With wide-spread usage of style transfer, numerous methods for style transfer draw an increasing attention. Several methods to enhance the efficiency of style transformers have been made, one of them is StarGANv2, a method for multiple-style transfer, which can transform a batch of source pictures into other pictures with different styles. The main difference of StarGANv2 with other style transformers is that it uses style code to represent the styles to enable StarGANv2 to complete multiple-style transformation. The authors of StarGANv2 use CelebA-HQ and AFHQ dataset to train the model and test the model, and the results are pretty better than other style transformers. The goal of this paper is to exploit the effectiveness of StarGANv2 in the real-world scenes, such as over exposure or the angle facing the camera. The results validate the power of StarGANv2 where the model is robust enough to transfer the pictures into other styles. To achieve this, the authors of StarGANv2 use the photo clipped in videos which record real-world animals and form a new dataset. Then, the authors of StarGANv2 use the dataset to test the pre-trained model which is trained by AFHQ dataset and evaluate it according to FID metric. The authors of StarGANv2 draw a conclusion that StarGANv2 is robust in real world scenes. The meaning of this paper is that the authors get the real-world usage of StarGANv2 and have a test of StarGANv2’s robustness in real world photos and validate the potential of StarGANv2 in real-world applications.
{"title":"Exploiting the power of StarGANv2 in the wild","authors":"Gengjun Huang, Xiaosheng Long, Yiming Mao","doi":"10.1117/12.2671374","DOIUrl":"https://doi.org/10.1117/12.2671374","url":null,"abstract":"With wide-spread usage of style transfer, numerous methods for style transfer draw an increasing attention. Several methods to enhance the efficiency of style transformers have been made, one of them is StarGANv2, a method for multiple-style transfer, which can transform a batch of source pictures into other pictures with different styles. The main difference of StarGANv2 with other style transformers is that it uses style code to represent the styles to enable StarGANv2 to complete multiple-style transformation. The authors of StarGANv2 use CelebA-HQ and AFHQ dataset to train the model and test the model, and the results are pretty better than other style transformers. The goal of this paper is to exploit the effectiveness of StarGANv2 in the real-world scenes, such as over exposure or the angle facing the camera. The results validate the power of StarGANv2 where the model is robust enough to transfer the pictures into other styles. To achieve this, the authors of StarGANv2 use the photo clipped in videos which record real-world animals and form a new dataset. Then, the authors of StarGANv2 use the dataset to test the pre-trained model which is trained by AFHQ dataset and evaluate it according to FID metric. The authors of StarGANv2 draw a conclusion that StarGANv2 is robust in real world scenes. The meaning of this paper is that the authors get the real-world usage of StarGANv2 and have a test of StarGANv2’s robustness in real world photos and validate the potential of StarGANv2 in real-world applications.","PeriodicalId":227528,"journal":{"name":"International Conference on Artificial Intelligence and Computer Engineering (ICAICE 2022)","volume":"21 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127453153","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}
Human-Object Interaction (HOI) detection is a fundamental task for understanding real-world scenes. In this paper, a graph model-based human-object interaction detection algorithm is proposed, which aims to make full use of the visual-spatial features and semantic information of human-object instances in the image, thereby improving the accuracy of interaction detection. Aiming at the characteristics of visual-spatial features and semantic information, we take the visual features of human and object instance boxes as nodes, and the corresponding spatial features of interaction relations as edges to construct an initial dense graph, and adaptively update the graph through the spatial and semantic information of instances. The V-COCO dataset is used to evaluate the algorithm, and the final accuracy is significantly improved, which proves the effectiveness of the algorithm.
