Recent advents in Neural Machine Translation (NMT) have shown improvements in low-resource language (LRL) translation tasks. In this work, we benchmark NMT between English and five African LRL pairs (Swahili, Amharic, Tigrigna, Oromo, Somali [SATOS]). We collected the available resources on the SATOS languages to evaluate the current state of NMT for LRLs. Our evaluation, comparing a baseline single language pair NMT model against semi-supervised learning, transfer learning, and multilingual modeling, shows significant performance improvements both in the En-LRL and LRL-En directions. In terms of averaged BLEU score, the multilingual approach shows the largest gains, up to +5 points, in six out of ten translation directions. To demonstrate the generalization capability of each model, we also report results on multi-domain test sets. We release the standardized experimental data and the test sets for future works addressing the challenges of NMT in under-resourced settings, in particular for the SATOS languages.
Despite the great promise of Transformers in many sequence modeling tasks (e.g., machine translation), their deterministic nature hinders them from generalizing to high entropy tasks such as dialogue response generation. Previous work proposes to capture the variability of dialogue responses with a recurrent neural network (RNN)-based conditional variational autoencoder (CVAE). However, the autoregressive computation of the RNN limits the training efficiency. Therefore, we propose the Variational Transformer (VT), a variational self-attentive feed-forward sequence model. The VT combines the parallelizability and global receptive field of the Transformer with the variational nature of the CVAE by incorporating stochastic latent variables into Transformers. We explore two types of the VT: 1) modeling the discourse-level diversity with a global latent variable; and 2) augmenting the Transformer decoder with a sequence of fine-grained latent variables. Then, the proposed models are evaluated on three conversational datasets with both automatic metric and human evaluation. The experimental results show that our models improve standard Transformers and other baselines in terms of diversity, semantic relevance, and human judgment.
Nigerian Pidgin is arguably the most widely spoken language in Nigeria. Variants of this language are also spoken across West and Central Africa, making it a very important language. This work aims to establish supervised and unsupervised neural machine translation (NMT) baselines between English and Nigerian Pidgin. We implement and compare NMT models with different tokenization methods, creating a solid foundation for future works.
Africa has the highest linguistic diversity in the world. On account of the importance of language to communication, and the importance of reliable, powerful and accurate machine translation models in modern inter-cultural communication, there have been (and still are) efforts to create state-of-the-art translation models for the many African languages. However, the low-resources, diacritical and tonal complexities of African languages are major issues facing African NLP today. The FFR is a major step towards creating a robust translation model from Fon, a very low-resource and tonal language, to French, for research and public use. In this paper, we describe our pilot project: the creation of a large growing corpora for Fon-to-French translations and our FFR v1.0 model, trained on this dataset. The dataset and model are made publicly available.
We report our experiments in building a domain-specific Tigrinya-to-English neural machine translation system. We use transfer learning from other Ge'ez script languages and report an improvement of 1.3 BLEU points over a classic neural baseline. We publish our development pipeline as an open-source library and also provide a demonstration application.
We present Felix --- a flexible text-editing approach for generation, designed to derive the maximum benefit from the ideas of decoding with bi-directional contexts and self-supervised pre-training. In contrast to conventional sequence-to-sequence (seq2seq) models, Felix is efficient in low-resource settings and fast at inference time, while being capable of modeling flexible input-output transformations. We achieve this by decomposing the text-editing task into two sub-tasks: tagging to decide on the subset of input tokens and their order in the output text and insertion to in-fill the missing tokens in the output not present in the input. The tagging model employs a novel Pointer mechanism, while the insertion model is based on a Masked Language Model. Both of these models are chosen to be non-autoregressive to guarantee faster inference. Felix performs favourably when compared to recent text-editing methods and strong seq2seq baselines when evaluated on four NLG tasks: Sentence Fusion, Machine Translation Automatic Post-Editing, Summarization, and Text Simplification.
