Annotated Training Data Research Articles

A Curriculum Learning Approach for Multi-Domain Text Classification Using Keyword Weight Ranking

Text classification is a well-established task in NLP, but it has two major limitations. Firstly, text classification is heavily reliant on domain-specific knowledge, meaning that a classifier that is trained on a given corpus may not perform well when presented with text from another domain. Secondly, text classification models require substantial amounts of annotated data for training, and in certain domains, there may be an insufficient quantity of labeled data available. Consequently, it is essential to explore methods for efficiently utilizing text data from various domains to improve the performance of models across a range of domains. One approach for achieving this is through the use of multi-domain text classification models that leverage adversarial training to extract domain-shared features among all domains as well as the specific features of each domain. After observing the varying distinctness of domain-specific features, our paper introduces a curriculum learning approach using a ranking system based on keyword weight to enhance the effectiveness of multi-domain text classification models. The experimental data from Amazon reviews and FDU-MTL datasets show that our method significantly improves the efficacy of multi-domain text classification models adopting adversarial learning and reaching state-of-the-art outcomes on these two datasets.

Semi-Supervised Modified-UNet for Lung Infection Image Segmentation

Automatic lung infection segmentation in computed tomography (CT) scans can offer great assistance in radiological diagnosis by improving accuracy and reducing time required for diagnosis. The biggest challenges for deep learning (DL) models in segmenting infection region are the high variances in infection characteristics, fuzzy boundaries between infected and normal tissues, and the troubles in getting large number of annotated data for training. To resolve such issues, we propose a Modified U-Net (Mod-UNet) model with minor architectural changes and significant modifications in the training process of vanilla 2D UNet. As part of these modifications, we updated the loss function, optimization function, and regularization methods, added a learning rate scheduler and applied advanced data augmentation techniques. Segmentation results on two Covid-19 Lung CT segmentation datasets show that the performance of Mod-UNet is considerably better than the baseline U-Net. Furthermore, to mitigate the issue of lack of annotated data, the Mod-UNet is used in a semi-supervised framework (Semi-Mod-UNet) which works on a random sampling approach to progressively enlarge the training dataset from a large pool of unannotated CT slices. Exhaustive experiments on the two Covid-19 CT segmentation datasets and on a real lung CT volume show that the Mod-UNet and Semi-Mod-UNet significantly outperform other state-of-theart approaches in automated lung infection segmentation.

ASPER: Answer Set Programming Enhanced Neural Network Models for Joint Entity-Relation Extraction

AbstractA plethora of approaches have been proposed for joint entity-relation (ER) extraction. Most of these methods largely depend on a large amount of manually annotated training data. However, manual data annotation is time-consuming, labor-intensive, and error-prone. Human beings learn using both data (through induction) and knowledge (through deduction). Answer Set Programming (ASP) has been a widely utilized approach for knowledge representation and reasoning that is elaboration tolerant and adept at reasoning with incomplete information. This paper proposes a new approach, ASP-enhanced Entity-Relation extraction (ASPER), to jointly recognize entities and relations by learning from both data and domain knowledge. In particular, ASPER takes advantage of the factual knowledge (represented as facts in ASP) and derived knowledge (represented as rules in ASP) in the learning process of neural network models. We have conducted experiments on two real datasets and compare our method with three baselines. The results show that our ASPER model consistently outperforms the baselines.

