem of the continuously growing size of the pretrained language models, which results in memory limitations, longer training time, and sometimes unexpectedly degraded performance. Specifically, they introduce A Lite BERT (ALBERT) architecture that incorporates two parameter-reduction techniques: factorized embedding parameterization and cross-layer parameter sharing. In addition, the suggested approach includes a self-supervised loss for sentence-order prediction to improve inter-sentence coherence. The experiments demonstrate that the best version of ALBERT sets new state-of-the-art results on GLUE, RACE, and SQuAD benchmarks while having fewer parameters than BERT-large.

What’s the core idea of this paper?

• It is not reasonable to further improve language models by making them larger because of memory limitations of available hardware, longer training times, and unexpected degradation of model performance with the increased number of parameters.
• To address this problem, the researchers introduce the ALBERT architecture that incorporates two parameter-reduction techniques:
  • factorized embedding parameterization, where the size of the hidden layers is separated from the size of vocabulary embeddings by decomposing the large vocabulary-embedding matrix into two small matrices;
  • cross-layer parameter sharing to prevent the number of parameters from growing with the depth of the network.
• The performance of ALBERT is further improved by introducing the self-supervised loss for sentence-order prediction to address BERT’s limitations with regard to inter-sentence coherence.

What’s the key achievement?

• With the introduced parameter-reduction techniques, the ALBERT configuration with 18× fewer parameters and 1.7× faster training compared to the original BERT-large model achieves only slightly worse performance.
• The much larger ALBERT configuration, which still has fewer parameters than BERT-large, outperforms all of the current state-of-the-art language modes by getting:
  • 89.4% accuracy on the RACE benchmark;
  • 89.4 score on the GLUE benchmark; and
  • An F1 score of 92.2 on the SQuAD 2.0 benchmark.

What does the AI community think?

• The paper has been submitted to ICLR 2020 and is available on the OpenReview forum, where you can see the reviews and comments of NLP experts. The reviewers are mainly very appreciative of the presented paper.

What are future research areas?

• Speeding up training and inference through methods like sparse attention and block attention.
• Further improving the model performance through hard example mining, more efficient model training, and other approaches.

What are possible business applications?

• The ALBERT language model can be leveraged in the business setting to improve performance on a wide range of downstream tasks, including chatbot performance, sentiment analysis, document mining, and text classification.

Where can you get implementation code?

• The original implementation of ALBERT is available on GitHub.
• A TensorFlow implementation of ALBERT is also available here.
• A PyTorch implementation of ALBERT can be found here and here.

6. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer, by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu

Original Abstract

Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.

Our Summary

The Google research team suggests a unified approach to transfer learning in NLP with the goal to set a new state of the art in the field. To this end, they propose treating each NLP problem as a “text-to-text” problem. Such a framework allows using the same model, objective, training procedure, and decoding process for different tasks, including summarization, sentiment analysis, question answering, and machine translation. The researchers call their model a Text-to-Text Transfer Transformer (T5) and train it on the large corpus of web-scraped data to get state-of-the-art results on a number of NLP tasks.

What’s the core idea of this paper?

• The paper has several important contributions:
  • Providing a comprehensive perspective on where the NLP field stands by exploring and comparing existing techniques.
  • Introducing a new approach to transfer learning in NLP by suggesting treating every NLP problem as a text-to-text task:
    • The model understands which tasks should be performed thanks to the task-specific prefix added to the original input sentence (e.g., “translate English to German:”, “summarize:”).
  • Presenting and releasing a new dataset consisting of hundreds of gigabytes of clean web-scraped English text, the Colossal Clean Crawled Corpus (C4).
  • Training a large (up to 11B parameters) model, called Text-to-Text Transfer Transformer (T5) on the C4 dataset.

What’s the key achievement?

• The T5 model with 11 billion parameters achieved state-of-the-art performance on 17 out of 24 tasks considered, including:
  • a GLUE score of 89.7 with substantially improved performance on CoLA, RTE, and WNLI tasks;
  • an Exact Match score of 90.06 on the SQuAD dataset;
  • a SuperGLUE score of 88.9, which is a very significant improvement over the previous state-of-the-art result (84.6) and very close to human performance (89.8);
  • a ROUGE-2-F score of 21.55 on the CNN/Daily Mail abstractive summarization task.

What are future research areas?

• Researching the methods to achieve stronger performance with cheaper models.
• Exploring more efficient knowledge extraction techniques.
• Further investigating the language-agnostic models.

