Owing to the lack of natural delimiters, Chinese named entity recognition (NER) is more challenging than it in English. While Chinese word segmentation (CWS) is generally regarded as key and open problem for Chinese NER, its accuracy is critical for the downstream models trainings and it also often suffers from out-of-vocabulary (OOV). In this paper, we propose an improved Chinese NER model called TB-LSTM-CRF, which introduces a Transformer Block on top of LSTM-CRF. The proposed model with Transformer Block exploits the self-attention mechanism to capture the information from adjacent characters and sentence contexts. It is more practical with using small-size character embeddings. Compared with the baseline using LSTM-CRF, experiment results show our method TB-LSTM-CRF is competitive without the need of any external resources, for instance other dictionaries.
{"title":"An Improved Chinese Named Entity Recognition Method with TB-LSTM-CRF","authors":"Jiazheng Li, Tao Wang, Weiwen Zhang","doi":"10.1145/3421515.3421534","DOIUrl":"https://doi.org/10.1145/3421515.3421534","url":null,"abstract":"Owing to the lack of natural delimiters, Chinese named entity recognition (NER) is more challenging than it in English. While Chinese word segmentation (CWS) is generally regarded as key and open problem for Chinese NER, its accuracy is critical for the downstream models trainings and it also often suffers from out-of-vocabulary (OOV). In this paper, we propose an improved Chinese NER model called TB-LSTM-CRF, which introduces a Transformer Block on top of LSTM-CRF. The proposed model with Transformer Block exploits the self-attention mechanism to capture the information from adjacent characters and sentence contexts. It is more practical with using small-size character embeddings. Compared with the baseline using LSTM-CRF, experiment results show our method TB-LSTM-CRF is competitive without the need of any external resources, for instance other dictionaries.","PeriodicalId":294293,"journal":{"name":"2020 2nd Symposium on Signal Processing Systems","volume":"7 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126193440","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}
Language identification is quite challenging when it comes to discriminating between closely related dialects of the same language. The fundamental issue is to explore the discriminative cue and effective representation. In this paper, the multi-dimensional language cues are used to distinguish languages, which includes the phonotactic and prosodic information and can be found in the unsupervised setting. Moreover, a novel supervised topic model is proposed to represent and learn the difference of languages. We built the system of language identification and reported the test results on the NIST LRE07 datasets and the Chinese dialect spoken corpus. Compared with other state-of-the-art methods, the experiment results show that the proposed method provides competitive performance and helps to capture robust discriminative information for short duration language identification.
{"title":"Automatic Language Identification using Suprasegmental Feature and Supervised Topic Model","authors":"Linjia Sun","doi":"10.1145/3421515.3421521","DOIUrl":"https://doi.org/10.1145/3421515.3421521","url":null,"abstract":"Language identification is quite challenging when it comes to discriminating between closely related dialects of the same language. The fundamental issue is to explore the discriminative cue and effective representation. In this paper, the multi-dimensional language cues are used to distinguish languages, which includes the phonotactic and prosodic information and can be found in the unsupervised setting. Moreover, a novel supervised topic model is proposed to represent and learn the difference of languages. We built the system of language identification and reported the test results on the NIST LRE07 datasets and the Chinese dialect spoken corpus. Compared with other state-of-the-art methods, the experiment results show that the proposed method provides competitive performance and helps to capture robust discriminative information for short duration language identification.","PeriodicalId":294293,"journal":{"name":"2020 2nd Symposium on Signal Processing Systems","volume":"39 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123711015","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}
Aiming at the problem that the traditional single convolutional neural network cannot completely extract comprehensive text features, this paper proposes a text sentiment classification based on the parallel TCN model of attention mechanism. First, obtain the comprehensive text features with the help of parallel Temporal Convolutional Network (TCN). Secondly, in the feature fusion layer, the features obtained by the parallel TCN are fused. Finally, it combines the attention mechanism to extract important feature information and improve the optimized text sentiment classification effect. And conducted multiple sets of comparative experiments on the two sets of Chinese data sets, the accuracy of the model in this paper reached 92.06% and 92.71%. Proved that the proposed model is better than the traditional single convolutional neural network.
