Justin Duchatellier, Tyler Holmes, Kun Suo, Yong Shi
Cloud-edge systems are vulnerable to thermal attacks as the increased energy consumption may remain undetected, while occurring alongside normal, CPU-intensive applications. The purpose of our research is to study thermal effects on modern edge systems. We also analyze how performance is affected from the increased heat and identify preventative measures. We speculate that due to the technology being a recent innovation, research on cloud-edge devices and thermal attacks is scarce. Other research focuses on server systems rather than edge platforms. In our paper, we use a Raspberry Pi 4 and a CPU-intensive application to represent thermal attacks on cloud-edge systems. We performed several experiments with the Raspberry Pi 4 and used stress-ng, a benchmarking tool available on Linux distributions, to simulate the attacks. The resulting effects displayed drastic increases in the temperature and power consumption. The key impact of our research is to highlight the following risks and mitigation plans: the vulnerability of cloud-edge systems from thermal attacks, the capability for the attacks to go unnoticed, to further the understanding of edge devices as well as the prevention of these attacks.
云边缘系统很容易受到热攻击，因为在正常的cpu密集型应用程序中，增加的能耗可能不会被检测到。我们的研究目的是研究现代边缘系统的热效应。我们还分析了性能如何受到热量增加的影响，并确定了预防措施。我们推测，由于该技术是最近的创新，对云边缘设备和热攻击的研究很少。其他研究关注的是服务器系统，而不是边缘平台。在我们的论文中，我们使用树莓派4和cpu密集型应用程序来表示对云边缘系统的热攻击。我们在Raspberry Pi 4上进行了几次实验，并使用了Linux发行版上可用的基准测试工具stress-ng来模拟攻击。由此产生的影响显示出温度和功耗的急剧增加。我们研究的主要影响是强调以下风险和缓解计划:云边缘系统在热攻击中的脆弱性，攻击不被注意的能力，进一步了解边缘设备以及预防这些攻击。
{"title":"An empirical study of thermal attacks on edge platforms","authors":"Justin Duchatellier, Tyler Holmes, Kun Suo, Yong Shi","doi":"10.1145/3409334.3452071","DOIUrl":"https://doi.org/10.1145/3409334.3452071","url":null,"abstract":"Cloud-edge systems are vulnerable to thermal attacks as the increased energy consumption may remain undetected, while occurring alongside normal, CPU-intensive applications. The purpose of our research is to study thermal effects on modern edge systems. We also analyze how performance is affected from the increased heat and identify preventative measures. We speculate that due to the technology being a recent innovation, research on cloud-edge devices and thermal attacks is scarce. Other research focuses on server systems rather than edge platforms. In our paper, we use a Raspberry Pi 4 and a CPU-intensive application to represent thermal attacks on cloud-edge systems. We performed several experiments with the Raspberry Pi 4 and used stress-ng, a benchmarking tool available on Linux distributions, to simulate the attacks. The resulting effects displayed drastic increases in the temperature and power consumption. The key impact of our research is to highlight the following risks and mitigation plans: the vulnerability of cloud-edge systems from thermal attacks, the capability for the attacks to go unnoticed, to further the understanding of edge devices as well as the prevention of these attacks.","PeriodicalId":148741,"journal":{"name":"Proceedings of the 2021 ACM Southeast Conference","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129825514","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The Internet of Things (IoT) is the term coined to encompass the myriad of devices that have some data processing and transmitting capabilities. Due to the increasing number of IoT devices connected to the Internet, network protocols intended for IoT technology have gained interest. This paper analyzes the performance of one of the most popular ones, named MQTT (Message Queuing Telemetry Transport), focused on Mosquitto, a widely used implementation. Our principal metric is the transmission time, defined as the time it takes a message to pass from one client through the broker to another client, since MQTT uses a publish/subscribe model with a broker. We evaluate different scenarios against some base configurations to give a firm comparison on how different factors affect the performance of an MQTT system based on Mosquitto, for payload sizes ranging from 512 to 1,048,576 bytes. For example, we assess how different network technologies (Ethernet, WiFi in the 2.4 GHz and 5 GHz bands) and QoS levels may yield better results at different message payload sizes. We also make a broker software comparison, evaluating Mosquitto against ActiveMQ and RabbitMQ. Our experiments exhibited similar results, with a slight advantage for RabbitMQ. Finally, we provide measurements on how DoS attacks can affect the Mosquitto broker, by flooding it with illegal MQTT petitions or making a TCP SYN flood attack. The goal of this study is to help MQTT implementers in making adequate decisions when selecting the different hardware and software solutions, for their MQTT systems.