Multilingual NLP

The subject area of multilingual natural language processing (NLP) is concerned with the processing of natural language data in several languages. NLP systems that can translate between languages are becoming more and more necessary as the globe gets more interconnected in order to promote understanding and communication among speakers of various languages. To be effective, communication must overcome a number of obstacles presented by multilingual NLP. Lack of language standardization, which results in major variations in the grammatical constructions, vocabulary, and writing systems used in many languages, is one of the fundamental problems. The requirement for substantial amounts of annotated data for machine learning model training presents another difficulty. The creation of high-quality annotated datasets in numerous languages is time- and money-consuming, which restricts the supply of multilingual NLP resources. The problem of creating NLP systems that can handle several languages at once is the last one. This necessitates the deployment of sophisticated algorithms that can handle and evaluate data in numerous languages while producing precise findings. Researchers and developers are working on a variety of methods to address these issues. Creating standardized formats for multilingual data representation, like Universal Dependencies, which offers a unified framework for annotating linguistic data in several languages, is one strategy. Using transfer learning techniques to transfer knowledge from high-resource languages to low-resource languages is an alternative strategy. The amount of annotated data required for training NLP models in low-resource languages can bede creased with the use of this method. Last but not least, researchers are working to create multilingual NLP models that can manage numerous languages at once. To deliver precise results across numerous languages, these models employ cutting-edge methodologies like neural machine translation and multilingual word embedding’s. Despite the fact that multilingual NLP presents a number of difficult issues, with continuing study and development, it is possible to create NLP systems that are capable of processing natural language data from several languages.

Current status of artificial intelligence analysis for the treatment of pancreaticobiliary diseases using endoscopic ultrasonography and endoscopic retrograde cholangiopancreatography.

Pancreatic and biliary diseases encompass a range of conditions requiring accurate diagnosis for appropriate treatment strategies. This diagnosis relies heavily on imaging techniques like endoscopic ultrasonography and endoscopic retrograde cholangiopancreatography. Artificial intelligence (AI), including machine learning and deep learning, is becoming integral in medical imaging and diagnostics, such as the detection of colorectal polyps. AI shows great potential in diagnosing pancreatobiliary diseases. Unlike machine learning, which requires feature extraction and selection, deep learning can utilize images directly as input. Accurate evaluation of AI performance is a complex task due to varied terminologies, evaluation methods, and development stages. Essential aspects of AI evaluation involve defining the AI's purpose, choosing appropriate gold standards, deciding on the validation phase, and selecting reliable validation methods. AI, particularly deep learning, is increasingly employed in endoscopic ultrasonography and endoscopic retrograde cholangiopancreatography diagnostics, achieving high accuracy levels in detecting and classifying various pancreatobiliary diseases. The AI often performs better than doctors, even in tasks like differentiating benign from malignant pancreatic tumors, cysts, and subepithelial lesions, identifying gallbladder lesions, assessing endoscopic retrograde cholangiopancreatography difficulty, and evaluating the biliary strictures. The potential for AI in diagnosing pancreatobiliary diseases, especially where other modalities have limitations, is considerable. However, a crucial constraint is the need for extensive, high-quality annotated data for AI training. Future advances in AI, such as large language models, promise further applications in the medical field.

Automatic License Plate Recognition

Abstract: In current situations many advancements have been made in License Plate (LP) recognition techniques to enhance accuracy, they are often limited to ideal scenarios with accurately annotated training data and restricted situations. Furthermore, monitoring systems frequently employ low-resolution (LR) images or videos. Our project focuses on addressing the challenge of LP detection in digital images captured in naturalistic environments. We aim to improve the quality of LP images by combining character segmentation and recognition with adversarial Super-Resolution (SR) methods. Specifically, we utilize the SRGAN approach, which is capable of processing tiny 72 × 72 LP images. This research investigates the performance of SRGAN across various aspects, including peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). By leveraging these methodologies, we aim to overcome the limitations of existing LP recognition techniques, especially in scenarios where training data is not perfectly annotated and LR images are prevalent. Our approach integrates character segmentation and recognition with SRGAN-based super-resolution techniques to enhance the quality and accuracy of LP images.

A Graph Fusion Approach for Cross-Lingual Machine Reading Comprehension

Although great progress has been made for Machine Reading Comprehension (MRC) in English, scaling out to a large number of languages remains a huge challenge due to the lack of large amounts of annotated training data in non-English languages. To address this challenge, some recent efforts of cross-lingual MRC employ machine translation to transfer knowledge from English to other languages, through either explicit alignment or implicit attention. For effective knowledge transition, it is beneficial to leverage both semantic and syntactic information. However, the existing methods fail to explicitly incorporate syntax information in model learning. Consequently, the models are not robust to errors in alignment and noises in attention. In this work, we propose a novel approach, which jointly models the cross-lingual alignment information and the mono-lingual syntax information using a graph. We develop a series of algorithms, including graph construction, learning, and pre-training. The experiments on two benchmark datasets for cross-lingual MRC show that our approach outperforms all strong baselines, which verifies the effectiveness of syntax information for cross-lingual MRC.