{"title":"Human-object interaction detection based on graph model","authors":"Qing Ye, Xiujuan Xu","doi":"10.1117/12.2671248","DOIUrl":"https://doi.org/10.1117/12.2671248","url":null,"abstract":"Human-Object Interaction (HOI) detection is a fundamental task for understanding real-world scenes. In this paper, a graph model-based human-object interaction detection algorithm is proposed, which aims to make full use of the visual-spatial features and semantic information of human-object instances in the image, thereby improving the accuracy of interaction detection. Aiming at the characteristics of visual-spatial features and semantic information, we take the visual features of human and object instance boxes as nodes, and the corresponding spatial features of interaction relations as edges to construct an initial dense graph, and adaptively update the graph through the spatial and semantic information of instances. The V-COCO dataset is used to evaluate the algorithm, and the final accuracy is significantly improved, which proves the effectiveness of the algorithm.","PeriodicalId":227528,"journal":{"name":"International Conference on Artificial Intelligence and Computer Engineering (ICAICE 2022)","volume":"20 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126913547","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 painted murals in Mogao Grottoes and Longmen Grottoes are symbols of China history and culture. However, most of the murals with complex texture and structure have suffered from different degrees of disease erosion after thousands of years. It is necessary to restore the damaged parts of the murals and to accurately restore their contents. In recent years, the use of new virtual technologies such as digital images to repair the damage can largely avoid secondary damage to the murals caused by manual restoration methods. Therefore, this paper takes the restoration of the most typical shedding diseases to the Mogao Caves murals in Dunhuang as an example. Furthermore, the research object of this paper is the shedding diseases including contour lines. For the traditional virtual methods of repairing shedding diseases, the structure and texture are usually restored at the same time, and these methods have little effect on the accurate removal of shedding disease through the contour line. It can be seen that shedding disease through the contour line is more difficult to repair, and more appropriate inpainting methods need to be explored. Considering the particularity of the shedding disease that passes through the contour line, this paper proposes a mural image inpainting algorithm based on structure priority to repair the shedding diseases. First, the structure repair problem is further converted into a optimization problem, and then the global optimization capability of the genetic algorithm is used to realize the connection of the structure information of the damaged area. Then, the texture is filled by subarea optimization to obtain an ideal repair effect, which can reasonably and effectively solve the problem of shedding disease repair through the contour line. Subjective and objective evaluation of experimental results is also better than other comparative methods.
{"title":"Mural image shedding diseases inpainting algorithm based on structure priority","authors":"Haibo Pen, Shuangshuang Wang, Zhuofan Zhang","doi":"10.1117/12.2671230","DOIUrl":"https://doi.org/10.1117/12.2671230","url":null,"abstract":"The painted murals in Mogao Grottoes and Longmen Grottoes are symbols of China history and culture. However, most of the murals with complex texture and structure have suffered from different degrees of disease erosion after thousands of years. It is necessary to restore the damaged parts of the murals and to accurately restore their contents. In recent years, the use of new virtual technologies such as digital images to repair the damage can largely avoid secondary damage to the murals caused by manual restoration methods. Therefore, this paper takes the restoration of the most typical shedding diseases to the Mogao Caves murals in Dunhuang as an example. Furthermore, the research object of this paper is the shedding diseases including contour lines. For the traditional virtual methods of repairing shedding diseases, the structure and texture are usually restored at the same time, and these methods have little effect on the accurate removal of shedding disease through the contour line. It can be seen that shedding disease through the contour line is more difficult to repair, and more appropriate inpainting methods need to be explored. Considering the particularity of the shedding disease that passes through the contour line, this paper proposes a mural image inpainting algorithm based on structure priority to repair the shedding diseases. First, the structure repair problem is further converted into a optimization problem, and then the global optimization capability of the genetic algorithm is used to realize the connection of the structure information of the damaged area. Then, the texture is filled by subarea optimization to obtain an ideal repair effect, which can reasonably and effectively solve the problem of shedding disease repair through the contour line. Subjective and objective evaluation of experimental results is also better than other comparative methods.","PeriodicalId":227528,"journal":{"name":"International Conference on Artificial Intelligence and Computer Engineering (ICAICE 2022)","volume":"5 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126191158","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}
Sarcasm is a special kind of linguistic sentiment that is widely used in a wide range of social media to express strong emotions in users. Therefore, the task of sarcasm recognition is particularly important for social media analysis. There are few studies on sarcasm sentiment recognition in Chinese, and they often ignore the complex interactions between different syntactic components of a sentence, such as sentiment words, entities, dummy words, and special punctuation that occur in the text. In order to improve the accuracy of Chinese sarcasm recognition, this paper proposes a multi-scale neural network sarcasm recognition algorithm incorporating a hierarchical representation of sentences, taking into account the semantic information of sentences and the relationship features between different syntactic components. The hierarchical syntactic tree is reconstructed to distinguish the key components of the sentence, and the multi-channel convolutional network is used to mine the relational features between syntactic levels and deeply fuse them with semantic information to perform the Chinese sarcastic sentiment recognition task. We have tested the method on a publicly available Chinese sarcastic comment dataset, and the results show that the method can effectively improve the accuracy rate of Chinese sarcastic sentiment recognition.
