Recent advances in natural language processing (NLP) in online social media are evidently owed to large-scale datasets. However, labeling, storing, and processing a large number of textual data points, e.g., tweets, has remained challenging. On top of that, in applications such as hate speech detection, labeling a sufficiently large dataset containing offensive content can be mentally and emotionally taxing for human annotators. Thus, NLP methods that can make the best use of significantly less labeled data points are of great interest. In this paper, we present a novel pool-based active learning method that can be used for the training of large unlabeled corpus with minimum annotation cost. For that, we propose to find the dominant sets of local clusters in the feature space. These sets represent maximally cohesive structures in the data. Then, the samples that do not belong to any of the dominant sets are selected to be used to train the model, as they represent the boundaries of the local clusters and are more challenging to classify. Our proposed method does not have any parameters to be tuned, making it dataset-independent, and it can approximately achieve the same classification accuracy as full training data, with significantly fewer data points. Additionally, our method achieves a higher performance in comparison to the state-of-the-art active learning strategies. Furthermore, our proposed algorithm is able to incorporate conventional active learning scores, such as uncertainty-based scores, into its selection criteria. We show the effectiveness of our method on different datasets and using different neural network architectures.
Sparse Transformer has recently attracted a lot of attention since the ability for reducing the quadratic dependency on the sequence length. We argue that two factors, information bottleneck sensitivity and inconsistency between different attention topologies, could affect the performance of the Sparse Transformer. This paper proposes a well-designed model named ERNIE-Sparse. It consists of two distinctive parts: (i) Hierarchical Sparse Transformer (HST) to sequentially unify local and global information. (ii) Self-Attention Regularization (SAR) method, a novel regularization designed to minimize the distance for transformers with different attention topologies. To evaluate the effectiveness of ERNIE-Sparse, we perform extensive evaluations. Firstly, we perform experiments on a multi-modal long sequence modeling task benchmark, Long Range Arena (LRA). Experimental results demonstrate that ERNIE-Sparse significantly outperforms a variety of strong baseline methods including the dense attention and other efficient sparse attention methods and achieves improvements by 2.77% (57.78% vs. 55.01%). Secondly, to further show the effectiveness of our method, we pretrain ERNIE-Sparse and verified it on 3 text classification and 2 QA downstream tasks, achieve improvements on classification benchmark by 0.83% (92.46% vs. 91.63%), on QA benchmark by 3.24% (74.67% vs. 71.43%). Experimental results continue to demonstrate its superior performance.
This paper provides an overview of the Omnidirectional Media Format (OMAF) standard, second edition, which has been recently finalized. OMAF specifies the media format for coding, storage, delivery, and rendering of omnidirectional media, including video, audio, images, and timed text. Additionally, OMAF supports multiple viewpoints corresponding to omnidirectional cameras and overlay images or video rendered over the omnidirectional background image or video. Many examples of usage scenarios for multiple viewpoints and overlays are described in the paper. OMAF provides a toolbox of features, which can be selectively used in virtual reality services. Consequently, the paper presents the interoperability points specified in the OMAF standard, which enable signaling which OMAF features are in use or required to be supported in implementations. Finally, the paper summarizes which OMAF interoperability points have been taken into use in virtual reality service specifications by the 3rd Generation Partnership Project (3GPP) and the Virtual Reality Industry Forum (VRIF).
Despite recent advances in neural text generation, encoding the rich diversity in human language remains elusive. We argue that the sub-optimal text generation is mainly attributable to the imbalanced token distribution, which particularly misdirects the learning model when trained with the maximum-likelihood objective. As a simple yet effective remedy, we propose two novel methods, F^2-Softmax and MefMax, for a balanced training even with the skewed frequency distribution. MefMax assigns tokens uniquely to frequency classes, trying to group tokens with similar frequencies and equalize frequency mass between the classes. F^2-Softmax then decomposes a probability distribution of the target token into a product of two conditional probabilities of (i) frequency class, and (ii) token from the target frequency class. Models learn more uniform probability distributions because they are confined to subsets of vocabularies. Significant performance gains on seven relevant metrics suggest the supremacy of our approach in improving not only the diversity but also the quality of generated texts.
We present L3Cube-MahaCorpus a Marathi monolingual data set scraped from different internet sources. We expand the existing Marathi monolingual corpus with 24.8M sentences and 289M tokens. We further present, MahaBERT, MahaAlBERT, and MahaRoBerta all BERT-based masked language models, and MahaFT, the fast text word embeddings both trained on full Marathi corpus with 752M tokens. We show the effectiveness of these resources on downstream classification and NER tasks. Marathi is a popular language in India but still lacks these resources. This work is a step forward in building open resources for the Marathi language. The data and models are available at https://github.com/l3cube-pune/MarathiNLP .