{"title":"Text Sentiment Analysis based on Parallel TCN Model and Attention Model","authors":"Dong Cao, Yujie Huang, Yunbin Fu","doi":"10.1145/3421515.3421524","DOIUrl":"https://doi.org/10.1145/3421515.3421524","url":null,"abstract":"Aiming at the problem that the traditional single convolutional neural network cannot completely extract comprehensive text features, this paper proposes a text sentiment classification based on the parallel TCN model of attention mechanism. First, obtain the comprehensive text features with the help of parallel Temporal Convolutional Network (TCN). Secondly, in the feature fusion layer, the features obtained by the parallel TCN are fused. Finally, it combines the attention mechanism to extract important feature information and improve the optimized text sentiment classification effect. And conducted multiple sets of comparative experiments on the two sets of Chinese data sets, the accuracy of the model in this paper reached 92.06% and 92.71%. Proved that the proposed model is better than the traditional single convolutional neural network.","PeriodicalId":294293,"journal":{"name":"2020 2nd Symposium on Signal Processing Systems","volume":"18 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129862917","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}
Yuxiang Peng, Guoheng Huang, Tao Peng, Lianglun Cheng, Hui-Shi Wu
In recent years, with the increasingly perfect monitoring system, how to make full use of the existing monitoring system to do security work has become a concern in the security field. Face recognition can be used in the security field, but it is difficult to play a role in the surveillance field because it usually requires the cooperation of pedestrians. Therefore, the pedestrian recognition technology without the cooperation of pedestrians has been widely concerned. In this paper, in order to realize a given sequence of monitoring pedestrian images and retrieve pedestrian images across devices, we proposed a new method to realize high- precision pedestrian recognition. First, because surveillance video is a series of pedestrian sequences, we proposed a Crossover Filtering Module (CFM) to screen video sequences for key frames. Then, we propose a network named Multi-frame Fusion Part- based Convolutional Baseline (MFPCB) to extract the features of screened keyframes. Finally, we use the cosine distance to measure the features and find the pedestrian image across the device. This paper mainly studies feature comparison and extraction, which can solve the problems of pedestrian occlusion and location under different cameras. Experiment confirms that MFPCB allows pedestrian recognition to gain another round of performance boost. For instance, on the Mars dataset, we achieve 77.3% mAP and 88.6% rank-1 accuracy, surpassing the state of the art by a large margin.
{"title":"A Pedestrian Re-identification Method Based on Multi-frame Fusion Part-based Convolutional Baseline Network","authors":"Yuxiang Peng, Guoheng Huang, Tao Peng, Lianglun Cheng, Hui-Shi Wu","doi":"10.1145/3421515.3421533","DOIUrl":"https://doi.org/10.1145/3421515.3421533","url":null,"abstract":"In recent years, with the increasingly perfect monitoring system, how to make full use of the existing monitoring system to do security work has become a concern in the security field. Face recognition can be used in the security field, but it is difficult to play a role in the surveillance field because it usually requires the cooperation of pedestrians. Therefore, the pedestrian recognition technology without the cooperation of pedestrians has been widely concerned. In this paper, in order to realize a given sequence of monitoring pedestrian images and retrieve pedestrian images across devices, we proposed a new method to realize high- precision pedestrian recognition. First, because surveillance video is a series of pedestrian sequences, we proposed a Crossover Filtering Module (CFM) to screen video sequences for key frames. Then, we propose a network named Multi-frame Fusion Part- based Convolutional Baseline (MFPCB) to extract the features of screened keyframes. Finally, we use the cosine distance to measure the features and find the pedestrian image across the device. This paper mainly studies feature comparison and extraction, which can solve the problems of pedestrian occlusion and location under different cameras. Experiment confirms that MFPCB allows pedestrian recognition to gain another round of performance boost. For instance, on the Mars dataset, we achieve 77.3% mAP and 88.6% rank-1 accuracy, surpassing the state of the art by a large margin.","PeriodicalId":294293,"journal":{"name":"2020 2nd Symposium on Signal Processing Systems","volume":"16 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116487963","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}
Text classification is a basic and important work in natural language processing (NLP). The existing text classification models are powerful. However, training such a model requires a large number of labeled training sets, but in the actual scene, insufficient data is often faced with. The lack of data is mainly divided into two categories: cold start and low resources. To solve this problem, text enhancement methods are usually used. In this paper, the source text enhancement and representation enhancement are combined to improve the enhancement effect. Five sets of experiments are designed to verify that our method is effective on different data sets and different classifiers. The simulation results show that the accuracy is improved and the generalization ability of the classifier is enhanced to some extent. We also find that the enhancement factor and the size of the training data set are not positively related to the enhancement effect. Therefore, the enhancement factor needs to be selected according to the characteristics of the data.
{"title":"Combined Method Based on Source Text and Representation for Text Enhancement","authors":"Xuelian Li, Weihai Li, Yunxiao Zu","doi":"10.1145/3421515.3421519","DOIUrl":"https://doi.org/10.1145/3421515.3421519","url":null,"abstract":"Text classification is a basic and important work in natural language processing (NLP). The existing text classification models are powerful. However, training such a model requires a large number of labeled training sets, but in the actual scene, insufficient data is often faced with. The lack of data is mainly divided into two categories: cold start and low resources. To solve this problem, text enhancement methods are usually used. In this paper, the source text enhancement and representation enhancement are combined to improve the enhancement effect. Five sets of experiments are designed to verify that our method is effective on different data sets and different classifiers. The simulation results show that the accuracy is improved and the generalization ability of the classifier is enhanced to some extent. We also find that the enhancement factor and the size of the training data set are not positively related to the enhancement effect. Therefore, the enhancement factor needs to be selected according to the characteristics of the data.","PeriodicalId":294293,"journal":{"name":"2020 2nd Symposium on Signal Processing Systems","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130957885","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}
Sergey Zotov, R. Dremliuga, A. Borshevnikov, Ksenia Krivosheeva
We analyzed the developed methods of computer vision in areas associated with recognition and detection of DeepFakes using various models and architectures of neural networks: mainly GAN and CNN. We also discussed the main types and models of these networks that are most effective in detecting and recognizing objects from different data sets, which were provided in the studied articles.