Advancements in Named Entity Recognition using Deep Learning Techniques: A Comprehensive Study on Emerging Trends

Because of the amount of computer-processable text material and processing hardware, deep learning architecture has been the focus of knowledge projects. The task of named entity recognition in natural language processing is no exception. Because of the volume of text material that computers can process and their processing capability, knowledge projects are now focusing on deep learning architecture. Named entity recognition, is a frequently occurring task in natural language processing. Deep learning models, particularly recurrent neural net- works (RNNs) and their variations such as gated recurrent units (GRUs) and long short-term memory (LSTM) , have revolutionised natural language pro- cessing (NER) by capturing intricate contextual connections. These models can accurately forecast the borders and types of named entities because they can effectively capture the sequential and hierarchical patterns inside text. However, using deep learning in NER is a challenging task in itself. Annotated training data is critical for building reliable models, but getting labelled data for all item kinds and domains can be time-consuming. To solve this difficulty, transfer learn- ing and domain adaptation strategies have arisen, using pre-trained models and adapting them to various domains or target tasks. This paper looks at recent deep learning methods for NER and how they evolved from older linear learning approaches. It also analyses the status of activities that are either upstream or downstream of NER, such as sequence tagging and entity linking, among other things

NISNet3D: three-dimensional nuclear synthesis and instance segmentation for fluorescence microscopy images

The primary step in tissue cytometry is the automated distinction of individual cells (segmentation). Since cell borders are seldom labeled, cells are generally segmented by their nuclei. While tools have been developed for segmenting nuclei in two dimensions, segmentation of nuclei in three-dimensional volumes remains a challenging task. The lack of effective methods for three-dimensional segmentation represents a bottleneck in the realization of the potential of tissue cytometry, particularly as methods of tissue clearing present the opportunity to characterize entire organs. Methods based on deep learning have shown enormous promise, but their implementation is hampered by the need for large amounts of manually annotated training data. In this paper, we describe 3D Nuclei Instance Segmentation Network (NISNet3D) that directly segments 3D volumes through the use of a modified 3D U-Net, 3D marker-controlled watershed transform, and a nuclei instance segmentation system for separating touching nuclei. NISNet3D is unique in that it provides accurate segmentation of even challenging image volumes using a network trained on large amounts of synthetic nuclei derived from relatively few annotated volumes, or on synthetic data obtained without annotated volumes. We present a quantitative comparison of results obtained from NISNet3D with results obtained from a variety of existing nuclei segmentation techniques. We also examine the performance of the methods when no ground truth is available and only synthetic volumes were used for training.

Towards fully synthetic training of 3D indoor object detectors: Ablation study

Object detection is a fundamental function of many intelligent autonomous systems such as service robots or autonomous vehicles. State-of-the-art 3D object detection methods are deep learning methods which require large amounts of annotated data for successful training. Acquiring and labeling real-world data is an expensive and time-consuming task. An alternative is to train object detectors using synthetically generated images. Since modern object detectors trained on synthetic data still do not achieve similar results to those trained on real data, generating realistic synthetic data is still an open issue. We address the problem of object detection in indoor environments based on 3D point clouds acquired by 3D cameras. We provide a detailed insight into the importance of five factors involved in generating synthetic data: camera noise, presence of background objects, positioning of objects, scene context, and object sizes. To investigate the importance of these factors, we have developed a fully modular method for generating realistic synthetic single-view point clouds for training object detectors. Our method can generate large amounts of customizable data in a short time. An interesting finding is that pre-training with our data and fine-tuning with real data improves the performance of 3D object detectors, enabling one of them to achieve state-of-the-art results on one of the benchmarks without the use of RGB data.