{"title":"A multi-scale sarcasm sentiment recognition algorithm incorporating sentence hierarchical representation","authors":"Yurong Hao, Long Zhang, Qiusheng Zheng, Liyue Niu","doi":"10.1117/12.2671064","DOIUrl":"https://doi.org/10.1117/12.2671064","url":null,"abstract":"Sarcasm is a special kind of linguistic sentiment that is widely used in a wide range of social media to express strong emotions in users. Therefore, the task of sarcasm recognition is particularly important for social media analysis. There are few studies on sarcasm sentiment recognition in Chinese, and they often ignore the complex interactions between different syntactic components of a sentence, such as sentiment words, entities, dummy words, and special punctuation that occur in the text. In order to improve the accuracy of Chinese sarcasm recognition, this paper proposes a multi-scale neural network sarcasm recognition algorithm incorporating a hierarchical representation of sentences, taking into account the semantic information of sentences and the relationship features between different syntactic components. The hierarchical syntactic tree is reconstructed to distinguish the key components of the sentence, and the multi-channel convolutional network is used to mine the relational features between syntactic levels and deeply fuse them with semantic information to perform the Chinese sarcastic sentiment recognition task. We have tested the method on a publicly available Chinese sarcastic comment dataset, and the results show that the method can effectively improve the accuracy rate of Chinese sarcastic sentiment recognition.","PeriodicalId":227528,"journal":{"name":"International Conference on Artificial Intelligence and Computer Engineering (ICAICE 2022)","volume":"12 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126069686","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}
Nowadays, target recognition, driverless, medical impact diagnosis, and other applications based on image recognition in life, scientific research, and work, rely mainly on a variety of large models with excellent performance, from the Convolutional Neural Network (CNN) at the beginning to the various variants of the classical model proposed now. In this paper, we will take the example of identifying catamount and canid datasets, comparing the efficiency and accuracy of CNN, Vision Transformer (ViT), and Swin Transformer laterally. We plan to run 25 epochs for each model and record the accuracy and time consumption separately. After the experiments we find that from the comparison of the epoch numbers and the real-time consumption, the CNN takes the least total time, followed by Swin Transformer. Also, ViT takes the least time to reach convergence, while Swin Transformer takes the most time. In terms of training accuracy, ViT has the highest training accuracy, followed by Swin Transformer, and CNN has the lowest training accuracy; the validation accuracy is similar to the training accuracy. ViT has the highest accuracy, but takes the longest time; conversely, CNN takes the shortest time and has the lowest accuracy. Swin Transformer, which seems a combination of CNN and ViT, is most complex but with ideal performance. In the future, ViT is indeed a promising model that deserves further research and exploration to contribute to the computer vision field.
{"title":"Comparison and analysis of computer vision models based on images of catamount and canid","authors":"Feng Jiang, Yueyufei Ma, Langyue Wang","doi":"10.1117/12.2671468","DOIUrl":"https://doi.org/10.1117/12.2671468","url":null,"abstract":"Nowadays, target recognition, driverless, medical impact diagnosis, and other applications based on image recognition in life, scientific research, and work, rely mainly on a variety of large models with excellent performance, from the Convolutional Neural Network (CNN) at the beginning to the various variants of the classical model proposed now. In this paper, we will take the example of identifying catamount and canid datasets, comparing the efficiency and accuracy of CNN, Vision Transformer (ViT), and Swin Transformer laterally. We plan to run 25 epochs for each model and record the accuracy and time consumption separately. After the experiments we find that from the comparison of the epoch numbers and the real-time consumption, the CNN takes the least total time, followed by Swin Transformer. Also, ViT takes the least time to reach convergence, while Swin Transformer takes the most time. In terms of training accuracy, ViT has the highest training accuracy, followed by Swin Transformer, and CNN has the lowest training accuracy; the validation accuracy is similar to the training accuracy. ViT has the highest accuracy, but takes the longest time; conversely, CNN takes the shortest time and has the lowest accuracy. Swin Transformer, which seems a combination of CNN and ViT, is most complex but with ideal performance. In the future, ViT is indeed a promising model that deserves further research and exploration to contribute to the computer vision field.","PeriodicalId":227528,"journal":{"name":"International Conference on Artificial Intelligence and Computer Engineering (ICAICE 2022)","volume":"72 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127130243","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}
Fangju Ran, Chen Xiong, Meng-yao Lu, Tianqing Yang
This paper proposes a method of emotion analysis based on BERT BiLSTM. Firstly, BERT is used to realize the word vectorization, and then Bilstm is constructed to extract semantic features for emotional analysis. In the experiment, the model designed in this paper is compared with the emotional dictionary, SVM, Word2vec LSTM, BERT TextCNN on the college online public opinion comment dataset, and the experiment proves that the accuracy of this model has been improved.