{"title":"Performance evaluation of a widely used implementation of the MQTT protocol with large payloads in normal operation and under a DoS attack","authors":"Eric Gamess, Trent N. Ford, Monica A. Trifas","doi":"10.1145/3409334.3452067","DOIUrl":"https://doi.org/10.1145/3409334.3452067","url":null,"abstract":"The Internet of Things (IoT) is the term coined to encompass the myriad of devices that have some data processing and transmitting capabilities. Due to the increasing number of IoT devices connected to the Internet, network protocols intended for IoT technology have gained interest. This paper analyzes the performance of one of the most popular ones, named MQTT (Message Queuing Telemetry Transport), focused on Mosquitto, a widely used implementation. Our principal metric is the transmission time, defined as the time it takes a message to pass from one client through the broker to another client, since MQTT uses a publish/subscribe model with a broker. We evaluate different scenarios against some base configurations to give a firm comparison on how different factors affect the performance of an MQTT system based on Mosquitto, for payload sizes ranging from 512 to 1,048,576 bytes. For example, we assess how different network technologies (Ethernet, WiFi in the 2.4 GHz and 5 GHz bands) and QoS levels may yield better results at different message payload sizes. We also make a broker software comparison, evaluating Mosquitto against ActiveMQ and RabbitMQ. Our experiments exhibited similar results, with a slight advantage for RabbitMQ. Finally, we provide measurements on how DoS attacks can affect the Mosquitto broker, by flooding it with illegal MQTT petitions or making a TCP SYN flood attack. The goal of this study is to help MQTT implementers in making adequate decisions when selecting the different hardware and software solutions, for their MQTT systems.","PeriodicalId":148741,"journal":{"name":"Proceedings of the 2021 ACM Southeast Conference","volume":"345 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116487413","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Vaidyanath Areyur Shanthakumar, Clark Barnett, Keith Warnick, P. A. Sudyanti, Vitalii Gerbuz, Tathagata Mukherjee
In this paper we describe a method for computing item based recommendations using matrix-factorization-like embeddings of the items computed using a neural network. Matrix factorizations (MF) compute near optimal item embeddings by minimizing a loss that measures the discrepancy between the predicted and known values of a sparse user-item rating matrix. Though useful for recommendation tasks, they are computationally intensive and hard to compute for large sets of users and items. Hence there is need to compute MF-like embeddings using other less computationally intensive methods, which can be substituted for the actual ones. In this work we explore the possibility of doing the same using a deep neural network (DNN). Our network is trained to learn matrix-factorization-like embeddings from easy to compute natural language processing (NLP) based semantic embeddings. The resulting MF-like embeddings are used to compute recommendations using an anonymized user product engagement dataset from the online retail company Overstock.com. We present the results of using our embeddings for computing recommendations with the Overstock.com production dataset consisting of ~3.5 million items and ~6 million users. Recommendations from Overstock.com's own recommendation system is compared against those obtained by using our MF-like embeddings, by comparing the results from both to the ground truth, which in our case is actual user co-clicks data. Our results show that it is possible to use DNNs for efficiently computing MF-like embeddings which can then be used in conjunction with the NLP based embeddings to improve the recommendations obtained from the NLP based embeddings.