Less Annotating, More Classifying: Addressing the Data Scarcity Issue of Supervised Machine Learning with Deep Transfer Learning and BERT-NLI

AbstractSupervised machine learning is an increasingly popular tool for analyzing large political text corpora. The main disadvantage of supervised machine learning is the need for thousands of manually annotated training data points. This issue is particularly important in the social sciences where most new research questions require new training data for a new task tailored to the specific research question. This paper analyses how deep transfer learning can help address this challenge by accumulating “prior knowledge” in language models. Models like BERT can learn statistical language patterns through pre-training (“language knowledge”), and reliance on task-specific data can be reduced by training on universal tasks like natural language inference (NLI; “task knowledge”). We demonstrate the benefits of transfer learning on a wide range of eight tasks. Across these eight tasks, our BERT-NLI model fine-tuned on 100 to 2,500 texts performs on average 10.7 to 18.3 percentage points better than classical models without transfer learning. Our study indicates that BERT-NLI fine-tuned on 500 texts achieves similar performance as classical models trained on around 5,000 texts. Moreover, we show that transfer learning works particularly well on imbalanced data. We conclude by discussing limitations of transfer learning and by outlining new opportunities for political science research.

Semisupervised Generative Adversarial Networks With Temporal Convolutions for Human Activity Recognition

Many potential applications of human activity recognition (HAR) can be found in health, surveillance, manufacturing, sports, etc. For instance, HAR can be exploited in ambient assisted living (AAL) systems to provide users with assistance services intend to improve their well-being, safety, and autonomy. Data annotation in HAR is a complex and time-consuming process that limits the availability of labeled samples. Furthermore, the classification performance of supervised deep neural networks depends on the availability of large, annotated training data. This paper proposes an alternative framework for semi-supervised generative adversarial networks (GANs) using temporal convolutions for semi-supervised action recognition to address several problems related to conventional approaches in the HAR context, such as high-dimensionality, scarcity of annotated data, scalability, and robustness. The proposed framework employs a single architecture on different datasets and its effectiveness is tested under four conditions reflecting real-world semi-supervised scenarios to investigate the impact of intersubject training, amount of labeled data, number of classes, and IMU positions on model performance. The evaluations are performed on the PAMAP2, Opportunity-locomotion, and LISSI HAR datasets to evaluate the generalizability of the framework. The results show the proposed framework’s high classification performance and generalization ability compared to baseline methods, achieving up to 25% improvement when only a small amount of annotated data is available. A comparison to the previous work using the same datasets has also validated the performance of the framework.

SPDC: A SUPER-POINT AND POINT COMBINING BASED DUAL-SCALE CONTRASTIVE LEARNING NETWORK FOR POINT CLOUD SEMANTIC SEGMENTATION

Abstract. Semantic segmentation of point clouds is one of the fundamental tasks of point cloud processing and is the basis for other downstream tasks. Deep learning has become the main method to solve point cloud processing. Most existing 3D deep learning models require large amounts of point cloud data to drive them, but annotating the data requires significant time and economic costs. To address the problem of semantic segmentation requiring large amounts of annotated data for training, this paper proposes a Super-point-level and Point-level Dual-scale Contrast learning network (SPDC). To solve the problem that contrastive learning is difficult to train and feature extraction is not sufficient, we introduce super-point maps to assist the network in feature extraction. We use a pre-trained super-point generation network to convert the original point cloud into a super-point map. A dynamic data augmentation(DDA) module is designed for the super-point maps for super-point-level contrastive learning. We map the extracted super-point-level features back to the original point-level scale and conduct secondary contrastive learning with the original point features. The whole feature extraction network is parameter sharing and to reduce the number of parameters we used the lightweight network DGCNN (encoder)+Self-attention as the backbone network. And we did a few-shot pre-training of the backbone network to make the network converge easily. Analogous to CutMix, we designed a new method for point cloud data augmentation called PointObjectMix (POM). This method solves the sample imbalance problem while preserving the overall characteristics of the objects in the scene. We conducted experiments on the S3DIS dataset and obtained 63.3% mIoU. We have also done a large number of ablation experiments to verify the effectiveness of the modules in our method. Experimental results show that our method outperforms the best-unsupervised network available.