The standard normalization method for neural network (NN) models used in Natural Language Processing (NLP) is layer normalization (LN). This is different than batch normalization (BN), which is widely-adopted in Computer Vision. The preferred use of LN in NLP is principally due to the empirical observation that a (naive/vanilla) use of BN leads to significant performance degradation for NLP tasks; however, a thorough understanding of the underlying reasons for this is not always evident. In this paper, we perform a systematic study of NLP transformer models to understand why BN has a poor performance, as compared to LN. We find that the statistics of NLP data across the batch dimension exhibit large fluctuations throughout training. This results in instability, if BN is naively implemented. To address this, we propose Power Normalization (PN), a novel normalization scheme that resolves this issue by (i) relaxing zero-mean normalization in BN, (ii) incorporating a running quadratic mean instead of per batch statistics to stabilize fluctuations, and (iii) using an approximate backpropagation for incorporating the running statistics in the forward pass. We show theoretically, under mild assumptions, that PN leads to a smaller Lipschitz constant for the loss, compared with BN. Furthermore, we prove that the approximate backpropagation scheme leads to bounded gradients. We extensively test PN for transformers on a range of NLP tasks, and we show that it significantly outperforms both LN and BN. In particular, PN outperforms LN by 0.4/0.6 BLEU on IWSLT14/WMT14 and 5.6/3.0 PPL on PTB/WikiText-103. We make our code publicly available at \url{https://github.com/sIncerass/powernorm}.
Humans can systematically generalize to novel compositions of existing concepts. Recent studies argue that neural networks appear inherently ineffective in such cognitive capacity, leading to a pessimistic view and a lack of attention to optimistic results. We revisit this controversial topic from the perspective of meaningful learning, an exceptional capability of humans to learn novel concepts by connecting them with known ones. We reassess the compositional skills of sequence-to-sequence models conditioned on the semantic links between new and old concepts. Our observations suggest that models can successfully one-shot generalize to novel concepts and compositions through semantic linking, either inductively or deductively. We demonstrate that prior knowledge plays a key role as well. In addition to synthetic tests, we further conduct proof-of-concept experiments in machine translation and semantic parsing, showing the benefits of meaningful learning in applications. We hope our positive findings will encourage excavating modern neural networks' potential in systematic generalization through more advanced learning schemes.
A typical architecture for end-to-end entity linking systems consists of three steps: mention detection, candidate generation and entity disambiguation. In this study we investigate the following questions: (a) Can all those steps be learned jointly with a model for contextualized text-representations, i.e. BERT (Devlin et al., 2019)? (b) How much entity knowledge is already contained in pretrained BERT? (c) Does additional entity knowledge improve BERT's performance in downstream tasks? To this end, we propose an extreme simplification of the entity linking setup that works surprisingly well: simply cast it as a per token classification over the entire entity vocabulary (over 700K classes in our case). We show on an entity linking benchmark that (i) this model improves the entity representations over plain BERT, (ii) that it outperforms entity linking architectures that optimize the tasks separately and (iii) that it only comes second to the current state-of-the-art that does mention detection and entity disambiguation jointly. Additionally, we investigate the usefulness of entity-aware token-representations in the text-understanding benchmark GLUE, as well as the question answering benchmarks SQUAD V2 and SWAG and also the EN-DE WMT14 machine translation benchmark. To our surprise, we find that most of those benchmarks do not benefit from additional entity knowledge, except for a task with very small training data, the RTE task in GLUE, which improves by 2%.
Several papers argue that wide minima generalize better than narrow minima. In this paper, through detailed experiments that not only corroborate the generalization properties of wide minima, we also provide empirical evidence for a new hypothesis that the density of wide minima is likely lower than the density of narrow minima. Further, motivated by this hypothesis, we design a novel explore-exploit learning rate schedule. On a variety of image and natural language datasets, compared to their original hand-tuned learning rate baselines, we show that our explore-exploit schedule can result in either up to 0.84% higher absolute accuracy using the original training budget or up to 57% reduced training time while achieving the original reported accuracy. For example, we achieve state-of-the-art (SOTA) accuracy for IWSLT'14 (DE-EN) dataset by just modifying the learning rate schedule of a high performing model.
Data-driven segmentation of words into subword units has been used in various natural language processing applications such as automatic speech recognition and statistical machine translation for almost 20 years. Recently it has became more widely adopted, as models based on deep neural networks often benefit from subword units even for morphologically simpler languages. In this paper, we discuss and compare training algorithms for a unigram subword model, based on the Expectation Maximization algorithm and lexicon pruning. Using English, Finnish, North Sami, and Turkish data sets, we show that this approach is able to find better solutions to the optimization problem defined by the Morfessor Baseline model than its original recursive training algorithm. The improved optimization also leads to higher morphological segmentation accuracy when compared to a linguistic gold standard. We publish implementations of the new algorithms in the widely-used Morfessor software package.
