Transformers are feedforward networks that can process input tokens in parallel. While this parallelization makes them computationally efficient, it restricts the model from fully exploiting the sequential nature of the input - the representation at a given layer can only access representations from lower layers, rather than the higher level representations already built in previous time steps. In this work, we propose the Feedback Transformer architecture that exposes all previous representations to all future representations, meaning the lowest representation of the current timestep is formed from the highest-level abstract representation of the past. We demonstrate on a variety of benchmarks in language modeling, neural machine translation, summarization, and reinforcement learning that the increased representation capacity can improve over Transformer baselines.
We study ResNet-, Time-Depth Separable ConvNets-, and Transformer-based acoustic models, trained with CTC or Seq2Seq criterions. We perform experiments on the LibriSpeech dataset, with and without LM decoding, optionally with beam rescoring. We reach 5.18% WER with external language models for decoding and rescoring. Additionally, we leverage the unlabeled data from LibriVox by doing semi-supervised training and show that it is possible to reach 5.29% WER on test-other without decoding, and 4.11% WER with decoding and rescoring, with only the standard 960 hours from LibriSpeech as labeled data.
Pre-training text representations have led to significant improvements in many areas of natural language processing. The quality of these models benefits greatly from the size of the pretraining corpora as long as its quality is preserved. In this paper, we describe an automatic pipeline to extract massive high-quality monolingual datasets from Common Crawl for a variety of languages. Our pipeline follows the data processing introduced in fastText (Mikolov et al., 2017; Grave et al., 2018), that deduplicates documents and identifies their language. We augment this pipeline with a filtering step to select documents that are close to high quality corpora like Wikipedia.
We show that margin-based bitext mining in a multilingual sentence space can be applied to monolingual corpora of billions of sentences. We are using ten snapshots of a curated common crawl corpus (Wenzek et al., 2019) totaling 32.7 billion unique sentences. Using one unified approach for 38 languages, we were able to mine 3.5 billions parallel sentences, out of which 661 million are aligned with English. 17 language pairs have more then 30 million parallel sentences, 82 more then 10 million, and most more than one million, including direct alignments between many European or Asian languages. To evaluate the quality of the mined bitexts, we train NMT systems for most of the language pairs and evaluate them on TED, WMT and WAT test sets. Using our mined bitexts only and no human translated parallel data, we achieve a new state-of-the-art for a single system on the WMT'19 test set for translation between English and German, Russian and Chinese, as well as German/French. In particular, our English/German system outperforms the best single one by close to 4 BLEU points and is almost on pair with best WMT'19 evaluation system which uses system combination and back-translation. We also achieve excellent results for distant languages pairs like Russian/Japanese, outperforming the best submission at the 2019 workshop on Asian Translation (WAT).
This paper shows that pretraining multilingual language models at scale leads to significant performance gains for a wide range of cross-lingual transfer tasks. We train a Transformer-based masked language model on one hundred languages, using more than two terabytes of filtered CommonCrawl data. Our model, dubbed XLM-R, significantly outperforms multilingual BERT (mBERT) on a variety of cross-lingual benchmarks, including +13.8% average accuracy on XNLI, +12.3% average F1 score on MLQA, and +2.1% average F1 score on NER. XLM-R performs particularly well on low-resource languages, improving 11.8% in XNLI accuracy for Swahili and 9.2% for Urdu over the previous XLM model. We also present a detailed empirical evaluation of the key factors that are required to achieve these gains, including the trade-offs between (1) positive transfer and capacity dilution and (2) the performance of high and low resource languages at scale. Finally, we show, for the first time, the possibility of multilingual modeling without sacrificing per-language performance; XLM-Ris very competitive with strong monolingual models on the GLUE and XNLI benchmarks. We will make XLM-R code, data, and models publicly available.
State of the art sequence-to-sequence models perform a fixed number of computations for each input sequence regardless of whether it is easy or hard to process. In this paper, we train Transformer models which can make output predictions at different stages of the network and we investigate different ways to predict how much computation is required for a particular sequence. Unlike dynamic computation in Universal Transformers, which applies the same set of layers iteratively, we apply different layers at every step to adjust both the amount of computation as well as the model capacity. Experiments on machine translation benchmarks show that this approach can match the accuracy of a baseline Transformer while using only half the number of decoder layers.
In this paper, we focus on the problem of adapting word vector-based models to new textual data. Given a model pre-trained on large reference data, how can we adapt it to a smaller piece of data with a slightly different language distribution? We frame the adaptation problem as a monolingual word vector alignment problem, and simply average models after alignment. We align vectors using the RCSLS criterion. Our formulation results in a simple and efficient algorithm that allows adapting general-purpose models to changing word distributions. In our evaluation, we consider applications to word embedding and text classification models. We show that the proposed approach yields good performance in all setups and outperforms a baseline consisting in fine-tuning the model on new data.
Overparameterized transformer networks have obtained state of the art results in various natural language processing tasks, such as machine translation, language modeling, and question answering. These models contain hundreds of millions of parameters, necessitating a large amount of computation and making them prone to overfitting. In this work, we explore LayerDrop, a form of structured dropout, which has a regularization effect during training and allows for efficient pruning at inference time. In particular, we show that it is possible to select sub-networks of any depth from one large network without having to finetune them and with limited impact on performance. We demonstrate the effectiveness of our approach by improving the state of the art on machine translation, language modeling, summarization, question answering, and language understanding benchmarks. Moreover, we show that our approach leads to small BERT-like models of higher quality compared to training from scratch or using distillation.