TomoTwin: generalized 3D localization of macromolecules in cryo-electron tomograms with structural data mining

Cryogenic-electron tomography enables the visualization of cellular environments in extreme detail, however, tools to analyze the full amount of information contained within these densely packed volumes are still needed. Detailed analysis of macromolecules through subtomogram averaging requires particles to first be localized within the tomogram volume, a task complicated by several factors including a low signal to noise ratio and crowding of the cellular space. Available methods for this task suffer either from being error prone or requiring manual annotation of training data. To assist in this crucial particle picking step, we present TomoTwin: an open source general picking model for cryogenic-electron tomograms based on deep metric learning. By embedding tomograms in an information-rich, high-dimensional space that separates macromolecules according to their three-dimensional structure, TomoTwin allows users to identify proteins in tomograms de novo without manually creating training data or retraining the network to locate new proteins.

SmartWoodID-an image collection of large end-grain surfaces to support wood identification systems.

Wood identification is a key step in the enforcement of laws and regulations aimed at combatting illegal timber trade. Robust wood identification tools, capable of distinguishing a large number of timbers, depend on a solid database of reference material. Reference material for wood identification is typically curated in botanical collections dedicated to wood consisting of samples of secondary xylem of lignified plants. Specimens from the Tervuren Wood Collection, one of the large institutional wood collections around the world, are used as a source of tree species data with potential application as timber. Here, we present SmartWoodID, a database of high-resolution optical scans of the end-grain surfaces enriched with expert wood anatomical descriptions of macroscopic features. These can serve as annotated training data to develop interactive identification keys and artificial intelligence for computer vision-based wood identification. The first edition of the database consists of images of 1190 taxa, with a focus on potential timber species from the Democratic Republic of the Congo with at least four different specimens per species included. Database URL https://hdl.handle.net/20.500.12624/SmartWoodID_first_edition.

Deep learning-based classification of organs at risk and delineation guideline in pelvic cancer radiation therapy.

Deep learning (DL) models for radiation therapy (RT) image segmentation require accurately annotated training data. Multiple organ delineation guidelines exist; however, information on the used guideline is not provided with the delineation. Extraction of training data with coherent guidelines can therefore be challenging. We present a supervised classification method for pelvis structure delineations where bowel cavity, femoral heads, bladder, and rectum data, with two guidelines, were classified. The impact on DL-based segmentation quality using mixed guideline training data was also demonstrated. Bowel cavity was manually delineated on CT images for anal cancer patients (n=170) according to guidelines Devisetty and RTOG. The DL segmentation quality from using training data with coherent or mixed guidelines was investigated. A supervised 3D squeeze-and-excite SENet-154 model was trained to classify two bowel cavity delineation guidelines. In addition, a pelvis CT dataset with manual delineations from prostate cancer patients (n=1854) was used where data with an alternative guideline for femoral heads, rectum, and bladder were generated using commercial software. The model was evaluated on internal (n=200) and external test data (n=99). By using mixed, compared to coherent, delineation guideline training data mean DICE score decreased 3% units, mean Hausdorff distance (95%) increased 5mm and mean surface distance (MSD) increased 1mm. The classification of bowel cavity test data achieved 99.8% unweighted classification accuracy, 99.9% macro average precision, 97.2% macro average recall, and 98.5% macro average F1. Corresponding metrics for the pelvis internal test data were all 99% or above and for the external pelvis test data they were 96.3%, 96.6%, 93.3%, and 94.6%. Impaired segmentation performance was observed for training data with mixed guidelines. The DL delineation classification models achieved excellent results on internal and external test data. This can facilitate automated guideline-specific data extraction while avoiding the need for consistent and correct structure labels.

