Combining Self-supervised Learning and Active Learning for Disfluency Detection

Author(s):  
Shaolei Wang ◽  
Zhongyuan Wang ◽  
Wanxiang Che ◽  
Sendong Zhao ◽  
Ting Liu

Spoken language is fundamentally different from the written language in that it contains frequent disfluencies or parts of an utterance that are corrected by the speaker. Disfluency detection (removing these disfluencies) is desirable to clean the input for use in downstream NLP tasks. Most existing approaches to disfluency detection heavily rely on human-annotated data, which is scarce and expensive to obtain in practice. To tackle the training data bottleneck, in this work, we investigate methods for combining self-supervised learning and active learning for disfluency detection. First, we construct large-scale pseudo training data by randomly adding or deleting words from unlabeled data and propose two self-supervised pre-training tasks: (i) a tagging task to detect the added noisy words and (ii) sentence classification to distinguish original sentences from grammatically incorrect sentences. We then combine these two tasks to jointly pre-train a neural network. The pre-trained neural network is then fine-tuned using human-annotated disfluency detection training data. The self-supervised learning method can capture task-special knowledge for disfluency detection and achieve better performance when fine-tuning on a small annotated dataset compared to other supervised methods. However, limited in that the pseudo training data are generated based on simple heuristics and cannot fully cover all the disfluency patterns, there is still a performance gap compared to the supervised models trained on the full training dataset. We further explore how to bridge the performance gap by integrating active learning during the fine-tuning process. Active learning strives to reduce annotation costs by choosing the most critical examples to label and can address the weakness of self-supervised learning with a small annotated dataset. We show that by combining self-supervised learning with active learning, our model is able to match state-of-the-art performance with just about 10% of the original training data on both the commonly used English Switchboard test set and a set of in-house annotated Chinese data.

2019 ◽  
Author(s):  
Yosuke Toda ◽  
Fumio Okura ◽  
Jun Ito ◽  
Satoshi Okada ◽  
Toshinori Kinoshita ◽  
...  

Incorporating deep learning in the image analysis pipeline has opened the possibility of introducing precision phenotyping in the field of agriculture. However, to train the neural network, a sufficient amount of training data must be prepared, which requires a time-consuming manual data annotation process that often becomes the limiting step. Here, we show that an instance segmentation neural network (Mask R-CNN) aimed to phenotype the barley seed morphology of various cultivars, can be sufficiently trained purely by a synthetically generated dataset. Our attempt is based on the concept of domain randomization, where a large amount of image is generated by randomly orienting the seed object to a virtual canvas. After training with such a dataset, performance based on recall and the average Precision of the real-world test dataset achieved 96% and 95%, respectively. Applying our pipeline enables extraction of morphological parameters at a large scale, enabling precise characterization of the natural variation of barley from a multivariate perspective. Importantly, we show that our approach is effective not only for barley seeds but also for various crops including rice, lettuce, oat, and wheat, and thus supporting the fact that the performance benefits of this technique is generic. We propose that constructing and utilizing such synthetic data can be a powerful method to alleviate human labor costs needed to prepare the training dataset for deep learning in the agricultural domain.


Author(s):  
Yi-Quan Li ◽  
Hao-Sen Chang ◽  
Daw-Tung Lin

In the field of computer vision, large-scale image classification tasks are both important and highly challenging. With the ongoing advances in deep learning and optical character recognition (OCR) technologies, neural networks designed to perform large-scale classification play an essential role in facilitating OCR systems. In this study, we developed an automatic OCR system designed to identify up to 13,070 large-scale printed Chinese characters by using deep learning neural networks and fine-tuning techniques. The proposed framework comprises four components, including training dataset synthesis and background simulation, image preprocessing and data augmentation, the process of training the model, and transfer learning. The training data synthesis procedure is composed of a character font generation step and a background simulation process. Three background models are proposed to simulate the factors of the background noise and anti-counterfeiting patterns on ID cards. To expand the diversity of the synthesized training dataset, rotation and zooming data augmentation are applied. A massive dataset comprising more than 19.6 million images was thus created to accommodate the variations in the input images and improve the learning capacity of the CNN model. Subsequently, we modified the GoogLeNet neural architecture by replacing the FC layer with a global average pooling layer to avoid overfitting caused by a massive amount of training data. Consequently, the number of model parameters was reduced. Finally, we employed the transfer learning technique to further refine the CNN model using a small number of real data samples. Experimental results show that the overall recognition performance of the proposed approach is significantly better than that of prior methods and thus demonstrate the effectiveness of proposed framework, which exhibited a recognition accuracy as high as 99.39% on the constructed real ID card dataset.


