In this paper, we proposed Adapitch, a multi-speaker TTS method that makes adaptation of the supervised module with untranscribed data. We design two self supervised modules to train the text encoder and mel decoder separately with untranscribed data to enhance the representation of text and mel. To better handle the prosody information in a synthesized voice, a supervised TTS module is designed conditioned on content disentangling of pitch, text, and speaker. The training phase was separated into two parts, pretrained and fixed the text encoder and mel decoder with unsupervised mode, then the supervised mode on the disentanglement of TTS. Experiment results show that the Adaptich achieved much better quality than baseline methods.
Modal verbs (e.g., "can", "should", or "must") occur highly frequently in scientific articles. Decoding their function is not straightforward: they are often used for hedging, but they may also denote abilities and restrictions. Understanding their meaning is important for various NLP tasks such as writing assistance or accurate information extraction from scientific text. To foster research on the usage of modals in this genre, we introduce the MIST (Modals In Scientific Text) dataset, which contains 3737 modal instances in five scientific domains annotated for their semantic, pragmatic, or rhetorical function. We systematically evaluate a set of competitive neural architectures on MIST. Transfer experiments reveal that leveraging non-scientific data is of limited benefit for modeling the distinctions in MIST. Our corpus analysis provides evidence that scientific communities differ in their usage of modal verbs, yet, classifiers trained on scientific data generalize to some extent to unseen scientific domains.
Table answering questions from business documents has many challenges that require understanding tabular structures, cross-document referencing, and additional numeric computations beyond simple search queries. This paper introduces a novel pipeline, named TabIQA, to answer questions about business document images. TabIQA combines state-of-the-art deep learning techniques 1) to extract table content and structural information from images and 2) to answer various questions related to numerical data, text-based information, and complex queries from structured tables. The evaluation results on VQAonBD 2023 dataset demonstrate the effectiveness of TabIQA in achieving promising performance in answering table-related questions. The TabIQA repository is available at https://github.com/phucty/itabqa.
Theory of convolutional neural networks suggests the property of shift equivariance, i.e., that a shifted input causes an equally shifted output. In practice, however, this is not always the case. This poses a great problem for scene text detection for which a consistent spatial response is crucial, irrespective of the position of the text in the scene. Using a simple synthetic experiment, we demonstrate the inherent shift variance of a state-of-the-art fully convolutional text detector. Furthermore, using the same experimental setting, we show how small architectural changes can lead to an improved shift equivariance and less variation of the detector output. We validate the synthetic results using a real-world training schedule on the text detection network. To quantify the amount of shift variability, we propose a metric based on well-established text detection benchmarks. While the proposed architectural changes are not able to fully recover shift equivariance, adding smoothing filters can substantially improve shift consistency on common text datasets. Considering the potentially large impact of small shifts, we propose to extend the commonly used text detection metrics by the metric described in this work, in order to be able to quantify the consistency of text detectors.
Natural Language Processing in the legal domain been benefited hugely by the emergence of Transformer-based Pre-trained Language Models (PLMs) pre-trained on legal text. There exist PLMs trained over European and US legal text, most notably LegalBERT. However, with the rapidly increasing volume of NLP applications on Indian legal documents, and the distinguishing characteristics of Indian legal text, it has become necessary to pre-train LMs over Indian legal text as well. In this work, we introduce transformer-based PLMs pre-trained over a large corpus of Indian legal documents. We also apply these PLMs over several benchmark legal NLP tasks over Indian legal documents, namely, Legal Statute Identification from facts, Semantic segmentation of court judgements, and Court Judgement Prediction. Our experiments demonstrate the utility of the India-specific PLMs developed in this work.