Quantifying the Simulation–Reality Gap for Deep Learning-Based Drone Detection

The detection of drones or unmanned aerial vehicles is a crucial component in protecting safety-critical infrastructures and maintaining privacy for individuals and organizations. The widespread use of optical sensors for perimeter surveillance has made optical sensors a popular choice for data collection in the context of drone detection. However, efficiently processing the obtained sensor data poses a significant challenge. Even though deep learning-based object detection models have shown promising results, their effectiveness depends on large amounts of annotated training data, which is time consuming and resource intensive to acquire. Therefore, this work investigates the applicability of synthetically generated data obtained through physically realistic simulations based on three-dimensional environments for deep learning-based drone detection. Specifically, we introduce a novel three-dimensional simulation approach built on Unreal Engine and Microsoft AirSim for generating synthetic drone data. Furthermore, we quantify the respective simulation–reality gap and evaluate established techniques for mitigating this gap by systematically exploring different compositions of real and synthetic data. Additionally, we analyze the adaptation of the simulation setup as part of a feedback loop-based training strategy and highlight the benefits of a simulation-based training setup for image-based drone detection, compared to a training strategy relying exclusively on real-world data.

SNER-CS: Self-training Named Entity Recognition in Computer Science

As the number of scientific publications grows, especially in computer science domain (CS), it is important to extract scientific entities from a large number of CS publications. Distantly supervised methods, generating distantly annotated training data by string match with external dictionary automatically, have been widely used in named entity recognition task. However, there are two challenges to use distantly supervised methods in computer science NER task. One is that more and more new tasks, methods and datasets in CS are proposed rapidly, which makes it difficult to build a computer science entity knowledge base with high coverage. The other is noisy annotation, because there is no uniform entity representation standard in computer science domain. To alleviate the two problems above, we propose a novel self-training method based pretraining language model with a distantly supervised label automatic construction system in CS (SNER-CS). Experimental results show that the proposed model SNER-CS performs previous state-of-the-art methods in computer science NER task.

Contrastive self-supervised learning for diabetic retinopathy early detection.

Diabetic Retinopathy (DR) is the major cause of blindness, which seriously threatens the world's vision health. Limited medical resources make early diagnosis and a large-scale screening of DR difficult. Most of the current automatic diagnostic methods are mostly based on deep learning and large-scale labeled data. However, the insufficiency of manual annotations for medical images still is a great challenge of training deep neural networks. Self-supervised learning methods are proposed to learn general features from dataset without manual annotations. Inspired by this, we proposed a deep learning based DR classification model (SimCLR-DR). In this paper, we first use contrastive self-learning algorithm to pre-train the encoder based on convolution network with unlabeled retinal images, then retrain the encoder with classifier on a small annotated training data to detect referable DR. The experimental results on Kaggle dataset show that this proposed method can overcome the training data insufficiency problem and performs better than transfer learning. SimCLR-DR is a good beginning for other deep learning based medical image detection approaches facing the challenge of insufficient annotated data. Figure presents an overview of the proposed framework, which contains three main steps: (i) Data preprocessing; (ii) Pretext task of SimCLR-DR based on contrastive learning; (iii) Downstream Task of SimCLRDR based on CNN.

Artificial intelligence for detection of intracranial haemorrhage on head computed tomography scans: diagnostic accuracy in Hong Kong.

The use of artificial intelligence (AI) to identify acute intracranial haemorrhage (ICH) on computed tomography (CT) scans may facilitate initial imaging interpretation in the accident and emergency department. However, AI model construction requires a large amount of annotated data for training, and validation with real-world data has been limited. We developed an algorithm using an open-access dataset of CT slices, then assessed its utility in clinical practice by validating its performance on CT scans from our institution. Using a publicly available international dataset of >750 000 expert-labelled CT slices, we developed an AI model which determines ICH probability for each CT scan and nominates five potential ICH-positive CT slices for review. We validated the model using retrospective data from 1372 non-contrast head CT scans (84 [6.1%] with ICH) collected at our institution. The model achieved an area under the curve of 0.842 (95% confidence interval=0.791-0.894; P<0.001) for scan-based detection of ICH. A pre-specified probability threshold of ≥50% for the presence of ICH yielded 78.6% accuracy, 73% sensitivity, 79% specificity, 18.6% positive predictive value, and 97.8% negative predictive value. There were 62 true-positive scans and 22 false-negative scans, which could be reduced to six false-negative scans by manual review of model-nominated CT slices. Our model exhibited good accuracy in the CT scan-based detection of ICH, considering the low prevalence of ICH in Hong Kong. Model refinement to allow direct localisation of ICH will facilitate the use of AI solutions in clinical practice.