2020 ◽  
Vol 2020 (10) ◽  
pp. 181-1-181-7
Author(s):  
Takahiro Kudo ◽  
Takanori Fujisawa ◽  
Takuro Yamaguchi ◽  
Masaaki Ikehara

Image deconvolution has been an important issue recently. It has two kinds of approaches: non-blind and blind. Non-blind deconvolution is a classic problem of image deblurring, which assumes that the PSF is known and does not change universally in space. Recently, Convolutional Neural Network (CNN) has been used for non-blind deconvolution. Though CNNs can deal with complex changes for unknown images, some CNN-based conventional methods can only handle small PSFs and does not consider the use of large PSFs in the real world. In this paper we propose a non-blind deconvolution framework based on a CNN that can remove large scale ringing in a deblurred image. Our method has three key points. The first is that our network architecture is able to preserve both large and small features in the image. The second is that the training dataset is created to preserve the details. The third is that we extend the images to minimize the effects of large ringing on the image borders. In our experiments, we used three kinds of large PSFs and were able to observe high-precision results from our method both quantitatively and qualitatively.


2020 ◽  
Vol 27 ◽  
Author(s):  
Zaheer Ullah Khan ◽  
Dechang Pi

Background: S-sulfenylation (S-sulphenylation, or sulfenic acid) proteins, are special kinds of post-translation modification, which plays an important role in various physiological and pathological processes such as cytokine signaling, transcriptional regulation, and apoptosis. Despite these aforementioned significances, and by complementing existing wet methods, several computational models have been developed for sulfenylation cysteine sites prediction. However, the performance of these models was not satisfactory due to inefficient feature schemes, severe imbalance issues, and lack of an intelligent learning engine. Objective: In this study, our motivation is to establish a strong and novel computational predictor for discrimination of sulfenylation and non-sulfenylation sites. Methods: In this study, we report an innovative bioinformatics feature encoding tool, named DeepSSPred, in which, resulting encoded features is obtained via n-segmented hybrid feature, and then the resampling technique called synthetic minority oversampling was employed to cope with the severe imbalance issue between SC-sites (minority class) and non-SC sites (majority class). State of the art 2DConvolutional Neural Network was employed over rigorous 10-fold jackknife cross-validation technique for model validation and authentication. Results: Following the proposed framework, with a strong discrete presentation of feature space, machine learning engine, and unbiased presentation of the underline training data yielded into an excellent model that outperforms with all existing established studies. The proposed approach is 6% higher in terms of MCC from the first best. On an independent dataset, the existing first best study failed to provide sufficient details. The model obtained an increase of 7.5% in accuracy, 1.22% in Sn, 12.91% in Sp and 13.12% in MCC on the training data and12.13% of ACC, 27.25% in Sn, 2.25% in Sp, and 30.37% in MCC on an independent dataset in comparison with 2nd best method. These empirical analyses show the superlative performance of the proposed model over both training and Independent dataset in comparison with existing literature studies. Conclusion : In this research, we have developed a novel sequence-based automated predictor for SC-sites, called DeepSSPred. The empirical simulations outcomes with a training dataset and independent validation dataset have revealed the efficacy of the proposed theoretical model. The good performance of DeepSSPred is due to several reasons, such as novel discriminative feature encoding schemes, SMOTE technique, and careful construction of the prediction model through the tuned 2D-CNN classifier. We believe that our research work will provide a potential insight into a further prediction of S-sulfenylation characteristics and functionalities. Thus, we hope that our developed predictor will significantly helpful for large scale discrimination of unknown SC-sites in particular and designing new pharmaceutical drugs in general.


2020 ◽  
Vol 10 (6) ◽  
pp. 2104
Author(s):  
Michał Tomaszewski ◽  
Paweł Michalski ◽  
Jakub Osuchowski

This article presents an analysis of the effectiveness of object detection in digital images with the application of a limited quantity of input. The possibility of using a limited set of learning data was achieved by developing a detailed scenario of the task, which strictly defined the conditions of detector operation in the considered case of a convolutional neural network. The described solution utilizes known architectures of deep neural networks in the process of learning and object detection. The article presents comparisons of results from detecting the most popular deep neural networks while maintaining a limited training set composed of a specific number of selected images from diagnostic video. The analyzed input material was recorded during an inspection flight conducted along high-voltage lines. The object detector was built for a power insulator. The main contribution of the presented papier is the evidence that a limited training set (in our case, just 60 training frames) could be used for object detection, assuming an outdoor scenario with low variability of environmental conditions. The decision of which network will generate the best result for such a limited training set is not a trivial task. Conducted research suggests that the deep neural networks will achieve different levels of effectiveness depending on the amount of training data. The most beneficial results were obtained for two convolutional neural networks: the faster region-convolutional neural network (faster R-CNN) and the region-based fully convolutional network (R-FCN). Faster R-CNN reached the highest AP (average precision) at a level of 0.8 for 60 frames. The R-FCN model gained a worse AP result; however, it can be noted that the relationship between the number of input samples and the obtained results has a significantly lower influence than in the case of other CNN models, which, in the authors’ assessment, is a desired feature in the case of a limited training set.


2020 ◽  
Author(s):  
Dongyu Xue ◽  
Han Zhang ◽  
Dongling Xiao ◽  
Yukang Gong ◽  
Guohui Chuai ◽  
...  