{"title":"Item based recommendation using matrix-factorization-like embeddings from deep networks","authors":"Vaidyanath Areyur Shanthakumar, Clark Barnett, Keith Warnick, P. A. Sudyanti, Vitalii Gerbuz, Tathagata Mukherjee","doi":"10.1145/3409334.3452041","DOIUrl":"https://doi.org/10.1145/3409334.3452041","url":null,"abstract":"In this paper we describe a method for computing item based recommendations using matrix-factorization-like embeddings of the items computed using a neural network. Matrix factorizations (MF) compute near optimal item embeddings by minimizing a loss that measures the discrepancy between the predicted and known values of a sparse user-item rating matrix. Though useful for recommendation tasks, they are computationally intensive and hard to compute for large sets of users and items. Hence there is need to compute MF-like embeddings using other less computationally intensive methods, which can be substituted for the actual ones. In this work we explore the possibility of doing the same using a deep neural network (DNN). Our network is trained to learn matrix-factorization-like embeddings from easy to compute natural language processing (NLP) based semantic embeddings. The resulting MF-like embeddings are used to compute recommendations using an anonymized user product engagement dataset from the online retail company Overstock.com. We present the results of using our embeddings for computing recommendations with the Overstock.com production dataset consisting of ~3.5 million items and ~6 million users. Recommendations from Overstock.com's own recommendation system is compared against those obtained by using our MF-like embeddings, by comparing the results from both to the ground truth, which in our case is actual user co-clicks data. Our results show that it is possible to use DNNs for efficiently computing MF-like embeddings which can then be used in conjunction with the NLP based embeddings to improve the recommendations obtained from the NLP based embeddings.","PeriodicalId":148741,"journal":{"name":"Proceedings of the 2021 ACM Southeast Conference","volume":"23 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124162117","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Stephen Gregory, Utkarsh Singh, Jeffrey G. Gray, Jon Hobbs
Monitoring and tracking inventory is one of the most important aspects of administrating any large-scale enterprise operation that involves physical goods. One of the most evident examples of such operations is automotive manufacturing, especially for servicing a global customer base. We present a software solution of Intelligent Process Automation (IPA) that utilizes state-of-the-art computer vision (CV) and other algorithmic techniques to locate, detect, and manage inventory storage logistics using label information from simple warehouse images. When used in conjunction with a recently developed robotic imaging system, our pipeline can be shown to replace the need for costly, error-prone human input to the inventory tracking system. This paper outlines the technical and practical application of IPA fueled by deep learning. The specific motivation for this project was to address a critical need of Mercedes-Benz U.S. International (MBUSI), but the techniques could be applied more generally to other inventory management contexts. We also discuss how our pipeline produces an inexpensive, efficient, and generalizable solution that provides the capability to retrieve data from an unpredictable environment, in contrast to previous approaches.
{"title":"A computer vision pipeline for automatic large-scale inventory tracking","authors":"Stephen Gregory, Utkarsh Singh, Jeffrey G. Gray, Jon Hobbs","doi":"10.1145/3409334.3452063","DOIUrl":"https://doi.org/10.1145/3409334.3452063","url":null,"abstract":"Monitoring and tracking inventory is one of the most important aspects of administrating any large-scale enterprise operation that involves physical goods. One of the most evident examples of such operations is automotive manufacturing, especially for servicing a global customer base. We present a software solution of Intelligent Process Automation (IPA) that utilizes state-of-the-art computer vision (CV) and other algorithmic techniques to locate, detect, and manage inventory storage logistics using label information from simple warehouse images. When used in conjunction with a recently developed robotic imaging system, our pipeline can be shown to replace the need for costly, error-prone human input to the inventory tracking system. This paper outlines the technical and practical application of IPA fueled by deep learning. The specific motivation for this project was to address a critical need of Mercedes-Benz U.S. International (MBUSI), but the techniques could be applied more generally to other inventory management contexts. We also discuss how our pipeline produces an inexpensive, efficient, and generalizable solution that provides the capability to retrieve data from an unpredictable environment, in contrast to previous approaches.","PeriodicalId":148741,"journal":{"name":"Proceedings of the 2021 ACM Southeast Conference","volume":"35 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127921531","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Md Khurram Monir Rabby, A. Islam, S. Belkasim, M. Bikdash
In this research, a wavelet transform-based feature extraction approach is proposed for the detection of epileptic seizures from the EEG raw dataset. The proposed approach uses the Wavelet Transform (WT) method to divide the seizure and non-seizure classes of signals into multiple sub-bands and extracts the features of the dataset following Petrosian Fractal Dimension (PFD), Higuchi Fractal Dimension (HFD), and Singular Value Decomposition Entropy (SVDE). The Kruskal-Wallis test is performed to determine the difference in the random sampling and the extracted features are leveraged to divide the dataset into the training and testing sets for developing the model in order to train the network. The proposed approach is applied to the EEG dataset of Bonn University. Hence, the Neural Network (NN), Artificial Neural Network (ANN), Support Vector Machine (SVM), and Convolutional Neural Network (CNN) are used as preliminary models in the proposed approach for training the networks. As a preliminary analysis of the proposed approach, the training and testing Area Under the Curve (AUC) is calculated in the Receiver Operating Characteristic (ROC) curve to measure the performances of the existing models. The primary results show that, in the proposed approach, the performance of ANN is better than NN, SVM, and CNN.