{"title":"Research on university network public opinion sentiment analysis based on BERT and Bi-LSTM","authors":"Fangju Ran, Chen Xiong, Meng-yao Lu, Tianqing Yang","doi":"10.1117/12.2671058","DOIUrl":"https://doi.org/10.1117/12.2671058","url":null,"abstract":"This paper proposes a method of emotion analysis based on BERT BiLSTM. Firstly, BERT is used to realize the word vectorization, and then Bilstm is constructed to extract semantic features for emotional analysis. In the experiment, the model designed in this paper is compared with the emotional dictionary, SVM, Word2vec LSTM, BERT TextCNN on the college online public opinion comment dataset, and the experiment proves that the accuracy of this model has been improved.","PeriodicalId":227528,"journal":{"name":"International Conference on Artificial Intelligence and Computer Engineering (ICAICE 2022)","volume":"199 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127301986","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 low matching accuracy of existing local stereo matching algorithms in weak texture areas, a local stereo matching algorithm based on multi-matching cost fusion and guided filtering cost aggregation with adaptive parameters is proposed. First, use the gradient direction to improve the gradient cost, and calculate the matching cost by combining the gradient cost with the Census transform and color cost. Secondly, the cost is aggregated by the guided filtering of adaptive parameters; Finally, the final disparity map is obtained through disparity calculation and multi-step disparity refinement. The improved algorithm is tested on 15 training sets on the Middlebury3 platform, and the average false matching rates of bad4.0 in all areas and non-occluded areas are 19.9% and 13.2%, respectively, which is improved compared with AD-Census and other algorithms.
{"title":"Multi-cost fusion stereo matching algorithm based on guided filter aggregation","authors":"Jingwen Liu, Xuedong Zhang","doi":"10.1117/12.2671218","DOIUrl":"https://doi.org/10.1117/12.2671218","url":null,"abstract":"Aiming at the low matching accuracy of existing local stereo matching algorithms in weak texture areas, a local stereo matching algorithm based on multi-matching cost fusion and guided filtering cost aggregation with adaptive parameters is proposed. First, use the gradient direction to improve the gradient cost, and calculate the matching cost by combining the gradient cost with the Census transform and color cost. Secondly, the cost is aggregated by the guided filtering of adaptive parameters; Finally, the final disparity map is obtained through disparity calculation and multi-step disparity refinement. The improved algorithm is tested on 15 training sets on the Middlebury3 platform, and the average false matching rates of bad4.0 in all areas and non-occluded areas are 19.9% and 13.2%, respectively, which is improved compared with AD-Census and other algorithms.","PeriodicalId":227528,"journal":{"name":"International Conference on Artificial Intelligence and Computer Engineering (ICAICE 2022)","volume":"13 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123727848","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}
Qin Zhong, Chunyan Zhao, Xin Zhou, Y. Wang, Ling Li
For the Hadamard product of the matrices with non-negative entries, we study the new upper bound for the spectral radius by applying the characteristic value containing the domain theorem. This estimating formula only involves the entries of two non-negative matrices. Hence, the upper bound is easy to calculate in practical examples. An example is considered to illustrate our results.