Answering compositional questions that require multiple steps of reasoning against text is challenging, especially when they involve discrete, symbolic operations. Neural module networks (NMNs) learn to parse such questions as executable programs composed of learnable modules, performing well on synthetic visual QA domains. However, we find that it is challenging to learn these models for non-synthetic questions on open-domain text, where a model needs to deal with the diversity of natural language and perform a broader range of reasoning. We extend NMNs by: (a) introducing modules that reason over a paragraph of text, performing symbolic reasoning (such as arithmetic, sorting, counting) over numbers and dates in a probabilistic and differentiable manner; and (b) proposing an unsupervised auxiliary loss to help extract arguments associated with the events in text. Additionally, we show that a limited amount of heuristically-obtained question program and intermediate module output supervision provides sufficient inductive bias for accurate learning. Our proposed model significantly outperforms state-of-the-art models on a subset of the DROP dataset that poses a variety of reasoning challenges that are covered by our modules.
Recent research on super-resolution (SR) has witnessed major developments with the advancements of deep convolutional neural networks. There is a need for information extraction from scenic text images or even document images on device, most of which are low-resolution (LR) images. Therefore, SR becomes an essential pre-processing step as Bicubic Upsampling, which is conventionally present in smartphones, performs poorly on LR images. To give the user more control over his privacy, and to reduce the carbon footprint by reducing the overhead of cloud computing and hours of GPU usage, executing SR models on the edge is a necessity in the recent times. There are various challenges in running and optimizing a model on resource-constrained platforms like smartphones. In this paper, we present a novel deep neural network that reconstructs sharper character edges and thus boosts OCR confidence. The proposed architecture not only achieves significant improvement in PSNR over bicubic upsampling on various benchmark datasets but also runs with an average inference time of 11.7 ms per image. We have outperformed state-of-the-art on the Text330 dataset. We also achieve an OCR accuracy of 75.89% on the ICDAR 2015 TextSR dataset, where ground truth has an accuracy of 78.10%.
Slide presentations are an effective and efficient tool used by the teaching community for classroom communication. However, this teaching model can be challenging for blind and visually impaired (VI) students. The VI student required personal human assistance for understand the presented slide. This shortcoming motivates us to design a Classroom Slide Narration System (CSNS) that generates audio descriptions corresponding to the slide content. This problem poses as an image-to-markup language generation task. The initial step is to extract logical regions such as title, text, equation, figure, and table from the slide image. In the classroom slide images, the logical regions are distributed based on the location of the image. To utilize the location of the logical regions for slide image segmentation, we propose the architecture, Classroom Slide Segmentation Network (CSSN). The unique attributes of this architecture differs from most other semantic segmentation networks. Publicly available benchmark datasets such as WiSe and SPaSe are used to validate the performance of our segmentation architecture. We obtained 9.54 segmentation accuracy improvement in WiSe dataset. We extract content (information) from the slide using four well-established modules such as optical character recognition (OCR), figure classification, equation description, and table structure recognizer. With this information, we build a Classroom Slide Narration System (CSNS) to help VI students understand the slide content. The users have given better feedback on the quality output of the proposed CSNS in comparison to existing systems like Facebooks Automatic Alt-Text (AAT) and Tesseract.
We consider a quantum version of the famous low-rank approximation problem. Specifically, we consider the distance $D(\rho,\sigma)$ between two normalized quantum states, $\rho$ and $\sigma$, where the rank of $\sigma$ is constrained to be at most $R$. For both the trace distance and Hilbert-Schmidt distance, we analytically solve for the optimal state $\sigma$ that minimizes this distance. For the Hilbert-Schmidt distance, the unique optimal state is $\sigma = \tau_R +N_R$, where $\tau_R = \Pi_R \rho \Pi_R$ is given by projecting $\rho$ onto its $R$ principal components with projector $\Pi_R$, and $N_R$ is a normalization factor given by $N_R = \frac{1- \text{Tr}(\tau_R)}{R}\Pi_R$. For the trace distance, this state is also optimal but not uniquely optimal, and we provide the full set of states that are optimal. We briefly discuss how our results have application for performing principal component analysis (PCA) via variational optimization on quantum computers.
Building natural language interfaces for databases has been a long-standing challenge for several decades. The major advantage of these so-called text-to-SQL systems is that end-users can query complex databases without the need to know SQL or the underlying database schema. Due to significant advancements in machine learning, the recent focus of research has been on neural networks to tackle this challenge on complex datasets like Spider. Several recent text-to-SQL systems achieve promising results on this dataset. However, none of them extracts and incorporates values from the user questions for generating SQL statements. Thus, the practical use of these systems in a real-world scenario has not been sufficiently demonstrated yet. In this paper we propose ValueNet light and ValueNet -- the first end-to-end text-to-SQL system incorporating values on the challenging Spider dataset. The main idea of our approach is to use not only metadata information about the underlying database but also information on the base data as input for our neural network architecture. In particular, we propose a novel architecture sketch to extract values from a user question and come up with possible value candidates which are not explicitly mentioned in the question. We then use a neural model based on an encoder-decoder architecture to synthesize the SQL query. Finally, we evaluate our model on the Spider challenge using the Execution Accuracy metric, a more difficult metric than used by most participants of the challenge. Our experimental evaluation demonstrates that ValueNet light and ValueNet reach state-of-the-art results of 64% and 60% accuracy, respectively, for translating from text to SQL, even when applying this more difficult metric than used by previous work.