What are possible business applications?

• Even though the introduced model has billions of parameters and can be too heavy to be applied in the business setting, the presented ideas can be used to improve the performance on different NLP tasks, including summarization, question answering, and sentiment analysis.

Where can you get implementation code?

• The pretrained models together with the dataset and code are released on GitHub.

7. Language Models are Few-Shot Learners, by Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, Dario Amodei

Original Abstract

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, 10× 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.

Our Summary

The OpenAI research team draws attention to the fact that the need for a labeled dataset for every new language task limits the applicability of language models. Considering that there is a wide range of possible tasks and it’s often difficult to collect a large labeled training dataset, the researchers suggest an alternative solution, which is scaling up language models to improve task-agnostic few-shot performance. They test their solution by training a 175B-parameter autoregressive language model, called GPT-3, and evaluating its performance on over two dozen NLP tasks. The evaluation under few-shot learning, one-shot learning, and zero-shot learning demonstrates that GPT-3 achieves promising results and even occasionally outperforms the state of the art achieved by fine-tuned models.

What’s the core idea of this paper?

• The GPT-3 model uses the same model and architecture as GPT-2, including the modified initialization, pre-normalization, and reversible tokenization.
• However, in contrast to GPT-2, it uses alternating dense and locally banded sparse attention patterns in the layers of the transformer, as in the Sparse Transformer.
• The model is evaluated in three different settings:
  • Few-shot learning, when the model is given a few demonstrations of the task (typically, 10 to 100) at inference time but with no weight updates allowed.
  • One-shot learning, when only one demonstration is allowed, together with a natural language description of the task.
  • Zero-shot learning, when no demonstrations are allowed and the model has access only to a natural language description of the task.

What’s the key achievement?

• The GPT-3 model without fine-tuning achieves promising results on a number of NLP tasks, and even occasionally surpasses state-of-the-art models that were fine-tuned for that specific task:
  • On the CoQA benchmark, 81.5 F1 in the zero-shot setting, 84.0 F1 in the one-shot setting, and 85.0 F1 in the few-shot setting, compared to the 90.7 F1 score achieved by fine-tuned SOTA.
  • On the TriviaQA benchmark, 64.3% accuracy in the zero-shot setting, 68.0% in the one-shot setting, and 71.2% in the few-shot setting, surpassing the state of the art (68%) by 3.2%.
  • On the LAMBADA dataset, 76.2 % accuracy in the zero-shot setting, 72.5% in the one-shot setting, and 86.4% in the few-shot setting, surpassing the state of the art (68%) by 18%.
• The news articles generated by the 175B-parameter GPT-3 model are hard to distinguish from real ones, according to human evaluations (with accuracy barely above the chance level at ~52%).

What are future research areas?

• Improving pre-training sample efficiency.
• Exploring how few-shot learning works.
• Distillation of large models down to a manageable size for real-world applications.

What does the AI community think?

• “The GPT-3 hype is way too much. It’s impressive (thanks for the nice compliments!) but it still has serious weaknesses and sometimes makes very silly mistakes. AI is going to change the world, but GPT-3 is just a very early glimpse. We have a lot still to figure out.” – Sam Altman, CEO and co-founder of OpenAI.
• “I’m shocked how hard it is to generate text about Muslims from GPT-3 that has nothing to do with violence… or being killed…” – Abubakar Abid, CEO and founder of Gradio.
• “No. GPT-3 fundamentally does not understand the world that it talks about. Increasing corpus further will allow it to generate a more credible pastiche but not fix its fundamental lack of comprehension of the world.  Demos of GPT-4 will still require human cherry picking.” – Gary Marcus, CEO and founder of Robust.ai.
• “Extrapolating the spectacular performance of GPT3 into the future suggests that the answer to life, the universe and everything is just 4.398 trillion parameters.” – Geoffrey Hinton, Turing Award winner.

What are possible business applications?

• The model with 175B parameters is hard to apply to real business problems due to its impractical resource requirements, but if the researchers manage to distill this model down to a workable size, it could be applied to a wide range of language tasks, including question answering and ad copy generation.

Where can you get implementation code?

• The code itself is not available, but some dataset statistics together with unconditional, unfiltered 2048-token samples from GPT-3 are released on GitHub.

8. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators, by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning

Original Abstract

Masked language modeling (MLM) pre-training methods such as BERT corrupt the input by replacing some tokens with [MASK] and then train a model to reconstruct the original tokens. While they produce good results when transferred to downstream NLP tasks, they generally require large amounts of compute to be effective. As an alternative, we propose a more sample-efficient pre-training task called replaced token detection. Instead of masking the input, our approach corrupts it by replacing some tokens with plausible alternatives sampled from a small generator network. Then, instead of training a model that predicts the original identities of the corrupted tokens, we train a discriminative model that predicts whether each token in the corrupted input was replaced by a generator sample or not. Thorough experiments demonstrate this new pre-training task is more efficient than MLM because the task is defined over all input tokens rather than just the small subset that was masked out. As a result, the contextual representations learned by our approach substantially outperform the ones learned by BERT given the same model size, data, and compute. The gains are particularly strong for small models; for example, we train a model on one GPU for 4 days that outperforms GPT (trained using 30× more compute) on the GLUE natural language understanding benchmark. Our approach also works well at scale, where it performs comparably to RoBERTa and XLNet while using less than 1/4 of their compute and outperforms them when using the same amount of compute.

Our Summary

The pre-training task for popular language models like BERT and XLNet involves masking a small subset of unlabeled input and then training the network to recover this original input. Even though it works quite well, this approach is not particularly data-efficient as it learns from only a small fraction of tokens (typically ~15%). As an alternative, the researchers from Stanford University and Google Brain propose a new pre-training task called replaced token detection. Instead of masking, they suggest replacing some tokens with plausible alternatives generated by a small language model. Then, the pre-trained discriminator is used to predict whether each token is an original or a replacement. As a result, the model learns from all input tokens instead of the small masked fraction, making it much more computationally efficient. The experiments confirm that the introduced approach leads to significantly faster training and higher accuracy on downstream NLP tasks.

What’s the core idea of this paper?

• Pre-training methods that are based on masked language modeling are computationally inefficient as they use only a small fraction of tokens for learning.
• Researchers propose a new pre-training task called replaced token detection, where:
  • some tokens are replaced by samples from a small generator network;
  • a model is pre-trained as a discriminator to distinguish between original and replaced tokens.
• The introduced approach, called ELECTRA (Efficiently Learning an Encoder that Classifies Token Replacements Accurately):
  • enables the model to learn from all input tokens instead of the small masked-out subset;
  • is not adversarial, despite the similarity to GAN, as the generator producing tokens for replacement is trained with maximum likelihood.

What’s the key achievement?

• Demonstrating that the discriminative task of distinguishing between real data and challenging negative samples is more efficient than existing generative methods for language representation learning.
• Introducing a model that substantially outperforms state-of-the-art approaches while requiring less pre-training compute:
  • ELECTRA-Small gets a GLUE score of 79.9 and outperforms a comparably small BERT model with a score of 75.1 and a much larger GPT model with a score of 78.8.
  • An ELECTRA model that performs comparably to XLNet and RoBERTa uses only 25% of their pre-training compute.
  • ELECTRA-Large outscores the alternative state-of-the-art models on the GLUE and SQuAD benchmarks while still requiring less pre-training compute.

What does the AI community think?

• The paper was selected for presentation at ICLR 2020, the leading conference in deep learning.

What are possible business applications?

• Because of its computational efficiency, the ELECTRA approach can make the application of pre-trained text encoders more accessible to business practitioners.

Where can you get implementation code?

• The original TensorFlow implementation and pre-trained weights are released on GitHub.

9. DeBERTa: Decoding-enhanced BERT with Disentangled Attention, by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen

Original Abstract 

Recent progress in pre-trained neural language models has significantly improved the performance of many natural language processing (NLP) tasks. In this paper we propose a new model architecture DeBERTa (Decoding-enhanced BERT with disentangled attention) that improves the BERT and RoBERTa models using two novel techniques. The first is the disentangled attention mechanism, where each word is represented using two vectors that encode its content and position, respectively, and the attention weights among words are computed using disentangled matrices on their contents and relative positions, respectively. Second, an enhanced mask decoder is used to incorporate absolute positions in the decoding layer to predict the masked tokens in model pre-training. In addition, a new virtual adversarial training method is used for fine-tuning to improve models’ generalization. We show that these techniques significantly improve the efficiency of model pre-training and the performance of both natural language understanding (NLU) and natural language generation (NLG) downstream tasks. Compared to RoBERTa-Large, a DeBERTa model trained on half of the training data performs consistently better on a wide range of NLP tasks, achieving improvements on MNLI by +0.9% (90.2% vs. 91.1%), on SQuAD v2.0 by +2.3% (88.4% vs. 90.7%) and RACE by +3.6% (83.2% vs. 86.8%). Notably, we scale up DeBERTa by training a larger version that consists of 48 Transform layers with 1.5 billion parameters. The significant performance boost makes the single DeBERTa model surpass the human performance on the SuperGLUE benchmark (Wang et al., 2019a) for the first time in terms of macro-average score (89.9 versus 89.8), and the ensemble DeBERTa model sits atop the SuperGLUE leaderboard as of January 6, 2021, outperforming the human baseline by a decent margin (90.3 versus 89.8).

Our Summary 

The authors from Microsoft Research propose DeBERTa, with two main improvements over BERT, namely disentangled attention and an enhanced mask decoder. DeBERTa has two vectors representing a token/word by encoding content and relative position respectively. The self-attention mechanism in DeBERTa processes self-attention of content-to-content, content-to-position, and also position-to-content, while the self-attention in BERT is equivalent to only having the first two components. The authors hypothesize that position-to-content self-attention is also needed to comprehensively model relative positions in a sequence of tokens. Furthermore, DeBERTa is equipped with an enhanced mask decoder, where the absolute position of the token/word is also given to the decoder along with the relative information. A single scaled-up variant of DeBERTa surpasses the human baseline on the SuperGLUE benchmark for the first time. The ensemble DeBERTa is the top-performing method on SuperGLUE at the time of this publication.

What’s the core idea of this paper?

• Disentangled attention: In the original BERT, the content embedding and position embedding are added before self-attention and the self-attention is applied only on the output of content and position vectors. The authors hypothesize that this only accounts for content-to-content self-attention and content-to-position self-attention and that we need position-to-content self-attention as well to model position information completely. DeBERTa has two separate vectors representing content and position and self-attention is calculated between all possible pairs, i.e., content-to-content, content-to-position, position-to-content, and position-to-position. Position-to-position self-attention is trivially 1 all the time and has no information, so it is not computed. 
• Enhanced mask decoder: The authors hypothesize that the model needs absolute position information to understand syntactical nuances such as subject-object characterization. So, DeBERTa is provided with absolute position information along with relative position information. The absolute position embedding is provided to the last decoder layer just before the softmax layer, which gives the output.

• Scale-invariant fine-tuning: A virtual adversarial training algorithm called scale-invariant fine-tuning is used as a regularization method to increase generalization. The word embeddings are perturbed to a small extent and trained to produce the same output as they would on non-perturbed word embeddings. The word embedding vectors are normalized to stochastic vectors (where the sum of the elements in a vector is 1) to be invariant to the number of parameters in the model. 

What’s the key achievement?

• Compared to the current state-of-the-art method RoBERTa-Large, the DeBERTA model trained on half the training data achieves:
  • an improvement of +0.9% in accuracy on MNLI (91.1% vs. 90.2%),
  • an improvement of +2.3% in accuracy on SQuAD v2.0 (90.7% vs. 88.4%),
  • an improvement of +3.6% in accuracy on RACE (86.8% vs. 83.2%)
• A single scaled-up variant of DeBERTa surpasses the human baseline on the SuperGLUE benchmark for the first time (89.9 vs. 89.8). The ensemble DeBERTa is the top-performing method on SuperGLUE at the time of this publication, outperforming the human baseline by a decent margin (90.3 versus 89.8).

What does the AI community think?

• The paper has been accepted to ICLR 2021, one of the key conferences in deep learning.

What are future research areas?

• Improving pretraining by introducing other useful information, in addition to positions, with the Enhanced Mask Decoder (EMD) framework.
• A more comprehensive study of scale-invariant fine-tuning (SiFT).

What are possible business applications?

• The contextual representations of pretrained language modeling could be used in search, question answering, summarization, virtual assistants, and chatbots, among other tasks.

Where can you get implementation code?

• The implementation of DeBERTa is available on GitHub.