{"title":"New estimation for spectral radius of Hadamard product","authors":"Qin Zhong, Chunyan Zhao, Xin Zhou, Y. Wang, Ling Li","doi":"10.1117/12.2671104","DOIUrl":"https://doi.org/10.1117/12.2671104","url":null,"abstract":"For the Hadamard product of the matrices with non-negative entries, we study the new upper bound for the spectral radius by applying the characteristic value containing the domain theorem. This estimating formula only involves the entries of two non-negative matrices. Hence, the upper bound is easy to calculate in practical examples. An example is considered to illustrate our results.","PeriodicalId":227528,"journal":{"name":"International Conference on Artificial Intelligence and Computer Engineering (ICAICE 2022)","volume":"57 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122704438","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}
To improve the real-time performance and robustness of traditional feature matching algorithms, an improved image feature matching algorithm E-OrbF based on ORB and FREAK is proposed. In E-OrbF, the original FAST feature points in ORB algorithm are distributed unevenly and redundant. The strategy of subregion and local threshold is adopted to improve the uniform distribution and stability of feature points. Then simplify the sampling mode of FREAK algorithm and design a new feature descriptor. While improving the matching speed, the sampling point pairs are further filtered to improve the matching accuracy. Finally, combine RANSAC matching algorithm to eliminate mismatches and reduce the rate of mismatches. The experimental results show that the algorithm has good real-time performance, while under the conditions of perspective transformation, rotation scale, complex illumination and blur. Both of them can well complete feature detection and feature matching and improve the robustness of existing methods. The algorithm can be applied to the fusion of virtual and real scenes on mobile terminals, and the average visual frame rate reaches 30 FPS, meeting the real-time requirements.
{"title":"E-OrbF: a robust image feature matching algorithm","authors":"Chang Liu, Huan Li","doi":"10.1117/12.2671148","DOIUrl":"https://doi.org/10.1117/12.2671148","url":null,"abstract":"To improve the real-time performance and robustness of traditional feature matching algorithms, an improved image feature matching algorithm E-OrbF based on ORB and FREAK is proposed. In E-OrbF, the original FAST feature points in ORB algorithm are distributed unevenly and redundant. The strategy of subregion and local threshold is adopted to improve the uniform distribution and stability of feature points. Then simplify the sampling mode of FREAK algorithm and design a new feature descriptor. While improving the matching speed, the sampling point pairs are further filtered to improve the matching accuracy. Finally, combine RANSAC matching algorithm to eliminate mismatches and reduce the rate of mismatches. The experimental results show that the algorithm has good real-time performance, while under the conditions of perspective transformation, rotation scale, complex illumination and blur. Both of them can well complete feature detection and feature matching and improve the robustness of existing methods. The algorithm can be applied to the fusion of virtual and real scenes on mobile terminals, and the average visual frame rate reaches 30 FPS, meeting the real-time requirements.","PeriodicalId":227528,"journal":{"name":"International Conference on Artificial Intelligence and Computer Engineering (ICAICE 2022)","volume":"6 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131290594","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}
With the wide application of deep learning, the abstractive text summary has become an important research topic in natural language processing. The abstractive text summary has high flexibility and can generate words that have not appeared in the text. However, the generated summary model will have factual errors, which significantly affect the usability of the summary. Therefore, this paper proposes a text summary model based on fact relationships and keyword fusion. We extract the fact relation triplet in the input text and automatically extract the keywords in the text to assist in the generation of the abstract. The fusion of fact relations and keywords can effectively alleviate the problem of factual errors in the abstract. Many experiments show that compared with other baseline models, our model (FRKFS) improves the performance of summaries generated on the data sets CNN/Daily Mail and XSum and alleviates the problem of factual errors.
{"title":"Fact relation and keywords fusion abstractive summarization","authors":"Shihao Tian, Long Zhang, Qiusheng Zheng","doi":"10.1117/12.2671188","DOIUrl":"https://doi.org/10.1117/12.2671188","url":null,"abstract":"With the wide application of deep learning, the abstractive text summary has become an important research topic in natural language processing. The abstractive text summary has high flexibility and can generate words that have not appeared in the text. However, the generated summary model will have factual errors, which significantly affect the usability of the summary. Therefore, this paper proposes a text summary model based on fact relationships and keyword fusion. We extract the fact relation triplet in the input text and automatically extract the keywords in the text to assist in the generation of the abstract. The fusion of fact relations and keywords can effectively alleviate the problem of factual errors in the abstract. Many experiments show that compared with other baseline models, our model (FRKFS) improves the performance of summaries generated on the data sets CNN/Daily Mail and XSum and alleviates the problem of factual errors.","PeriodicalId":227528,"journal":{"name":"International Conference on Artificial Intelligence and Computer Engineering (ICAICE 2022)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126450009","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：