In light of the recent breakthroughs in automatic machine translation systems, we propose a novel approach that we term as "Face-to-Face Translation". As today's digital communication becomes increasingly visual, we argue that there is a need for systems that can automatically translate a video of a person speaking in language A into a target language B with realistic lip synchronization. In this work, we create an automatic pipeline for this problem and demonstrate its impact on multiple real-world applications. First, we build a working speech-to-speech translation system by bringing together multiple existing modules from speech and language. We then move towards "Face-to-Face Translation" by incorporating a novel visual module, LipGAN for generating realistic talking faces from the translated audio. Quantitative evaluation of LipGAN on the standard LRW test set shows that it significantly outperforms existing approaches across all standard metrics. We also subject our Face-to-Face Translation pipeline, to multiple human evaluations and show that it can significantly improve the overall user experience for consuming and interacting with multimodal content across languages. Code, models and demo video are made publicly available. Demo video: https://www.youtube.com/watch?v=aHG6Oei8jF0 Code and models: https://github.com/Rudrabha/LipGAN
There have been very few attempts to benchmark performances of state-of-the-art algorithms for Neural Machine Translation task on Indian Languages. Google, Bing, Facebook and Yandex are some of the very few companies which have built translation systems for few of the Indian Languages. Among them, translation results from Google are supposed to be better, based on general inspection. Bing-Translator do not even support Marathi language which has around 95 million speakers and ranks 15th in the world in terms of combined primary and secondary speakers. In this exercise, we trained and compared variety of Neural Machine Marathi to English Translators trained with BERT-tokenizer by huggingface and various Transformer based architectures using Facebook's Fairseq platform with limited but almost correct parallel corpus to achieve better BLEU scores than Google on Tatoeba and Wikimedia open datasets.
Optimization in machine learning, both theoretical and applied, is presently dominated by first-order gradient methods such as stochastic gradient descent. Second-order optimization methods, that involve second derivatives and/or second order statistics of the data, are far less prevalent despite strong theoretical properties, due to their prohibitive computation, memory and communication costs. In an attempt to bridge this gap between theoretical and practical optimization, we present a scalable implementation of a second-order preconditioned method (concretely, a variant of full-matrix Adagrad), that along with several critical algorithmic and numerical improvements, provides significant convergence and wall-clock time improvements compared to conventional first-order methods on state-of-the-art deep models. Our novel design effectively utilizes the prevalent heterogeneous hardware architecture for training deep models, consisting of a multicore CPU coupled with multiple accelerator units. We demonstrate superior performance compared to state-of-the-art on very large learning tasks such as machine translation with Transformers, language modeling with BERT, click-through rate prediction on Criteo, and image classification on ImageNet with ResNet-50.
Transformers have been successfully applied to sequential, auto-regressive tasks despite being feedforward networks. Unlike recurrent neural networks, Transformers use attention to capture temporal relations while processing 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 available. 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, machine translation, and reinforcement learning that the increased representation capacity can create small, shallow models with much stronger performance than comparable Transformers.
Many sequence-to-sequence generation tasks, including machine translation and text-to-speech, can be posed as estimating the density of the output y given the input x: p(y|x). Given this interpretation, it is natural to evaluate sequence-to-sequence models using conditional log-likelihood on a test set. However, the goal of sequence-to-sequence generation (or structured prediction) is to find the best output y^ given an input x, and each task has its own downstream metric R that scores a model output by comparing against a set of references y*: R(y^, y* | x). While we hope that a model that excels in density estimation also performs well on the downstream metric, the exact correlation has not been studied for sequence generation tasks. In this paper, by comparing several density estimators on five machine translation tasks, we find that the correlation between rankings of models based on log-likelihood and BLEU varies significantly depending on the range of the model families being compared. First, log-likelihood is highly correlated with BLEU when we consider models within the same family (e.g. autoregressive models, or latent variable models with the same parameterization of the prior). However, we observe no correlation between rankings of models across different families: (1) among non-autoregressive latent variable models, a flexible prior distribution is better at density estimation but gives worse generation quality than a simple prior, and (2) autoregressive models offer the best translation performance overall, while latent variable models with a normalizing flow prior give the highest held-out log-likelihood across all datasets. Therefore, we recommend using a simple prior for the latent variable non-autoregressive model when fast generation speed is desired.