在本研究中，提出了一种基于小波变换的特征提取方法，用于从EEG原始数据集中检测癫痫发作。该方法利用小波变换(Wavelet Transform, WT)方法将信号的癫痫和非癫痫类别划分为多个子带，并根据Petrosian分形维数(PFD)、Higuchi分形维数(HFD)和奇异值分解熵(SVDE)提取数据集的特征。通过Kruskal-Wallis测试来确定随机抽样的差异，并利用提取的特征将数据集划分为训练集和测试集，用于开发模型，以训练网络。将该方法应用于德国波恩大学的脑电图数据集。因此，在提出的方法中，神经网络(NN)、人工神经网络(ANN)、支持向量机(SVM)和卷积神经网络(CNN)被用作训练网络的初步模型。作为对所提出方法的初步分析，在Receiver Operating Characteristic (ROC)曲线中计算训练和测试曲线下面积(Area Under the Curve, AUC)来衡量现有模型的性能。初步结果表明，在提出的方法中，神经网络的性能优于神经网络、支持向量机和CNN。
{"title":"Wavelet transform-based feature extraction approach for epileptic seizure classification","authors":"Md Khurram Monir Rabby, A. Islam, S. Belkasim, M. Bikdash","doi":"10.1145/3409334.3452078","DOIUrl":"https://doi.org/10.1145/3409334.3452078","url":null,"abstract":"In this research, a wavelet transform-based feature extraction approach is proposed for the detection of epileptic seizures from the EEG raw dataset. The proposed approach uses the Wavelet Transform (WT) method to divide the seizure and non-seizure classes of signals into multiple sub-bands and extracts the features of the dataset following Petrosian Fractal Dimension (PFD), Higuchi Fractal Dimension (HFD), and Singular Value Decomposition Entropy (SVDE). The Kruskal-Wallis test is performed to determine the difference in the random sampling and the extracted features are leveraged to divide the dataset into the training and testing sets for developing the model in order to train the network. The proposed approach is applied to the EEG dataset of Bonn University. Hence, the Neural Network (NN), Artificial Neural Network (ANN), Support Vector Machine (SVM), and Convolutional Neural Network (CNN) are used as preliminary models in the proposed approach for training the networks. As a preliminary analysis of the proposed approach, the training and testing Area Under the Curve (AUC) is calculated in the Receiver Operating Characteristic (ROC) curve to measure the performances of the existing models. The primary results show that, in the proposed approach, the performance of ANN is better than NN, SVM, and CNN.","PeriodicalId":148741,"journal":{"name":"Proceedings of the 2021 ACM Southeast Conference","volume":"30 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128493138","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Edge computing or Internet of Things (IoT) comprises a set of devices that are interconnected ranging from our daily used objects to advanced networked equipment. It is constantly evolving as the number of devices owned by users is increasing at a rapid speed. These devices are used for various scenarios such as health care, monitoring, autonomous vehicles etc. However, as the edges perform more complex operations and IoTs carry increasing heavy workloads, they demand more energy to perform such tasks. In this paper, we perform a comprehensive study of state-of-the-art energy saving on edge platforms. Specifically, energy efficiency of the devices that run on the edges as well as corresponding solutions including hardware, software, algorithms, etc. will be thoroughly analyzed and we also presented the strengths and weakness of various researches in each area.