{"title":"DeepFake Detection Algorithms: A Meta-Analysis","authors":"Sergey Zotov, R. Dremliuga, A. Borshevnikov, Ksenia Krivosheeva","doi":"10.1145/3421515.3421532","DOIUrl":"https://doi.org/10.1145/3421515.3421532","url":null,"abstract":"We analyzed the developed methods of computer vision in areas associated with recognition and detection of DeepFakes using various models and architectures of neural networks: mainly GAN and CNN. We also discussed the main types and models of these networks that are most effective in detecting and recognizing objects from different data sets, which were provided in the studied articles.","PeriodicalId":294293,"journal":{"name":"2020 2nd Symposium on Signal Processing Systems","volume":"52 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122986743","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}
Tyre pattern image classification plays an important role in traffic accidents and criminal scene investigation, and it contains rich texture structure information. Classic deep learning models, such as VGG, are often not targeted to represent the texture structure of tyre pattern images, and often cause over-fitting training due to large-scale parameters and insufficient training samples. To improve classification performance of tyre pattern image and solve the model overfitting problem, an efficient tyre pattern image classification model based on multi-scale Gabor convolutional neural network (MS-GCN) is proposed. First, a bank of large-scale directional Gabor filters are used to modulate the convolution kernel to extract more accurate texture features for large-size tyre pattern images, which greatly reduces the number of the training parameters and makes the model more streamlined. Secondly, due to the multi-scale texture similarity of the tyre pattern image, the multi-scale features from different convolutional layers are fused to produce the precise feature representation of the image, following by the optimal feature dimension selection. A large number of experiments were carried out on the real tyre pattern image data set. The results showed that the classification accuracy of the proposed algorithm is 95.9%, which is greatly improved compared with the handcrafted feature extraction algorithm and increased by 17.3% compared with deep learning-based model VGG16. In addition, the classification accuracy of the proposed algorithm on the GHIM-10K data set is 92%, which is also significantly improved compared to other methods. Overall, it shows the effectiveness and superiority of the proposed algorithm.
{"title":"Tyre Pattern Classification Based on Multi-scale GCN Model","authors":"Fuping Wang, Xiaoxia Ding, Y. Liu","doi":"10.1145/3421515.3421520","DOIUrl":"https://doi.org/10.1145/3421515.3421520","url":null,"abstract":"Tyre pattern image classification plays an important role in traffic accidents and criminal scene investigation, and it contains rich texture structure information. Classic deep learning models, such as VGG, are often not targeted to represent the texture structure of tyre pattern images, and often cause over-fitting training due to large-scale parameters and insufficient training samples. To improve classification performance of tyre pattern image and solve the model overfitting problem, an efficient tyre pattern image classification model based on multi-scale Gabor convolutional neural network (MS-GCN) is proposed. First, a bank of large-scale directional Gabor filters are used to modulate the convolution kernel to extract more accurate texture features for large-size tyre pattern images, which greatly reduces the number of the training parameters and makes the model more streamlined. Secondly, due to the multi-scale texture similarity of the tyre pattern image, the multi-scale features from different convolutional layers are fused to produce the precise feature representation of the image, following by the optimal feature dimension selection. A large number of experiments were carried out on the real tyre pattern image data set. The results showed that the classification accuracy of the proposed algorithm is 95.9%, which is greatly improved compared with the handcrafted feature extraction algorithm and increased by 17.3% compared with deep learning-based model VGG16. In addition, the classification accuracy of the proposed algorithm on the GHIM-10K data set is 92%, which is also significantly improved compared to other methods. Overall, it shows the effectiveness and superiority of the proposed algorithm.","PeriodicalId":294293,"journal":{"name":"2020 2nd Symposium on Signal Processing Systems","volume":"19 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131552722","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}
At present, as a direction of intelligent transportation, the research results of car body color detection are still relatively lacking, and the current car body color detection is still easy to be affected by light, shielding, pollution and other factors. This paper proposes a color recognition of vehicle based on low light enhancement and Pixel-wise Contextual Attention, including low light intensity enhancement based on dual Fully Convolutional Networks (FCN), vehicle body detection based on Pixel-wise Contextual Attention Networks (PiCANet), and color classification of vehicle based on Convolutional Neural Network (CNN). The method of low light enhancement has better robustness and adaptability, and can better process the dark image. We use Pixel-wise Contextual Attention Networks, which better identify main area of vehicle with context information. Experiments show that our method is more accurate than the state-of-the-art method with 0.6% under insufficient light.
