We introduce a new approach for speech pre-training named SPIRAL which works by learning denoising representation of perturbed data in a teacher-student framework. Specifically, given a speech utterance, we first feed the utterance to a teacher network to obtain corresponding representation. Then the same utterance is perturbed and fed to a student network. The student network is trained to output representation resembling that of the teacher. At the same time, the teacher network is updated as moving average of student's weights over training steps. In order to prevent representation collapse, we apply an in-utterance contrastive loss as pre-training objective and impose position randomization on the input to the teacher. SPIRAL achieves competitive or better results compared to state-of-the-art speech pre-training method wav2vec 2.0, with significant reduction of training cost (80% for Base model, 65% for Large model). Furthermore, we address the problem of noise-robustness that is critical to real-world speech applications. We propose multi-condition pre-training by perturbing the student's input with various types of additive noise. We demonstrate that multi-condition pre-trained SPIRAL models are more robust to noisy speech (9.0% - 13.3% relative word error rate reduction on real noisy test data), compared to applying multi-condition training solely in the fine-tuning stage. The code will be released after publication.
Language-specific pre-trained models have proven to be more accurate than multilingual ones in a monolingual evaluation setting, Arabic is no exception. However, we found that previously released Arabic BERT models were significantly under-trained. In this technical report, we present JABER and SABER, Junior and Senior Arabic BERt respectively, our pre-trained language model prototypes dedicated for Arabic. We conduct an empirical study to systematically evaluate the performance of models across a diverse set of existing Arabic NLU tasks. Experimental results show that JABER and SABER achieve state-of-the-art performances on ALUE, a new benchmark for Arabic Language Understanding Evaluation, as well as on a well-established NER benchmark.
Vast efforts have been devoted to creating high-performance few-shot learners, i.e., models that perform well with little training data. Training large-scale pretrained language models (PLMs) has incurred significant cost, but utilizing PLM-based few-shot learners is still challenging due to their enormous size. This work focuses on a crucial question: How to make effective use of these few-shot learners? We propose LMTurk, a novel approach that treats few-shot learners as crowdsourcing workers. The rationale is that crowdsourcing workers are in fact few-shot learners: They are shown a few illustrative examples to learn about a task and then start annotating. LMTurk employs few-shot learners built upon PLMs as workers. We show that the resulting annotations can be utilized to train models that solve the task well and are small enough to be deployable in practical scenarios. Altogether, LMTurk is an important step towards making effective use of current PLM-based few-shot learners.
Language-specific pre-trained models have proven to be more accurate than multilingual ones in a monolingual evaluation setting, Arabic is no exception. However, we found that previously released Arabic BERT models were significantly under-trained. In this technical report, we present JABER, Junior Arabic BERt, our pretrained language model prototype dedicated for Arabic. We conduct an empirical study to systematically evaluate the performance of models across a diverse set of existing Arabic NLU tasks. Experimental results show that JABER achieves the state-of-the-art performances on ALUE, a new benchmark for Arabic Language Understanding Evaluation, as well as on a well-established NER benchmark
End-to-end models are becoming popular approaches for mispronunciation detection and diagnosis (MDD). A streaming MDD framework which is demanded by many practical applications still remains a challenge. This paper proposes a streaming end-to-end MDD framework called CCA-MDD. CCA-MDD supports online processing and is able to run strictly in real-time. The encoder of CCA-MDD consists of a conv-Transformer network based streaming acoustic encoder and an improved cross-attention named coupled cross-attention (CCA). The coupled cross-attention integrates encoded acoustic features with pre-encoded linguistic features. An ensemble of decoders trained from multi-task learning is applied for final MDD decision. Experiments on publicly available corpora demonstrate that CCA-MDD achieves comparable performance to published offline end-to-end MDD models.
