Error correction techniques have been used to refine the output sentences from automatic speech recognition (ASR) models and achieve a lower word error rate (WER) than original ASR outputs. Previous works usually use a sequence-to-sequence model to correct an ASR output sentence autoregressively, which causes large latency and cannot be deployed in online ASR services. A straightforward solution to reduce latency, inspired by non-autoregressive (NAR) neural machine translation, is to use an NAR sequence generation model for ASR error correction, which, however, comes at the cost of significantly increased ASR error rate. In this paper, observing distinctive error patterns and correction operations (i.e., insertion, deletion, and substitution) in ASR, we propose FastCorrect, a novel NAR error correction model based on edit alignment. In training, FastCorrect aligns each source token from an ASR output sentence to the target tokens from the corresponding ground-truth sentence based on the edit distance between the source and target sentences, and extracts the number of target tokens corresponding to each source token during edition/correction, which is then used to train a length predictor and to adjust the source tokens to match the length of the target sentence for parallel generation. In inference, the token number predicted by the length predictor is used to adjust the source tokens for target sequence generation. Experiments on the public AISHELL-1 dataset and an internal industrial-scale ASR dataset show the effectiveness of FastCorrect for ASR error correction: 1) it speeds up the inference by 6-9 times and maintains the accuracy (8-14% WER reduction) compared with the autoregressive correction model; and 2) it outperforms the popular NAR models adopted in neural machine translation and text edition by a large margin.
Mean opinion score (MOS) is a popular subjective metric to assess the quality of synthesized speech, and usually involves multiple human judges to evaluate each speech utterance. To reduce the labor cost in MOS test, multiple methods have been proposed to automatically predict MOS scores. To our knowledge, for a speech utterance, all previous works only used the average of multiple scores from different judges as the training target and discarded the score of each individual judge, which did not well exploit the precious MOS training data. In this paper, we propose MBNet, a MOS predictor with a mean subnet and a bias subnet to better utilize every judge score in MOS datasets, where the mean subnet is used to predict the mean score of each utterance similar to that in previous works, and the bias subnet to predict the bias score (the difference between the mean score and each individual judge score) and capture the personal preference of individual judges. Experiments show that compared with MOSNet baseline that only leverages mean score for training, MBNet improves the system-level spearmans rank correlation co-efficient (SRCC) by 2.9% on VCC 2018 dataset and 6.7% on VCC 2016 dataset.
Machine teaching uses a meta/teacher model to guide the training of a student model (which will be used in real tasks) through training data selection, loss function design, etc. Previously, the teacher model only takes shallow/surface information as inputs (e.g., training iteration number, loss and accuracy from training/validation sets) while ignoring the internal states of the student model, which limits the potential of learning to teach. In this work, we propose an improved data teaching algorithm, where the teacher model deeply interacts with the student model by accessing its internal states. The teacher model is jointly trained with the student model using meta gradients propagated from a validation set. We conduct experiments on image classification with clean/noisy labels and empirically demonstrate that our algorithm makes significant improvement over previous data teaching methods.
While the multi-branch architecture is one of the key ingredients to the success of computer vision tasks, it has not been well investigated in natural language processing, especially sequence learning tasks. In this work, we propose a simple yet effective variant of Transformer called multi-branch attentive Transformer (briefly, MAT), where the attention layer is the average of multiple branches and each branch is an independent multi-head attention layer. We leverage two training techniques to regularize the training: drop-branch, which randomly drops individual branches during training, and proximal initialization, which uses a pre-trained Transformer model to initialize multiple branches. Experiments on machine translation, code generation and natural language understanding demonstrate that such a simple variant of Transformer brings significant improvements. Our code is available at \url{https://github.com/HA-Transformer}.
Labeling data (e.g., labeling the people, objects, actions and scene in images) comprehensively and efficiently is a widely needed but challenging task. Numerous models were proposed to label various data and many approaches were designed to enhance the ability of deep learning models or accelerate them. Unfortunately, a single machine-learning model is not powerful enough to extract various semantic information from data. Given certain applications, such as image retrieval platforms and photo album management apps, it is often required to execute a collection of models to obtain sufficient labels. With limited computing resources and stringent delay, given a data stream and a collection of applicable resource-hungry deep-learning models, we design a novel approach to adaptively schedule a subset of these models to execute on each data item, aiming to maximize the value of the model output (e.g., the number of high-confidence labels). Achieving this lofty goal is nontrivial since a model's output on any data item is content-dependent and unknown until we execute it. To tackle this, we propose an Adaptive Model Scheduling framework, consisting of 1) a deep reinforcement learning-based approach to predict the value of unexecuted models by mining semantic relationship among diverse models, and 2) two heuristic algorithms to adaptively schedule the model execution order under a deadline or deadline-memory constraints respectively. The proposed framework doesn't require any prior knowledge of the data, which works as a powerful complement to existing model optimization technologies. We conduct extensive evaluations on five diverse image datasets and 30 popular image labeling models to demonstrate the effectiveness of our design: our design could save around 53\% execution time without loss of any valuable labels.
In this paper, we develop a novel weighted Laplacian method, which is partially inspired by the theory of graph Laplacian, to study recent popular graph problems, such as multilevel graph partitioning and balanced minimum cut problem, in a more convenient manner. Since the weighted Laplacian strategy inherits the virtues of spectral methods, graph algorithms designed using weighted Laplacian will necessarily possess more robust theoretical guarantees for algorithmic performances, comparing with those existing algorithms that are heuristically proposed. In order to illustrate its powerful utility both in theory and in practice, we also present two effective applications of our weighted Laplacian method to multilevel graph partitioning and balanced minimum cut problem, respectively. By means of variational methods and theory of partial differential equations (PDEs), we have established the equivalence relations among the weighted cut problem, balanced minimum cut problem and the initial clustering problem that arises in the middle stage of graph partitioning algorithms under a multilevel structure. These equivalence relations can indeed provide solid theoretical support for algorithms based on our proposed weighted Laplacian strategy. Moreover, from the perspective of the application to the balanced minimum cut problem, weighted Laplacian can make it possible for research of numerical solutions of PDEs to be a powerful tool for the algorithmic study of graph problems. Experimental results also indicate that the algorithm embedded with our strategy indeed outperforms other existing graph algorithms, especially in terms of accuracy, thus verifying the efficacy of the proposed weighted Laplacian.
Unsupervised neural machine translation (NMT) has attracted a lot of attention recently. While state-of-the-art methods for unsupervised translation usually perform well between similar languages (e.g., English-German translation), they perform poorly between distant languages, because unsupervised alignment does not work well for distant languages. In this work, we introduce unsupervised pivot translation for distant languages, which translates a language to a distant language through multiple hops, and the unsupervised translation on each hop is relatively easier than the original direct translation. We propose a learning to route (LTR) method to choose the translation path between the source and target languages. LTR is trained on language pairs whose best translation path is available and is applied on the unseen language pairs for path selection. Experiments on 20 languages and 294 distant language pairs demonstrate the advantages of the unsupervised pivot translation for distant languages, as well as the effectiveness of the proposed LTR for path selection. Specifically, in the best case, LTR achieves an improvement of 5.58 BLEU points over the conventional direct unsupervised method.