{"title":"Color Recognition of Vehicle Based on Low Light Enhancement and Pixel-wise Contextual Attention","authors":"Pengkang Zeng, JinTao Zhu, Guoheng Huang, Lianglun Cheng","doi":"10.1145/3421515.3421527","DOIUrl":"https://doi.org/10.1145/3421515.3421527","url":null,"abstract":"At present, as a direction of intelligent transportation, the research results of car body color detection are still relatively lacking, and the current car body color detection is still easy to be affected by light, shielding, pollution and other factors. This paper proposes a color recognition of vehicle based on low light enhancement and Pixel-wise Contextual Attention, including low light intensity enhancement based on dual Fully Convolutional Networks (FCN), vehicle body detection based on Pixel-wise Contextual Attention Networks (PiCANet), and color classification of vehicle based on Convolutional Neural Network (CNN). The method of low light enhancement has better robustness and adaptability, and can better process the dark image. We use Pixel-wise Contextual Attention Networks, which better identify main area of vehicle with context information. Experiments show that our method is more accurate than the state-of-the-art method with 0.6% under insufficient light.","PeriodicalId":294293,"journal":{"name":"2020 2nd Symposium on Signal Processing Systems","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115359225","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}
In machine vision, to correct the distortion of image is required. For improving the performance of the real-time distortion, this paper proposes an algorithm that can compress the inverse mapping table while conduct on-line reconstruction for the inverse mapping table by using interpolation method on FPGA platform in order to overcome the problems that FPGA, when be used to implement algorithm correcting image distortion, will become complexity in the on-line computation of the inverse mapping coordinate and perform insufficient in the capacity of on-chip ROM. The inverse mapping table is used to obtain inverse mapping coordinates that reduce both the amount of on-line computation of FPGA and the need of capacity of on chip ROM. The simulation on MATLAB show the results that when the compression parameters n is 4, 8, or 16, the distortion image can be corrected well and the information will not be lost. A FPGA-based double-sided visual image acquisition platform is built, and the algorithm is tested on the platform. Results show that the proposed algorithm can correct the nonlinear distortion of the image well.
{"title":"Image Correction Based on Field-Programmable Gate Array","authors":"Xinrong Mao, Kaiming Liu","doi":"10.1145/3421515.3421530","DOIUrl":"https://doi.org/10.1145/3421515.3421530","url":null,"abstract":"In machine vision, to correct the distortion of image is required. For improving the performance of the real-time distortion, this paper proposes an algorithm that can compress the inverse mapping table while conduct on-line reconstruction for the inverse mapping table by using interpolation method on FPGA platform in order to overcome the problems that FPGA, when be used to implement algorithm correcting image distortion, will become complexity in the on-line computation of the inverse mapping coordinate and perform insufficient in the capacity of on-chip ROM. The inverse mapping table is used to obtain inverse mapping coordinates that reduce both the amount of on-line computation of FPGA and the need of capacity of on chip ROM. The simulation on MATLAB show the results that when the compression parameters n is 4, 8, or 16, the distortion image can be corrected well and the information will not be lost. A FPGA-based double-sided visual image acquisition platform is built, and the algorithm is tested on the platform. Results show that the proposed algorithm can correct the nonlinear distortion of the image well.","PeriodicalId":294293,"journal":{"name":"2020 2nd Symposium on Signal Processing Systems","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129756728","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 paper proposes a method to quantify the typicality of lexical bundles in sections of academic articles, specifically in the field of Natural Language Processing papers. Typicality is defined as the product of individual KL-divergence scores and the probability of a bundle to appear in a type of section. An evaluation of our typicality measure against two other baselines shows slight improvements according to the Silhouette coefficient.
{"title":"Typicality of Lexical Bundles in Different Sections of Scientific Articles","authors":"Haotong Wang, Y. Lepage, Chooi-Ling Goh","doi":"10.1145/3421515.3421517","DOIUrl":"https://doi.org/10.1145/3421515.3421517","url":null,"abstract":"This paper proposes a method to quantify the typicality of lexical bundles in sections of academic articles, specifically in the field of Natural Language Processing papers. Typicality is defined as the product of individual KL-divergence scores and the probability of a bundle to appear in a type of section. An evaluation of our typicality measure against two other baselines shows slight improvements according to the Silhouette coefficient.","PeriodicalId":294293,"journal":{"name":"2020 2nd Symposium on Signal Processing Systems","volume":"23 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132144344","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：