The recently proposed BERT has shown great power on a variety of natural language understanding tasks, such as text classification, reading comprehension, etc. However, how to effectively apply BERT to neural machine translation (NMT) lacks enough exploration. While BERT is more commonly used as fine-tuning instead of contextual embedding for downstream language understanding tasks, in NMT, our preliminary exploration of using BERT as contextual embedding is better than using for fine-tuning. This motivates us to think how to better leverage BERT for NMT along this direction. We propose a new algorithm named BERT-fused model, in which we first use BERT to extract representations for an input sequence, and then the representations are fused with each layer of the encoder and decoder of the NMT model through attention mechanisms. We conduct experiments on supervised (including sentence-level and document-level translations), semi-supervised and unsupervised machine translation, and achieve state-of-the-art results on seven benchmark datasets. Our code is available at \url{https://github.com/bert-nmt/bert-nmt}.
Though early successes of Statistical Machine Translation (SMT) systems are attributed in part to the explicit modelling of the interaction between any two source and target units, e.g., alignment, the recent Neural Machine Translation (NMT) systems resort to the attention which partially encodes the interaction for efficiency. In this paper, we employ Joint Representation that fully accounts for each possible interaction. We sidestep the inefficiency issue by refining representations with the proposed efficient attention operation. The resulting Reformer models offer a new Sequence-to- Sequence modelling paradigm besides the Encoder-Decoder framework and outperform the Transformer baseline in either the small scale IWSLT14 German-English, English-German and IWSLT15 Vietnamese-English or the large scale NIST12 Chinese-English translation tasks by about 1 BLEU point.We also propose a systematic model scaling approach, allowing the Reformer model to beat the state-of-the-art Transformer in IWSLT14 German-English and NIST12 Chinese-English with about 50% fewer parameters. The code is publicly available at https://github.com/lyy1994/reformer.
To date, most state-of-the-art sequence modeling architectures use attention to build generative models for language based tasks. Some of these models use all the available sequence tokens to generate an attention distribution which results in time complexity of $O(n^2)$. Alternatively, they utilize depthwise convolutions with softmax normalized kernels of size $k$ acting as a limited-window self-attention, resulting in time complexity of $O(k{\cdot}n)$. In this paper, we introduce Time-aware Large Kernel (TaLK) Convolutions, a novel adaptive convolution operation that learns to predict the size of a summation kernel instead of using a fixed-sized kernel matrix. This method yields a time complexity of $O(n)$, effectively making the sequence encoding process linear to the number of tokens. We evaluate the proposed method on large-scale standard machine translation, abstractive summarization and language modeling datasets and show that TaLK Convolutions constitute an efficient improvement over other attention/convolution based approaches.
We present a deep generative model for unsupervised text style transfer that unifies previously proposed non-generative techniques. Our probabilistic approach models non-parallel data from two domains as a partially observed parallel corpus. By hypothesizing a parallel latent sequence that generates each observed sequence, our model learns to transform sequences from one domain to another in a completely unsupervised fashion. In contrast with traditional generative sequence models (e.g. the HMM), our model makes few assumptions about the data it generates: it uses a recurrent language model as a prior and an encoder-decoder as a transduction distribution. While computation of marginal data likelihood is intractable in this model class, we show that amortized variational inference admits a practical surrogate. Further, by drawing connections between our variational objective and other recent unsupervised style transfer and machine translation techniques, we show how our probabilistic view can unify some known non-generative objectives such as backtranslation and adversarial loss. Finally, we demonstrate the effectiveness of our method on a wide range of unsupervised style transfer tasks, including sentiment transfer, formality transfer, word decipherment, author imitation, and related language translation. Across all style transfer tasks, our approach yields substantial gains over state-of-the-art non-generative baselines, including the state-of-the-art unsupervised machine translation techniques that our approach generalizes. Further, we conduct experiments on a standard unsupervised machine translation task and find that our unified approach matches the current state-of-the-art.