Unsupervised large-scale vision-language pre-training has shown promising advances on various downstream tasks. Existing methods often model the cross-modal interaction either via the similarity of the global feature of each modality which misses sufficient information, or finer-grained interactions using cross/self-attention upon visual and textual tokens. However, cross/self-attention suffers from inferior efficiency in both training and inference. In this paper, we introduce a large-scale Fine-grained Interactive Language-Image Pre-training (FILIP) to achieve finer-level alignment through a cross-modal late interaction mechanism, which uses a token-wise maximum similarity between visual and textual tokens to guide the contrastive objective. FILIP successfully leverages the finer-grained expressiveness between image patches and textual words by modifying only contrastive loss, while simultaneously gaining the ability to pre-compute image and text representations offline at inference, keeping both large-scale training and inference efficient. Furthermore, we construct a new large-scale image-text pair dataset called FILIP300M for pre-training. Experiments show that FILIP achieves state-of-the-art performance on multiple downstream vision-language tasks including zero-shot image classification and image-text retrieval. The visualization on word-patch alignment further shows that FILIP can learn meaningful fine-grained features with promising localization ability.
Point set registration is an essential step in many computer vision applications, such as 3D reconstruction and SLAM. Although there exist many registration algorithms for different purposes, however, this topic is still challenging due to the increasing complexity of various real-world scenarios, such as heavy noise and outlier contamination. In this paper, we propose a novel probabilistic generative method to simultaneously align multiple point sets based on the heavy-tailed Laplacian distribution. The proposed method assumes each data point is generated by a Laplacian Mixture Model (LMM), where its centers are determined by the corresponding points in other point sets. Different from the previous Gaussian Mixture Model (GMM) based method, which minimizes the quadratic distance between points and centers of Gaussian probability density, LMM minimizes the sparsity-induced L1 distance, thereby it is more robust against noise and outliers. We adopt Expectation-Maximization (EM) framework to solve LMM parameters and rigid transformations. We approximate the L1 optimization as a linear programming problem by exponential mapping in Lie algebra, which can be effectively solved through the interior point method. To improve efficiency, we also solve the L1 optimization by Alternating Direction Multiplier Method (ADMM). We demonstrate the advantages of our method by comparing it with representative state-of-the-art approaches on benchmark challenging data sets, in terms of robustness and accuracy.
With the advances in deep learning, tremendous progress has been made with chit-chat dialogue systems and task-oriented dialogue systems. However, these two systems are often tackled separately in current methods. To achieve more natural interaction with humans, a dialogue agent needs to be capable of both chatting and accomplishing tasks. To this end, we propose a unified dialogue system (UniDS) with the two aforementioned skills. In particular, we design a unified dialogue data schema, compatible for both chit-chat and task-oriented dialogues, and we train UniDS with mixed dialogue data from a pretrained chit-chat dialogue model. Without adding extra parameters to SOTA baselines, UniDS can alternatively handle chit-chat and task-oriented dialogues in a unified framework. Experimental results demonstrate that the proposed UniDS works comparably well as the pure chit-chat system, and it outperforms state-of-the-art task-oriented dialogue systems. More importantly, UniDS achieves better robustness as it is able to smoothly switch between two types of dialogues. These results demonstrate the feasibility and potential of building an one-for-all dialogue system.
In recent years, researchers tend to pre-train ever-larger language models to explore the upper limit of deep models. However, large language model pre-training costs intensive computational resources and most of the models are trained from scratch without reusing the existing pre-trained models, which is wasteful. In this paper, we propose bert2BERT, which can effectively transfer the knowledge of an existing smaller pre-trained model (e.g., BERT_BASE) to a large model (e.g., BERT_LARGE) through parameter initialization and significantly improve the pre-training efficiency of the large model. Specifically, we extend the previous function-preserving on Transformer-based language model, and further improve it by proposing advanced knowledge for large model's initialization. In addition, a two-stage pre-training method is proposed to further accelerate the training process. We did extensive experiments on representative PLMs (e.g., BERT and GPT) and demonstrate that (1) our method can save a significant amount of training cost compared with baselines including learning from scratch, StackBERT and MSLT; (2) our method is generic and applicable to different types of pre-trained models. In particular, bert2BERT saves about 45% and 47% computational cost of pre-training BERT_BASE and GPT_BASE by reusing the models of almost their half sizes. The source code will be publicly available upon publication.