{"title":"A study of state-of-the-art energy saving on edges","authors":"Kousalya Banka, Kun Suo, Yong Shi, S. Baidya","doi":"10.1145/3409334.3452079","DOIUrl":"https://doi.org/10.1145/3409334.3452079","url":null,"abstract":"Edge computing or Internet of Things (IoT) comprises a set of devices that are interconnected ranging from our daily used objects to advanced networked equipment. It is constantly evolving as the number of devices owned by users is increasing at a rapid speed. These devices are used for various scenarios such as health care, monitoring, autonomous vehicles etc. However, as the edges perform more complex operations and IoTs carry increasing heavy workloads, they demand more energy to perform such tasks. In this paper, we perform a comprehensive study of state-of-the-art energy saving on edge platforms. Specifically, energy efficiency of the devices that run on the edges as well as corresponding solutions including hardware, software, algorithms, etc. will be thoroughly analyzed and we also presented the strengths and weakness of various researches in each area.","PeriodicalId":148741,"journal":{"name":"Proceedings of the 2021 ACM Southeast Conference","volume":"172 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122361548","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Neuroscience provides a rich source of inspiration for new types of algorithms and architectures to employ when building AI and the resulting biologically-plausible approaches that provide formal, testable models of brain function. The working memory toolkit (WMtk), was developed to assist the integration of an artificial neural network (ANN)-based computational neuroscience model of working memory into reinforcement learning (RL) agents, mitigating the details of ANN design and providing a simple symbolic encoding interface. While the WMtk allows RL agents to perform well in partially-observable domains, it requires prefiltering of sensory information by the programmer: a task often delegated to dimensional attention mechanisms in other cognitive architectures. To fill this gap, we develop and test a biologically-plausible dimensional attention filter for the WMtk and validate model performance using a partially-observable 1D maze task. We show that the attention filter improves learning behavior in two ways by: 1) speeding up learning in the short-term, early in training and 2) developing emergent alternative strategies which optimize performance over the long-term.
{"title":"Benefits of combining dimensional attention and working memory for partially observable reinforcement learning problems","authors":"Ngozi Omatu, Joshua L. Phillips","doi":"10.1145/3409334.3452072","DOIUrl":"https://doi.org/10.1145/3409334.3452072","url":null,"abstract":"Neuroscience provides a rich source of inspiration for new types of algorithms and architectures to employ when building AI and the resulting biologically-plausible approaches that provide formal, testable models of brain function. The working memory toolkit (WMtk), was developed to assist the integration of an artificial neural network (ANN)-based computational neuroscience model of working memory into reinforcement learning (RL) agents, mitigating the details of ANN design and providing a simple symbolic encoding interface. While the WMtk allows RL agents to perform well in partially-observable domains, it requires prefiltering of sensory information by the programmer: a task often delegated to dimensional attention mechanisms in other cognitive architectures. To fill this gap, we develop and test a biologically-plausible dimensional attention filter for the WMtk and validate model performance using a partially-observable 1D maze task. We show that the attention filter improves learning behavior in two ways by: 1) speeding up learning in the short-term, early in training and 2) developing emergent alternative strategies which optimize performance over the long-term.","PeriodicalId":148741,"journal":{"name":"Proceedings of the 2021 ACM Southeast Conference","volume":"346 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116532768","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
This article introduced the modification of an incremental learning algorithm and summarized its performance via the complexity analysis. The algorithm was originally proposed in the context of classic rough set theory, utilizing the hierarchy of probabilistic decision tables as the classifier. The variable precision rough set model (VPRS model) is an extension of the classic rough set theory with unique features. When implemented under the VPRS model, the algorithm has to be modified; for example, some of its strategies can be merged and additional operations are required. Initially, the algorithm was modified into a version specifically suitable for the field of face recognition. This article further reformulated the algorithm so that it can be potentially applied in different areas and, after that, analyzed its complexity.
{"title":"Modification and complexity analysis of an incremental learning algorithm under the VPRS model","authors":"Xuguang Chen","doi":"10.1145/3409334.3452076","DOIUrl":"https://doi.org/10.1145/3409334.3452076","url":null,"abstract":"This article introduced the modification of an incremental learning algorithm and summarized its performance via the complexity analysis. The algorithm was originally proposed in the context of classic rough set theory, utilizing the hierarchy of probabilistic decision tables as the classifier. The variable precision rough set model (VPRS model) is an extension of the classic rough set theory with unique features. When implemented under the VPRS model, the algorithm has to be modified; for example, some of its strategies can be merged and additional operations are required. Initially, the algorithm was modified into a version specifically suitable for the field of face recognition. This article further reformulated the algorithm so that it can be potentially applied in different areas and, after that, analyzed its complexity.","PeriodicalId":148741,"journal":{"name":"Proceedings of the 2021 ACM Southeast Conference","volume":"7 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123744102","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Sentiment analysis is a fascinating area as a natural language understanding benchmark to evaluate customers' feedback and needs. Moreover, sentiment analysis can be applied to understand the people's reactions to public events such as the presidential elections and disease pandemics. Recent works in sentiment analysis on COVID-19 present a domain-targeted Bidirectional Encoder Representations from Transformer (BERT) language model, COVID-Twitter BERT (CT-BERT). However, there is little improvement in text classification using a BERT-based language model directly. Therefore, an auxiliary approach using BERT was proposed. This method converts single-sentence classification into pair-sentence classification, which solves the performance issue of BERT in text classification tasks. In this paper, we combine a pre-trained BERT model from COVID-related tweets and the auxiliary-sentence method to achieve better classification performance on COVID tweets sentiment analysis. We show that converting single-sentence classification into pair-sentence classification extends the dataset and obtains higher accuracies and F1 scores. However, we expect a domain-specific language model would perform better than a general language model. In our results, we show that the performance of CT-BERT does not necessarily outperform BERT specifically in understanding sentiments.