10. PaLM: Scaling Language Modeling with Pathways, by Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, Noah Fiedel

Original Abstract

Large language models have been shown to achieve remarkable performance across a variety of natural language tasks using few-shot learning, which drastically reduces the number of task-specific training examples needed to adapt the model to a particular application. To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model PaLM. We trained PaLM on 6144 TPU v4 chips using Pathways, a new ML system which enables highly efficient training across multiple TPU Pods. We demonstrate continued benefits of scaling by achieving state-of-the-art few-shot learning results on hundreds of language understanding and generation benchmarks. On a number of these tasks, PaLM 540B achieves breakthrough performance, outperforming the finetuned state-of-the-art on a suite of multi-step reasoning tasks, and outperforming average human performance on the recently released BIG-bench benchmark. A significant number of BIG-bench tasks showed discontinuous improvements from model scale, meaning that performance steeply increased as we scaled to our largest model. PaLM also has strong capabilities in multilingual tasks and source code generation, which we demonstrate on a wide array of benchmarks. We additionally provide a comprehensive analysis on bias and toxicity, and study the extent of training data memorization with respect to model scale. Finally, we discuss the ethical considerations related to large language models and discuss potential mitigation strategies.

Our Summary

The Google Research team contributed a lot in the area of pre-trained language models with their BERT, ALBERT, and T5 models. One of their latest contributions is the Pathways Language Model (PaLM), a 540-billion parameter, dense decoder-only Transformer model trained with the Pathways system. The goal of the Pathways system is to orchestrate distributed computation for accelerators. With its help, the team was able to efficiently train a single model across multiple TPU v4 Pods. The experiments on hundreds of language understanding and generation tasks demonstrated that PaLM achieves state-of-the-art few-shot performance across most tasks with breakthrough capabilities demonstrated in language understanding, language generation, reasoning, and code-related tasks.

What’s the core idea of this paper?

• The main idea of the paper is to scale training of a 540-billion parameter language model with the Pathways system:
  • The team was using data parallelism at the Pod level across two Cloud TPU v4 Pods while using standard data and model parallelism within each Pod.
  • They were able to scale training to 6144 TPU v4 chips, the largest TPU-based system configuration used for training to date.
  • The model achieved a training efficiency of 57.8% hardware FLOPs utilization, which, as the authors claim, is the highest yet achieved training efficiency for large language models at this scale. 
• The training data for the PaLM model included a combination of English and multilingual datasets containing high-quality web documents, books, Wikipedia, conversations, and GitHub code.

What’s the key achievement?

• Numerous experiments demonstrate that model performance steeply increased as the team scaled to their largest model.
• PaLM 540B achieved breakthrough performance on multiple very difficult tasks:
  • Language understanding and generation. The introduced model surpassed the few-shot performance of prior large models on 28 out of 29 tasks that include question-answering tasks, cloze and sentence-completion tasks, in-context reading comprehension tasks, common-sense reasoning tasks, SuperGLUE tasks, and more. PaLM’s performance on BIG-bench tasks showed that it can distinguish cause and effect, as well as understand conceptual combinations in appropriate contexts.
  • Reasoning. With 8-shot prompting, PaLM solves 58% of the problems in GSM8K, a benchmark of thousands of challenging grade school level math questions, outperforming the prior top score of 55% achieved by fine-tuning the GPT-3 175B model. PaLM also demonstrates the ability to generate explicit explanations in situations that require a complex combination of multi-step logical inference, world knowledge, and deep language understanding.
  • Code generation. PaLM performs on par with the fine-tuned Codex 12B while using 50 times less Python code for training, confirming that large language models transfer learning from both other programming languages and natural language data more effectively.

What are future research areas?

• Combining the scaling capabilities of the Pathways system with novel architectural choices and training schemes.

What are possible business applications?

• Similarly to other recently introduced pre-trained language models, PaLM can be applied in a wide range of downstream tasks, including conversational AI, question answering, machine translation, document classification, ad copy generation, code bug fixing, and more.

Where can you get implementation code?

• So far, there was no official code implementation release for PaLM but it actually uses a standard Transformer model architecture, with some customizations. 
• Pytorch implementation of the specific Transformer architecture from PaLM can be accessed on GitHub.

If you like these research summaries, you might be also interested in the following articles:

• 2020’s Top AI & Machine Learning Research Papers
• GPT-3 & Beyond: 10 NLP Research Papers You Should Read
• The Latest Breakthroughs in Conversational AI Agents
• What Every NLP Engineer Needs To Know About Pre-Trained Language Models

Enjoy this article? Sign up for more AI research updates.

We’ll let you know when we release more summary articles like this one.