Deep learning inference brings together the data and the Convolutional Neural Network (CNN). This is problematic in case the user wants to preserve the privacy of the data and the service provider does not want to reveal the weights of his CNN. Secure Inference allows the two parties to engage in a protocol that preserves their respective privacy concerns, while revealing only the inference result to the user. This is known as Multi-Party Computation (MPC). A major bottleneck of MPC algorithms is communication, as the parties must send data back and forth. The linear component of a CNN (i.e. convolutions) can be done efficiently with minimal communication, but the non-linear part (i.e., ReLU) requires the bulk of communication bandwidth. We propose two ways to accelerate Secure Inference. The first is based on the observation that the ReLU outcome of many convolutions is highly correlated. Therefore, we replace the per pixel ReLU operation by a ReLU operation per patch. Each layer in the network will benefit from a patch of a different size and we devise an algorithm to choose the optimal set of patch sizes through a novel reduction of the problem to a knapsack problem. The second way to accelerate Secure Inference is based on cutting the number of bit comparisons required for a secure ReLU operation. We demonstrate the cumulative effect of these tools in the semi-honest secure 3-party setting for four problems: Classifying ImageNet using ResNet50 backbone, classifying CIFAR100 using ResNet18 backbone, semantic segmentation of ADE20K using MobileNetV2 backbone and semantic segmentation of Pascal VOC 2012 using ResNet50 backbone. Our source code is publicly available: $\href{https://github.com/yg320/secure_inference}{\text{https://github.com/yg320/secure_inference}}$
Recently, more and more people study online for the convenience of access to massive learning materials (e.g. test questions/notes), thus accurately understanding learning materials became a crucial issue, which is essential for many educational applications. Previous studies focus on using language models to represent the question data. However, test questions (TQ) are usually heterogeneous and multi-modal, e.g., some of them may only contain text, while others half contain images with information beyond their literal description. In this context, both supervised and unsupervised methods are difficult to learn a fused representation of questions. Meanwhile, this problem cannot be solved by conventional methods such as image caption, as the images may contain information complementary rather than duplicate to the text. In this paper, we first improve previous text-only representation with a two-stage unsupervised instance level contrastive based pre-training method (MCL: Mixture Unsupervised Contrastive Learning). Then, TQ-Net was proposed to fuse the content of images to the representation of heterogeneous data. Finally, supervised contrastive learning was conducted on relevance prediction-related downstream tasks, which helped the model to learn the representation of questions effectively. We conducted extensive experiments on question-based tasks on large-scale, real-world datasets, which demonstrated the effectiveness of TQ-Net and improve the precision of downstream applications (e.g. similar questions +2.02% and knowledge point prediction +7.20%). Our code will be available, and we will open-source a subset of our data to promote the development of relative studies.
Background: Encouraged by the success of pretrained Transformer models in many natural language processing tasks, their use for International Classification of Diseases (ICD) coding tasks is now actively being explored. In this study, we investigate three types of Transformer-based models, aiming to address the extreme label set and long text classification challenges that are posed by automated ICD coding tasks. Methods: The Transformer-based model PLM-ICD achieved the current state-of-the-art (SOTA) performance on the ICD coding benchmark dataset MIMIC-III. It was chosen as our baseline model to be further optimised. XR-Transformer, the new SOTA model in the general extreme multi-label text classification domain, and XR-LAT, a novel adaptation of the XR-Transformer model, were also trained on the MIMIC-III dataset. XR-LAT is a recursively trained model chain on a predefined hierarchical code tree with label-wise attention, knowledge transferring and dynamic negative sampling mechanisms. Results: Our optimised PLM-ICD model, which was trained with longer total and chunk sequence lengths, significantly outperformed the current SOTA PLM-ICD model, and achieved the highest micro-F1 score of 60.8%. The XR-Transformer model, although SOTA in the general domain, did not perform well across all metrics. The best XR-LAT based model obtained results that were competitive with the current SOTA PLM-ICD model, including improving the macro-AUC by 2.1%. Conclusion: Our optimised PLM-ICD model is the new SOTA model for automated ICD coding on the MIMIC-III dataset, while our novel XR-LAT model performs competitively with the previous SOTA PLM-ICD model.
Incremental text-to-speech, also known as streaming TTS, has been increasingly applied to online speech applications that require ultra-low response latency to provide an optimal user experience. However, most of the existing speech synthesis pipelines deployed on GPU are still non-incremental, which uncovers limitations in high-concurrency scenarios, especially when the pipeline is built with end-to-end neural network models. To address this issue, we present a highly efficient approach to perform real-time incremental TTS on GPUs with Instant Request Pooling and Module-wise Dynamic Batching. Experimental results demonstrate that the proposed method is capable of producing high-quality speech with a first-chunk latency lower than 80ms under 100 QPS on a single NVIDIA A10 GPU and significantly outperforms the non-incremental twin in both concurrency and latency. Our work reveals the effectiveness of high-performance incremental TTS on GPUs.
We enhance the vanilla adversarial training method for unsupervised Automatic Speech Recognition (ASR) by a diffusion-GAN. Our model (1) injects instance noises of various intensities to the generator's output and unlabeled reference text which are sampled from pretrained phoneme language models with a length constraint, (2) asks diffusion timestep-dependent discriminators to separate them, and (3) back-propagates the gradients to update the generator. Word/phoneme error rate comparisons with wav2vec-U under Librispeech (3.1% for test-clean and 5.6% for test-other), TIMIT and MLS datasets, show that our enhancement strategies work effectively.