AbstractIn silico modelling and analysis of small molecules substantially accelerates the process of drug development. Representing and understanding molecules is the fundamental step for various in silico molecular analysis tasks. Traditionally, these molecular analysis tasks have been investigated individually and separately. In this study, we presented X-MOL, which applies large-scale pre-training technology on 1.1 billion molecules for molecular understanding and representation, and then, carefully designed fine-tuning was performed to accommodate diverse downstream molecular analysis tasks, including molecular property prediction, chemical reaction analysis, drug-drug interaction prediction, de novo generation of molecules and molecule optimization. As a result, X-MOL was proven to achieve state-of-the-art results on all these molecular analysis tasks with good model interpretation ability. Collectively, taking advantage of super large-scale pre-training data and super-computing power, our study practically demonstrated the utility of the idea of “mass makes miracles” in molecular representation learning and downstream in silico molecular analysis, indicating the great potential of using large-scale unlabelled data with carefully designed pre-training and fine-tuning strategies to unify existing molecular analysis tasks and substantially enhance the performance of each task.


2020 ◽  
Vol 34 (05) ◽  
pp. 9193-9200
Author(s):  
Shaolei Wang ◽  
Wangxiang Che ◽  
Qi Liu ◽  
Pengda Qin ◽  
Ting Liu ◽  
...  

Most existing approaches to disfluency detection heavily rely on human-annotated data, which is expensive to obtain in practice. To tackle the training data bottleneck, we investigate methods for combining multiple self-supervised tasks-i.e., supervised tasks where data can be collected without manual labeling. First, we construct large-scale pseudo training data by randomly adding or deleting words from unlabeled news data, and propose two self-supervised pre-training tasks: (i) tagging task to detect the added noisy words. (ii) sentence classification to distinguish original sentences from grammatically-incorrect sentences. We then combine these two tasks to jointly train a network. The pre-trained network is then fine-tuned using human-annotated disfluency detection training data. Experimental results on the commonly used English Switchboard test set show that our approach can achieve competitive performance compared to the previous systems (trained using the full dataset) by using less than 1% (1000 sentences) of the training data. Our method trained on the full dataset significantly outperforms previous methods, reducing the error by 21% on English Switchboard.


2021 ◽  
Author(s):  
Noor Ahmad ◽  
Muhammad Aminu ◽  
Mohd Halim Mohd Noor

Deep learning approaches have attracted a lot of attention in the automatic detection of Covid-19 and transfer learning is the most common approach. However, majority of the pre-trained models are trained on color images, which can cause inefficiencies when fine-tuning the models on Covid-19 images which are often grayscale. To address this issue, we propose a deep learning architecture called CovidNet which requires a relatively smaller number of parameters. CovidNet accepts grayscale images as inputs and is suitable for training with limited training dataset. Experimental results show that CovidNet outperforms other state-of-the-art deep learning models for Covid-19 detection.


Author(s):  
Adil Hussain Mohammed

Cloud provide support to manage, control, monitor different organization. Due to flexible nature f cloud chance of attack on it increases by means of some software attack in form of ransomware. Many of researcher has proposed various model to prevent such attacks or to identify such activities. This paper has proposed a ransomware detection model by use of trained neural network. Training of neural network was done by filter or optimized feature set obtained from the feature reduction algorithm. Paper has proposed a Invasive Weed Optimization algorithm that filter good set of feature from the available input training dataset. Proposed model test was performed on real dataset, have set sessions related to cloud ransomware attacks. Result shows that proposed model has increase the comparing parameter values.


Author(s):  
M. Kölle ◽  
V. Walter ◽  
S. Schmohl ◽  
U. Soergel

Abstract. Automated semantic interpretation of 3D point clouds is crucial for many tasks in the domain of geospatial data analysis. For this purpose, labeled training data is required, which has often to be provided manually by experts. One approach to minimize effort in terms of costs of human interaction is Active Learning (AL). The aim is to process only the subset of an unlabeled dataset that is particularly helpful with respect to class separation. Here a machine identifies informative instances which are then labeled by humans, thereby increasing the performance of the machine. In order to completely avoid involvement of an expert, this time-consuming annotation can be resolved via crowdsourcing. Therefore, we propose an approach combining AL with paid crowdsourcing. Although incorporating human interaction, our method can run fully automatically, so that only an unlabeled dataset and a fixed financial budget for the payment of the crowdworkers need to be provided. We conduct multiple iteration steps of the AL process on the ISPRS Vaihingen 3D Semantic Labeling benchmark dataset (V3D) and especially evaluate the performance of the crowd when labeling 3D points. We prove our concept by using labels derived from our crowd-based AL method for classifying the test dataset. The analysis outlines that by labeling only 0:4% of the training dataset by the crowd and spending less than 145 $, both our trained Random Forest and sparse 3D CNN classifier differ in Overall Accuracy by less than 3 percentage points compared to the same classifiers trained on the complete V3D training set.


Sign in / Sign up

Export Citation Format

Share Document