{"title":"Sentiment analysis on COVID tweets using COVID-Twitter-BERT with auxiliary sentence approach","authors":"Hung-Yeh Lin, Teng-Sheng Moh","doi":"10.1145/3409334.3452074","DOIUrl":"https://doi.org/10.1145/3409334.3452074","url":null,"abstract":"Sentiment analysis is a fascinating area as a natural language understanding benchmark to evaluate customers' feedback and needs. Moreover, sentiment analysis can be applied to understand the people's reactions to public events such as the presidential elections and disease pandemics. Recent works in sentiment analysis on COVID-19 present a domain-targeted Bidirectional Encoder Representations from Transformer (BERT) language model, COVID-Twitter BERT (CT-BERT). However, there is little improvement in text classification using a BERT-based language model directly. Therefore, an auxiliary approach using BERT was proposed. This method converts single-sentence classification into pair-sentence classification, which solves the performance issue of BERT in text classification tasks. In this paper, we combine a pre-trained BERT model from COVID-related tweets and the auxiliary-sentence method to achieve better classification performance on COVID tweets sentiment analysis. We show that converting single-sentence classification into pair-sentence classification extends the dataset and obtains higher accuracies and F1 scores. However, we expect a domain-specific language model would perform better than a general language model. In our results, we show that the performance of CT-BERT does not necessarily outperform BERT specifically in understanding sentiments.","PeriodicalId":148741,"journal":{"name":"Proceedings of the 2021 ACM Southeast Conference","volume":"7 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125301135","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Machine Translation is the usage of machine learning techniques in translation from one language to another. It has recently been applied to streaming translation, also known as automatic subtitling. The most common challenge in this area is the trade-off between correctness and speed. Due to its real-time feature, streaming translation needs high speed as it has strict playtime constraints. This paper proposes an enhanced Transformer model for fast streaming translation. The proposed machine-learning method is described, implemented, and evaluated based on a common German-English bilingual dataset. The evaluation results have shown that the proposed system successfully achieved a good speed in the training phase, and a high speed in the actual translating phrase that is fast enough for real-time applications, while also maintaining robust correctness. We believe the proposed Transformer model is a significant contribution to natural-language processing, and would be useful for other real-time translation applications.
{"title":"Fast streaming translation using machine learning with transformer","authors":"Jiabao Qiu, M. Moh, Teng-Sheng Moh","doi":"10.1145/3409334.3452059","DOIUrl":"https://doi.org/10.1145/3409334.3452059","url":null,"abstract":"Machine Translation is the usage of machine learning techniques in translation from one language to another. It has recently been applied to streaming translation, also known as automatic subtitling. The most common challenge in this area is the trade-off between correctness and speed. Due to its real-time feature, streaming translation needs high speed as it has strict playtime constraints. This paper proposes an enhanced Transformer model for fast streaming translation. The proposed machine-learning method is described, implemented, and evaluated based on a common German-English bilingual dataset. The evaluation results have shown that the proposed system successfully achieved a good speed in the training phase, and a high speed in the actual translating phrase that is fast enough for real-time applications, while also maintaining robust correctness. We believe the proposed Transformer model is a significant contribution to natural-language processing, and would be useful for other real-time translation applications.","PeriodicalId":148741,"journal":{"name":"Proceedings of the 2021 ACM Southeast Conference","volume":"73 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130352171","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：